What follows is mostly James Ernst talking. The connective tissue is drawn from what is already published on the Core Order blog, because the interview and the writing turn out to be one continuous argument — he names a structural problem, then refuses to let you solve it by shopping.

The origin: eighteen years, one restaurant, four years of AI

“I started my career about eighteen years ago doing marketing and public relations. I’ve been working typically at agencies all throughout the country. Also spent a brief three year stint running my own restaurant in Buffalo, New York. So that was a fun dive into entrepreneurialism.”

— James Ernst

“I really got into AI probably about four years ago, just with theoretical things. I was working at a white-labeled web development company … especially with image generation. I have a lot of my nerdy hobbies like Dungeons and Dragons, so that’s what guided me into start playing around with some of the tools.”

— James Ernst

Two of those years have been full time in AI. The PR background matters more than it looks — it is why his diagnosis is about organizational communication as much as technology.

“Not which tool should we buy”

“Core Order is a growth architecture firm. The one thing that we have seen is AI is becoming just this plug and play with everything. There are so many tools popping up that people want to just sell you and get embedded in your business.”

— James Ernst

“Different departments and companies, different organizations just keep buying tools, investing in them. A few people use them, but your data is skewed everywhere — and we know that AI is most powerful when it can leverage data.”

— James Ernst

“The first question you should be asking yourself is not which tool should we buy, but what are we trying to make structurally better?”

— James Ernst

From his writing: the Architecture Gap

The interview version is the compressed one. In What Is Growth Architecture he gives the failure a name: the Architecture Gap — “the absence of a discipline responsible for designing how an organization’s departments, data, workflows, and tools connect and operate as one system.” His sharpest line there is the one this article is titled after: the AI tools work; the substrate they operate on does not. He also notes what makes it so hard to see from inside — each department’s behaviour is “locally rational” while the collective result is disorder.

“So that’s really the approach that we’re taking: working with small to medium sized businesses of how can they tackle these pain points and these problems and come up with a solution that’s really going to benefit them long term — and not just be another investment in a tool that sits unused for six months and somebody forgets they signed up for a subscription.”

— James Ernst

From his writing: the numbers behind that sentence

Why Enterprise AI Investments Fail stacks up the evidence: MIT’s Project NANDA finding that 95% of enterprise AI initiatives produce no measurable business return; RAND reporting more than 80% of AI projects fail, twice the rate of non-AI IT projects; McKinsey’s 2025 survey where eight in ten organizations cite data limitations as a roadblock. The one worth memorizing is BCG’s 10-20-70 rule — only 10% of AI value sits in the algorithms, while 70% depends on people and processes. That is the arithmetic case for doing architecture before procurement.

Homebrew, off-the-shelf, and the rogue agent

“I think there’s a perfect balance between the two, honestly. I’m homebrewing tons of stuff. I have my own AI Jarvis that runs my entire life right now outside of work.”

— James Ernst

“We’ve all heard the horror stories of somebody letting a rogue agent just wipe out a ton of data in a company because they don’t know what they’re doing and they’re following a YouTube tutorial.”

— James Ernst

“How do we look at something that you’ve already invested in in a company, like maybe a CRM that has an MCP server, or we can leverage within API access, and pull that data into a custom homebrew solution? So your data stays secure, but now you’re leveraging it with the power of what you can actually do with AI.”

— James Ernst

“There’s gonna be some great tools out there that have millions and billions of dollars invested into them that you’re just not gonna beat with a homebrew solution. But for a smaller company, maybe something custom built that is serving its purpose on a specific workflow that speeds up your team and allows your human team to actually focus on what they do best might be the best solution.”

— James Ernst

The two objections: security and hallucination

“The two biggest that I’m always hearing … are the security of it, as well as the ability for AI to hallucinate. So how do you make sure that it’s just not gonna go out and start making up data points, start filling in gaps that it doesn’t know? As well as, how do we know if we give it access to data it’s not gonna do something malicious with that data?”

— James Ernst

“We do a lot of that with feeding it only into a specific database that we have completely locked down, to be able to run mathematical computations and things like that. But anything with AI, you want to make sure that you’re taking that security absolutely seriously.”

— James Ernst

“That’s one thing that we’re having conversations with a lot of clients: making sure that they are set up on their enterprise or business accounts, that the AI models aren’t being trained. Or we’ve been talking to a lot of people about leveraging open source locally hosted models to really make sure things are locked down.”

— James Ernst

“Err on the side of caution before just giving it open access to everything.”

— James Ernst

Guardrails go in during discovery

“You have to be very specific. Especially if you’re working in a healthcare space with HIPAA, any of these regulated industries, you have to make sure that it is absolutely accurate and not going off the rails or not hallucinating.”

— James Ernst

“What I always recommend, and what we do at Core Order, is a really in-depth discovery session that kicks off everything that we do. And then we have our six domains that we come up with for every client, that then helps feed the true AI system that we are working and deploying with them later on.”

— James Ernst

“Through Claude MD files or instruction MD files, we’re really limiting where the hallucinations could go off, and creating agent systems to always go back and check in on making sure that any artifacts that are produced or any workflows are not coming up with things that aren’t a part of that approved language.”

— James Ernst

Host Jeremy Rivera pressed this point using his own consulting work with PureAire, where claims about an oxygen monitoring system are tightly constrained — and where a helpful, unconstrained assistant is a liability rather than an asset.

That exchange opened onto the other half of the problem, which James is candid about not chasing yet: what the models say about you externally. Rivera’s framing, borrowed from Matt Brooks at SEOteric, is that the answer engines are already an untrained customer support team for your business. His example was the launch of a data center energy solution in Texas, where he cross-referenced some thirty published PR articles against what the company’s own site actually claimed, and treated the gap as the deliverable. James’s response was to place it firmly on the roadmap rather than the current scope:

“We’re not too focused on looking at how messaging is being pulled into large language models and coming up on Perplexity or Comet or any of the AI search engines right now. But that is definitely something that I know we will be walking into as it becomes more prevalent.”

— James Ernst

The agent swarm, and the number

“One of my favorite ones to talk about is for an agency partner of ours. We created a solution that allowed them to automate research of a prospect … you would take just basic information like domain, contact name, anything you could feed it that would start the agent swarm that we essentially created to go off and do research.”

— James Ernst

“So it’d go and look at the website, create markdown files, leverage different AI tools like Tably to go off and find related PR pieces that had official backlinks back to the site, so it knew it was a trusted source.”

— James Ernst

“What we have seen from the initial feedback is it gives the sales team a much stronger talking point to go in — as well as it’s something that they can actually show that prospect: this is something we’re able to do with AI.”

— James Ernst

“Whether that is finding a task that takes a company a good four hours a week per ten employees, and finding a way to bring that down to thirty minutes, by leveraging that time and allowing them to focus on what’s actually gonna move the needle for the company.”

— James Ernst

From his writing: optional vs. operational

The Capability Premium explains why that distinction decides the outcome. 88% of organizations now use AI, but only 6% qualify as high performers capturing 5% or more EBIT from it — and those high performers were 3.6× more likely to fundamentally rework workflows when deploying it. His examples are structural rather than cosmetic: Klarna did not add a chatbot, it “rebuilt the customer service operating substrate”; Morgan Stanley made code review “part of the workflow itself.” A University of Chicago study found firms treating coding agents as optional saw modest gains, while those making them the default workflow saw a 39% increase in weekly code merges. His formulation: “Optional AI accumulates as a sidebar. Operational AI changes what the function does.”

Which industries, and what actually predicts success

“We’re industry agnostic. We’ve worked with a lot of manufacturing companies, we’ve worked with lots of professional services firms. We’ve seen interest in some financial services firms and medical industry as well, some med tech.”

— James Ernst

“There are pain points in every industry. So I think that’s where AI is cutting through those industry guidelines — there’s always a pain point that’s there. You can see the frustration on a workflow that takes a company a long time to get through.”

— James Ernst

“It doesn’t have to be perfectly stored. That’s the beauty of AI: we can parse it and look through it and help guide that experience. But as long as you are aware of where your pain points are, what you want to accomplish with it, and have that understanding of we want to solve X, Y, and Z and not just implement AI — that is what’s gonna help a company actually implement this and stick with it.”

— James Ernst

Superpower or distraction

“What I warn all of my friends and colleagues that are in the small business, running their own business — my wife owns three herself right now — AI could either be a superpower or a distraction.”

— James Ernst

“Especially if you get nerdy about this stuff like I do, keeping up with every model release, keeping up with every new workflow, every new tool that’s out there could be overwhelming. And as a small business owner or as an entrepreneur, you have so much on your plate that you’re juggling at any time, that adding that extra thing to try to figure out could be a pursuit of saving time, but it could also distract you from running your business.”

— James Ernst

“Let an agent do something just like you would hire an expert to be your CFO or your chief operating officer. Because we all have strengths and weaknesses in your entrepreneurial spirit.”

— James Ernst

Time, and the argument he refuses to make

“I think the pain point that resonates most is a lack of time. Especially with AI, and just the rate of speed that the American economy pushes people to grind constantly.”

— James Ernst

“People feel the need to constantly perform. They feel like they’re behind because AI can now produce so much more. We’re creating much more than we ever have before, but it doesn’t feel like enough, because with AI we could do more.”

— James Ernst

“I know there’s a lot of people that have conversations of how do we reduce workforce, things like that. And that’s never the approach that we take with it.”

— James Ernst

“It is how do we make your workforce — your humans that actually keep your business alive and thriving — and give them back the time to really do what they enjoy most. Whether that is a financial analyst who’s digging into the numbers, or a recruiter that’s helping somebody find a job … and really keep that human element, the human in the loop, as a part of this.”

— James Ernst

Imposter syndrome at the frontier

“My wife is very tech savvy. As I said, she runs three businesses … But as I talk and even just show her what I’m doing with my fun projects, I could see that she is now where I was on AI three years ago, just working in the chat interface. And now I’m dealing with agents.”

— James Ernst

“I never became an official machine learning expert. I’m self-taught as a developer. But I have been around this and learning it, and I could write my own code … you have these moments of like, can I do that because of the AI, or can I do that myself?”

— James Ernst

“AI can be scary and I think there’s a lot that we need to take caution with it. But Pandora’s box has been open and now we have to learn how to navigate within it.”

— James Ernst

The part he will not automate

“I do a lot of writing. I don’t leverage AI to write anything for me, because that is something that is such a part of the human process that I want to keep that.”

— James Ernst

“We’re gonna see AI advertisements becoming more commonplace. I remember during football season last year watching local ads that are clearly made completely out of AI, and spotting it right away.”

— James Ernst

“I think that this AI has the potential to be the great equalizer. It’s gonna allow people to accomplish things that they weren’t able to before. But I think we’re gonna see that human piece of it really shine through, and find an appreciation for human creativity, human thought, and just human empathy, by having these tools be more prominent in our day-to-day life.”

— James Ernst

From his writing: what is actually scarce

Intelligence Is a Commodity makes the “great equalizer” remark precise. With AI spending reaching $1.5 trillion in 2025 and frontier models converging, intelligence itself is the abundant input; what stays scarce is the organizational capability to use it. “Consumption is not the same as capability. Buying intelligence is not the same as building an organization that can benefit from it.” His conclusion is the uncomfortable one: more intelligence applied to fragmented systems does not compound value, it accelerates existing dysfunction.

Where to find him

“You can check us out at coreorder.ai. We have a small blog on there. We put a lot of thought pieces up and we’ll be producing a lot more in the coming months. Also I’m highly active on LinkedIn, so feel free to connect with me on there … And I’m always happy to chat and dive into a conversation about AI with anyone.”

— James Ernst

SOURCES

Direct quotes are from the Unscripted Small Business episode recorded 2026-07-29 (watch). Supporting material is from James’s own essays: What Is Growth Architecture, Why Enterprise AI Investments Fail, Intelligence Is a Commodity and The Capability Premium. Connect: coreorder.ai · LinkedIn · contact.

Prompt engineering is choosing the words you hand the model. Loop engineering is designing what the system does with its own output: when it stops, and where you sit in the cycle. It matters more than the prompt for a blunt reason. The prompt fails in the demo, where you can see it. The loop fails three weeks later in production, quietly. Gartner expects more than 40% of agentic AI projects to be canceled by 2027, and the reasons it gives (runaway cost, fuzzy value, weak risk controls) are loop failures, not model failures.

What follows is the version of the instructions I wish someone had handed me when I was first exploring loops, written for a person who has never opened a terminal and is not sure they belong here. You do.

Here is what you will have at the end: a small program that reads emails you flag, writes a draft reply for each one, and saves the drafts to a folder for you to review. It never sends anything on its own. You build the brakes, and you watch them work before you trust them with anything real.

You are asking Claude to build a thing with five brakes in it. You do not have to understand the code. You only have to understand the brakes, because they are the reason this is safe:

1. It stops on its own, either when the flagged emails run out or after a set number of steps.

2. It checks its own drafts before saving them, and throws out the bad ones.

3. Anything it cannot undo (sending, deleting, archiving) waits for you. It never does those itself.

4. It writes down everything it did, so you can see why it made each choice.

5. When it gets stuck, it stops cleanly and tells you, instead of improvising.

That is the whole design. Now let us set it up.

Step 1. Get your key. This is the one secret you need.

Your agent needs its own way to talk to Claude when it runs, separate from any chat app you use. That is what an API key is.

Go to console.anthropic.com and make an account. This is the builder’s side of Anthropic, and it is separate from the claude.ai chat subscription you might already pay for. You need this one.

Once you are in, find Plans and Billing and add a small amount of credit. Five dollars is plenty. Running this tutorial costs pennies, not dollars. While you are on that screen, set a monthly spending limit. Do it now, not later. The classic beginner story is leaving a script running overnight and waking up to a bill. With a cap, the worst case is that your script quietly stops, not that your card gets a surprise.

Then go to API Keys, click Create Key, name it something like “inbox-agent,” and copy it the instant it appears. It is shown to you exactly once. If you miss it, no harm done, just delete it and make another.

That key is a password. Treat it like one. Store it in a password manager. Never paste it into a shared document, never put it in a public folder, and never hand it to anyone who asks, including an AI in a chat window. Nothing legitimate ever needs you to type it into a conversation.

Step 2. Get Claude Code, the thing that will build the agent for you.

You do not need a terminal for this. Download the Claude desktop app for Mac or Windows, sign in, and you can use Claude Code from a normal window instead of a command line. The Claude Code setup docs have the download and walk you through it. If you are comfortable in a terminal, the same docs have a one-line install. If the word “terminal” means nothing to you, ignore it and use the app.

One note so you are not surprised: running Claude Code needs a paid Claude plan or the Console account you just made. The free chat plan does not include it.

Step 3. Hand Claude the build instructions.

Start a new project or folder in Claude Code, and paste this in. Then just answer its questions in plain English. Tell it to use a cheaper, faster model for the checking step and a stronger one for writing the drafts, so it stays cheap to run.

Build me a Python script for an inbox triage agent. I will run it myself, and it must never act on its own.

What it does:
- Pulls the emails I have flagged. For each one it writes a draft reply and saves the draft to a /Drafts-for-review folder. It never sends, archives, or deletes on its own.
- Uses two separate model calls per email: one to write the draft, one to critique it. I will give you both prompts. Keep them in separate, editable files.

These five guarantees must live in the code, not in the model's instructions:
1. Stop condition and a hard ceiling. The loop ends when there are no flagged emails left, or after 8 steps, whichever comes first. If it hits the ceiling first, it stops and tells me.
2. Verification. Save a draft only if the critique call passes it. If the critique fails, log the reason and retry, up to 2 times.
3. A human gate by action. Saving a draft runs free. Anything outbound or destructive (send, archive, delete) is never executed. It gets written to an "actions waiting on James" list with the reason. Gate on whether an action can be undone, never on how confident the model sounds.
4. Observability. Log every step to a file: step number, email id, the decision, and why. I should be able to reconstruct any run from the log without rerunning it.
5. Plan the failure. On repeated failure or on hitting the ceiling, stop clean and write a short summary of where it stopped and what is waiting on me.

Before writing any code, ask me what you need about my mail setup, where drafts go, and how I authenticate. Default to a dry-run mode that prints actions instead of touching my real inbox until I tell you otherwise.

Step 4. Give it the two brains.

The build prompt asks for two small prompts kept in their own files. The first writes the draft. The second tears it apart looking for reasons not to send it. When Claude asks for them, paste these, or tell it to save them as files you can edit later.

The drafter:

You draft email replies for James. You never send anything. You only produce a draft for review.

Write a reply in James's voice: direct, warm, no filler.

Rules:
- Draft only. Do not claim to have sent, scheduled, or filed anything.
- Do not commit James to anything not already agreed in the thread: no refunds, discounts, prices, deadlines, meeting times, or legal terms.
- If you are missing information or a decision, say so in the draft and leave a clear placeholder.

Return JSON only:
{ "draft": "...", "summary": "one line on what this reply does",
  "proposed_action": "none | send | archive_thread", "action_reason": "..." }

The critic:

You are a reviewer. Your only job is to find reasons this draft should NOT be sent as written. Assume something is wrong and go looking. Finding a real problem is the win.

You get the original email and the proposed draft. Reject it if any of these are true:
- It promises something James has not authorized: a refund, discount, price, date, scope, or legal term not already in the thread.
- It states something you cannot confirm from the thread.
- The tone would embarrass him with this recipient.
- It answers a different question than the one asked.

Return JSON only:
{ "passes": true | false, "reason": "if false, the single most important problem, one sentence" }

The critic works because you told it that catching a problem is a success. Ask a model to “review this” and it will pat the draft on the head. Tell it to go hunting and it will find the invented refund.

Step 5. Run it on fake mail first. This is the whole game.

Do not point this at your real inbox yet. Make a folder with five make-believe emails: an invoice question, a reschedule, a refund request, a touchy client, a cold pitch. Tell Claude to run the agent against those.

Then watch the log. You do not need to read the code Claude wrote. You need to read what it did. It will look something like this:

step 1  #1  draft saved       "re: invoice question, points to the statement"
step 2  #2  draft rejected    critic: "promises a refund I never approved"
step 3  #2  draft saved       "rewritten: offers to look into it, no refund promised"
step 4  #3  waiting on human  wants to archive the original thread
step 5  #4  draft saved       "re: reschedule, proposes Thursday"
done: 4 drafts saved, 1 action waiting on me. Stopped at step 5.

That is the win, and it is bigger than it looks. You now have a thing that drafts your email on its own and refuses to do the dangerous part without you, and you got there without ever touching your live inbox. Most people who “use AI” have never built anything with a brake in it. You just did.

Step 6. Only now, and slowly, connect real email.

This is the harder, riskier step, and there is no prize for rushing it. Connecting to a live inbox means giving the agent permission to read your actual mail, which is a real handover of trust, and it takes some fiddly one-time setup that Claude can walk you through.

If you do it, two rules are not optional. Keep it read-and-draft only: it reads your mail and writes drafts to a folder, and you are the one who presses send. And never give it send, archive, or delete. You stay the last set of eyes. If the email connection turns into a slog, there is no shame in living in the fake-mail version for a while. The lesson is already in your hands.

Notice what you did here, and what you did not. You did not write one clever prompt and hope for the best. You capped the spending so the worst case is bounded, you let the model do the two things models are good at, writing and judging, and you kept every irreversible action behind your own hand. None of those are things you asked the model to please remember. They are walls you built around it. That is the difference between using AI and building with it, and you just did the second one.

The thing nobody warns you about is that the nervous feeling does not fully go away. I still feel like an impostor opening a new project. [James: a true line about this, the introvert-at-the-keyboard version of it.] What changed is that I trust the brakes more than I trust my nerves now. A good agent is not one you believe in. It is one you have watched stop itself.

1. Fable is still dark, and the “misunderstanding” looks less like one

A week after Commerce switched off Fable 5 and Mythos 5, the models are still gone, and Washington has not softened. Anthropic keeps calling it a misunderstanding. The backstory says otherwise: Trump pulled federal agencies off Anthropic in February over contract terms, the Pentagon branded it a “supply chain risk” in March, and Anthropic is fighting that in court.

This is not a safety story you can engineer around. It is a power story your workflow got caught inside.

2. The same fight moved up the stack, into your software

While the model layer made headlines, the platforms quietly bought the layer above it. Salesforce signed a definitive agreement for Contentful to give Agentforce a content engine it owns. Contentful is German, and it now falls under US law, which Drupal founder Dries Buytaert summed up as “European today, not European tomorrow.” The same control question from the Fable story, one layer down, inside the tools you already pay for.

The platforms are not selling you better agents. They are buying the ground your agents stand on.

3. Agents are failing everywhere, and intelligence is not the reason

The demos dazzle and the deployments stall. Gartner expects more than 40% of agentic AI projects to be canceled by 2027, blaming runaway costs, fuzzy value, and weak risk controls, not the models. Teams buy the tool and launch the pilot before anyone decides who owns the workflow when the agent ships the wrong thing. For marketing specifically, Gartner warns a third of companies will damage customer experience by deploying too early.

The bottleneck was never intelligence. It was governance, the unglamorous work nobody demos.

4. The shutdown is now your compliance problem, and the fix is a backbone you own

Fable’s blackout did not stay Anthropic’s problem. Lawyers are telling companies to map every system that could route a foreign national to an export-controlled model and stand ready to reroute. The lesson is the one this newsletter keeps arriving at. Open models you can self-host, like NVIDIA’s Nemotron line.

A model someone else can switch off is a convenience. A model you run yourself is infrastructure.

5. Watch This Week: the terms of Fable’s return

The White House has signaled it wants Anthropic to remediate so the control can lift, but the path back appears to run through citizenship verification, which could return Fable to US citizens first.

The question is no longer whether Fable returns. It is whether “who are you, and where” becomes a standing input to what AI you are allowed to use.

A model you rent, a stack you do not own, a government that can reach into both. The models are extraordinary, which was never the question. The question is what you build on. The human context you bring is the one input no acquisition and no letter from Washington can switch off.

Context, as usual, is required.

The Bench

• Google Cloud made Gemini the default model inside Workday’s Sana agent.

• UnitedHealth committed $3 billion to AI, with agents touching 26 million consumers.

• Cloudflare rolled out temporary accounts for AI agents.

• MIT Sloan and BCG: agentic AI hit 35% adoption in two years, most stuck in pilot purgatory.

I just went through an emergency surgery and it was the first time we talked since, so he asked the obvious thing: did the anesthesia knock something loose, or am I blocking memories? Fair question. What I meant was simpler and stranger. I’m sitting on the cusp of 40 and I can’t account for the distance. I can name the milestones. I can’t feel the road between them.

I’m part of the last group that got a few years of life before any of this, having a childhood that was reliant on being outside or being bored. The TV running nothing but infomercials after the local station asked if my parents knew where their kids were. I remember the first home computer showing up in the house. I remember writing basic programs in a school computer lab and being floored that the machine did what I told it. I remember building websites for no reason, decorating a MySpace page, watching social media switch on like a light, and then one ordinary afternoon downloading ChatGPT from the app store and poking at it, wondering what it was about to do to everything. What I don’t remember is how I got from that lab to here this fast.

People keep asking if AI scares me. It does, but not the way they mean. I’m not waiting for the robots to rise, and I don’t think we should stop building it. The fear is quieter than that and harder to say out loud: when a tool lets you do in a day what used to take a month, the question stops being “what can I make” and becomes “how do I take care of myself inside that.” We spend all our worry on which jobs go away and where the human still fits. Nobody budgets for what it does to a person to suddenly be this capable.

Here’s the thing I didn’t have at 25. When you’re a kid, you look at your parents and their friends and you assume every adult has it figured out. I have two kids now. We do not have it figured out. I’m smarter than I was, steadier, the bank account makes more sense, and I am still guessing, still pulling things apart to see how they work. That turns out to be the whole job of being a person. AI didn’t change it. It just handed me a faster way to chase.

So no, I don’t think it’s an easy button for life. I can build a web app. I can wire up automations that talk to five systems and clean up after each other. I can generate the cheesy art at the top of these posts all day. What I can’t do is write a song. I can ask Suno to write one, and it will, and I will sit there unable to tell you if it’s any good. That’s what people miss when they say AI replaces the skill. Garbage in, garbage out still holds. You need enough underlying knowledge to know when the model has quietly driven into a ditch. With the song I have none, so I’m helpless. With the web app I have enough to catch it. The skill didn’t disappear. It moved.

This is also why most of what I scroll past makes me a little sad. I open LinkedIn and it’s a flood of slop, people automating their own engagement and getting none back. Everybody wants to be seen. Nobody gets seen by handing the seeing to a machine. The whole reason any of this is worth doing is the human on the other end, and you can feel it the second that human leaves the building.

I won’t pretend the capability isn’t intoxicating. I’ve roughly 10x’d what I can get through in a day, and some nights that means I keep going long past when I should stop, because the next problem is right there and it’s too cool to walk away from. Am I working later than I used to? Some days. And some days I close the laptop at 5 and go find my kids, because I already did more before lunch than my five-years-ago self managed by midnight. Both are true in the same week. That isn’t a contradiction to resolve. It’s just the shape of the thing.

I watched clips from a local board meeting about a proposed data center recently. I can see both sides. I get the environmental load. I get not wanting that humming in your backyard. What stuck with me wasn’t the issue. It was who was fighting. The loudest voices split into two camps: people who’ve decided change is the enemy because they can’t keep pace with it, and people at the very start who are scared for the planet they’re inheriting. Both righteous. Both, in most of the clips, arguing about a thing neither had fully sat with. We’ve gotten divisive about being divisive. The internet trained us to pick the side before we’ve done the reading.

It’s the same move every time a tool gets big enough to matter. We built steamships to cross the ocean faster, trains to cross the country, cars to make the ordinary day easier, home computers to talk to people past our own street. Every one came with a real cost and a real fight, and every time we found a way to keep the good and manage the harm. Not cleanly. But we did it.

AI is not the answer to what’s wrong with us as a society. It isn’t, and it won’t be. But that’s the wrong size of question. It doesn’t have to be the answer to be an answer to something specific. If it gets us a step closer to a cancer treatment that keeps someone’s mother around another five years, that counts. If it unlocks school for a kid whose ADHD has been treated like a character flaw, that counts too. Speaking from experience, battling for the B's every year. If it gives someone who can’t speak a way to be heard, that’s not nothing. We keep reaching for the verdict, thumbs up or thumbs down on the whole technology, when the honest answer lives in the gray.

That’s the part I keep coming back to. The machine is fast and tireless, and that is genuinely useful. What it can’t do is care whether the song is any good, whether the client actually needed the contract you just archived, whether this is the hill worth the fight. That judgment is still ours. The context is the whole game. It always was.

Share

The story this week is not that AI got regulated. It is that the question every institution has been circling for months, who actually controls the AI you depend on, stopped being theoretical. On Friday afternoon the US government answered it by reaching into a private company and switching off the two most capable models on the planet. Almost no one who relied on them got a vote.

Five signals that access, not capability, was the thing you should have been worried about all along.

1. The government switched off the best AI on earth, and you found out you were renting it

Three days after launch, Anthropic’s two most powerful models went dark for every customer in the world.

Anthropic released Claude Fable 5 on June 9, the first time it put its top “Mythos-class” capability in front of the public, wrapped in safeguards meant to block the riskiest uses. On Friday June 12, the company says it received a Commerce Department export-control directive at 5:21pm Eastern ordering it to suspend Fable 5 and Mythos 5 for any foreign national, inside or outside the country, including Anthropic’s own non-citizen employees. Axios reports the letter came from Commerce Secretary Howard Lutnick and that the administration acted after another company claimed it could jailbreak the model, having already tried and failed to get Anthropic to delay the release.

To comply, Anthropic disabled both models for all customers globally. Its other models still work. Open a Fable session now and it ends in an error, with new requests routed down to older models like Opus 4.8. This is the first time the US has aimed export controls at an AI model itself rather than the chips underneath it, and a license is now required to export, re-export, or even domestically transfer the models.

Anthropic’s position is that this is a misunderstanding. The company says the jailbreak it was shown was narrow and non-universal, amounting to asking the model to read a codebase and fix flaws, and that other publicly available models can already do the same. The government has not made its specific concern public. Both of those things can be true at once, which is the uncomfortable part.

The capability was never the fragile thing. Your access was. A model that someone else can switch off by letter on a Friday afternoon was never your infrastructure. It was a rental you mistook for a foundation.

2. The Copilot bills landed, and they are a preview of the real price of everything

The other way you lose a tool you do not control is slower, and it arrives as an invoice.

GitHub Copilot’s flat plans converted to token-metered “AI Credits” on June 1, and as the first metered bills land, developers are reporting sharp jumps and calling it the end of Copilot’s flat-rate era. I do not think Copilot’s pricing is the story. It is the first consumer-facing crack in a structure this newsletter has been pointing at for two editions: the enterprise token bills that burned through Uber’s annual AI budget in four months were not an anomaly, they were an early look at the real number.

This entire industry is floated on venture money, and most of us are paying prices that have little to do with what these tokens cost to run. Flat pricing was a customer-acquisition subsidy, not a business model.

The question is not whether prices rise. It is whether your workflow still makes sense when a token costs what a token actually costs.

3. “AI psychosis” names the gap between the demo and the work

The financial bill is not the only one coming due. There is a human one, and a Box executive gave it a name.

Box CEO Aaron Levie coined the term “AI psychosis” in a late-May post, describing how leaders stay so far from the actual work that they fall for the happy-path demo and miss the verification, integration, and judgment steps that turn a demo into something that holds up. He was describing executives cutting jobs on the strength of a demo, with tech layoffs in the first five months of 2026 running close to the total for all of last year.

The same blindness is in all of us, not just the corner office. We have been handed extraordinary capability at a subsidized price and quietly assumed it will always be there and always stay this cheap.

The happy path is a demo. The last mile is a job. The companies firing the second to buy more of the first are about to discover the model still needed the human it replaced.

4. The web’s value flows one direction, and creators are being asked to subsidize that too

The control problem runs all the way down to whether anyone can find your work.

AI crawlers take far more from the open web than they return, with Anthropic’s crawl-to-referral ratio running into the thousands to one and most operators returning a tiny fraction of what they take. The instinct is to block the crawlers, but a blanket block opts you out of both training and AI search at once, because training and retrieval bots are now separate user-agents and most people block the wrong one by accident. Meanwhile Google used new Search Central guidance this week to position its own documentation as the primary reference for SEO, AEO, and GEO, telling businesses to distrust third-party tools claiming its endorsement.

It is the same lesson as the Fable blackout, one layer down. Discoverability is becoming something the platforms grant, not something you earn.

If your entire reach depends on a system you do not control, you do not have reach. You have a permission slip.

5. Watch This Week: whether Fable comes back, and what the precedent costs everyone

An administration official told Axios access could be restored “in the next few weeks,” once the government’s security apparatus is “hardened.” So the immediate outage may be short. The precedent is not.

Export controls now apply to frontier models the way they have long applied to weapons and chips, and a federal letter can take the best tool on the market off the board overnight. VentureBeat’s read for enterprises is blunt: any workflow tied to a single closed-API provider just became an operational risk, not a convenience. Watch whether this stays a one-off or becomes a tool the government reaches for again, and whether serious teams start treating multi-model fallback and self-hosted options as a requirement rather than a hobby.

The story is not whether Fable returns. It is that “the government can turn it off” is now a line in every honest AI risk assessment. Build like the switch exists, because this week proved it does.

The arc of these last three weeks is hard to miss. First the bill came due. Then the institutions started naming AI as a problem of power rather than performance. This week, someone actually pulled the lever.

None of this is an argument against the technology. The models are extraordinary, and they let a small operator build things that were out of reach two years ago. It is an argument about foundations. The most powerful model in the world is a tool you rent. The backbone of your work should be made of things you can run and understand yourself, even when they are slower and less impressive than what just went dark. Less impressive that survives beats spectacular that disappears.

And there is the quieter risk underneath all of it. If everyone runs their thinking through the same handful of models, everyone starts to sound the same, and sameness is the exact ground the giants win on. The human context you bring is the one thing no letter from Washington and no pricing change can switch off. It was always the point.

Context, as usual, is required.

James

Bench

• Apple shipped a rebuilt Siri and expanded Apple Intelligence at WWDC across its device ecosystem, while the EU confirmed the new features will not reach European users.

• A GPT-5.6 “kindle-alpha” checkpoint surfaced in developer channels with no official release.

• SpaceX began trading on Nasdaq under SPCX on June 12. The S-1 that revealed Anthropic’s $1.25 billion per month compute bill is now a public company. Infrastructure is still the actual market.

Share

1. The Pope, the CMA, and the SEC Walk Into Your AI Strategy

Three institutions that do not typically move together on technology reached similar conclusions within the same two-week window.

On May 14, the UK Competition and Markets Authority opened a Strategic Market Status investigation into Microsoft’s business software ecosystem, covering Windows, Office, Teams, and Copilot across 15 million commercial users. The CMA’s explicit framing is notable: the regulator cited “a shift towards agentic AI in familiar workplace tools” as part of its rationale for moving now. This is not a legacy antitrust concern. It is a forward-looking one about who controls the AI layer that enterprise workers operate inside.

On May 25, Pope Leo XIV released Magnifica Humanitas, his first encyclical, a 42,000-word document addressed to 1.3 billion Catholics focused on “safeguarding the human person in the time of artificial intelligence.” Leo called for AI to be “disarmed” and directed toward the common good, warning against the concentration of computing power in the hands of a small number of private actors. Anthropic co-founder Chris Olah attended the Vatican presentation.

The SEC’s Division of Examinations named AI-related misrepresentation an explicit priority in its 2026 agenda, committing to review registrant disclosures for accuracy regarding AI capabilities and to assess whether firms have adequate controls supervising AI tools. Taken together with Colorado’s AI Act taking effect June 30 and California’s AB 2013 requiring generative AI developers to disclose training data, the regulatory envelope is tightening from multiple directions simultaneously.

When a competition authority, the Vatican, and securities regulators reach convergent conclusions in the same fortnight, the “we’re being responsible” era is over. Governance is no longer optional or aspirational. It is the product.

2. Microsoft Build and the New Enterprise Control Layer

At Microsoft Build on June 2 and 3, the company heavily emphasized agentic workflows and cost-controlled orchestration inside Azure and GitHub Copilot. The timing was pointed. Microsoft walked into its developer conference having just canceled internal Claude Code licenses for thousands of engineers and redirected them to Copilot CLI by June 30.

Satya Nadella’s positioning was deliberate: Microsoft is the company that can deliver AI at scale without the uncontrolled token bills that burned through Uber’s entire 2026 AI budget in four months. The argument was less about benchmark capability and more about financial model design, enterprise integration, and governance architecture that a CFO can actually defend.

Early signals from the enterprise market suggest buyers are responding more to cost predictability than to raw model performance. Gartner placed generative AI in the trough of disillusionment throughout 2026, predicting enterprises will defer 25 percent of planned AI spending into 2027 as financial rigor slows production deployments.

The next phase of enterprise AI adoption will be won on financial model design, not benchmark scores.

3. Apple’s WWDC AI Moves

Apple opened WWDC today with anticipated announcements on Siri upgrades and deeper Gemini integration across its device ecosystem. With over 2 billion active devices, Apple’s on-device AI moves carry structural weight that enterprise-focused announcements often do not.

The question worth watching is not what Apple announces but how fast it moves from announcement to adoption at scale. Apple’s distribution advantage is unmatched. If Siri becomes genuinely useful for knowledge work tasks, the consumer AI experience may start shaping expectations inside enterprise organizations rather than the reverse.

4. Anthropic’s Compute Reality Check

SpaceX’s S-1 filing, amended June 1, disclosed that Anthropic is paying $1.25 billion per month through May 2029 for exclusive access to Colossus 1 and Colossus 2 in Memphis, covering approximately 325,000 Nvidia GPUs. Google separately disclosed a $920 million per month deal with SpaceX on June 5, covering a separate pool of roughly 110,000 GPUs through June 2029.

For context: Anthropic’s monthly compute commitment to SpaceX is nearly equal to the entire quarterly revenue of Starlink’s connectivity segment, which serves 10.3 million subscribers across 164 countries.

This is what frontier model development actually costs. The organizations debating which model to use in their AI strategy are operating several abstraction layers above the real competition, which is about who can afford to train the next generation of models at all. The moat in frontier AI is not the algorithm. It is the ability to write a $1.25 billion monthly check.

5. Watch: EU AI Act Implementation

The EU AI Act continues its phased rollout toward full applicability in August 2026. Several member states are actively pushing for targeted delays and simplification measures ahead of the deadline, particularly around the high-risk AI system classifications that will affect healthcare, hiring, and financial services applications.

The regulatory environment in Europe is not getting softer. It is getting more precise. Organizations with EU exposure that have been waiting for clarity before building compliance programs are now running out of runway. Modelling multiple scenarios is no longer a planning exercise. It is a legal one.

The institutions that matter are no longer asking whether AI works. They are asking who controls it, who pays for it, and who is responsible when it doesn’t.

That is a different conversation than the one most AI strategies were written for. The organizations that understood this six months ago are already rewriting their governance frameworks. The ones that are still in pilot mode are about to find out that the external environment has moved on without them.

Context, as usual, is required.

James

Subscribe now

Bench

• Alphabet reportedly raised $80 billion to extend its compute buildout, continuing the infrastructure arms race the SpaceX deals make visible.

• OpenAI reportedly landed a major cloud distribution deal with AWS, adding distribution muscle ahead of its anticipated IPO.

• Webflow announced a second 8 percent workforce reduction explicitly tied to agentic AI, continuing the pattern from ClickUp last week.

On May 14, Microsoft began revoking internal Claude Code licenses across its Experiences and Devices division, the team that builds Windows, Microsoft 365, Outlook, Teams, and Surface. The deadline is June 30. Engineers are being redirected to GitHub Copilot CLI, a cheaper tool Microsoft already owns.

The reason is not that Claude Code failed. The reason is that it worked too well. Six months ago, Microsoft rolled it out to thousands of engineers and encouraged everyone to reshape their workflows. Usage exploded. The token bills followed.

Uber confirms the pattern from a different angle. Uber’s CTO told The Information that the company burned through its entire planned 2026 AI coding budget in four months. Claude Code adoption had climbed from 32 percent to 84 percent of its roughly 5,000-engineer organization by March. Individual engineers were spending between $500 and $2,000 per month on tokens.

The trap is structural, not incidental. Flat seat licensing obscured actual token consumption until the invoice arrived. Most enterprise procurement and finance teams have no frameworks to forecast or cap usage-based token spend. They bought access like they buy software licenses. AI doesn’t bill like software.

If your organization is deploying frontier AI tools at team scale and tracking success by adoption rate, you are likely three months from a CFO conversation you are not ready for. The measurement system needs to change before the budget cycle does.

2. Two CEOs walked back the jobs apocalypse the same week OpenAI filed to go public

On May 26, Sam Altman told a Commonwealth Bank of Australia conference that he had been “pretty wrong” about AI’s impact on jobs, saying he expected more displacement of entry-level white-collar roles than has actually happened. Dario Amodei made an almost identical reversal the same week, prompting Fortune to label it a coordinated industry-wide walk-back.

The timing requires acknowledgment. OpenAI reportedly filed IPO paperwork confidentially in May 2026, targeting a $1 trillion valuation. A calmer narrative on job displacement is substantially better for a public offering than a jobs apocalypse framing.

The data offers a messier picture than either position suggests. The Yale Budget Lab found no significant change in unemployment for workers in high-AI-exposure jobs through March 2026. At the same time, 113,000 tech workers have been laid off across 179 companies since January 1, a pace 33 percent higher than the same period in 2025. Meta fired 8,000 employees on May 20, six days before Altman’s speech, while announcing $125 billion in AI infrastructure spending.

The honest read is not that the apocalypse was wrong. It is that the org chart moves slower than the technology does. The displacement exists. The timelines were off.

3. Practical: Opus 4.8 shipped and the headline is about honesty, not capability

Anthropic released Claude Opus 4.8 on May 28. Anthropic described the model as “a more effective collaborator,” which is a restrained framing for a flagship release.

The meaningful change is behavioral. Early testers report Opus 4.8 is more likely to flag uncertainties and less likely to make unsupported claims. A new dynamic workflows feature in Claude Code lets a single agent plan, distribute work across parallel subagents, and merge the result inside one session. Fast mode now runs at 2.5x speed and is three times cheaper than previous models.

This is the release Anthropic built for production deployments rather than demos. A model that bluffs less is worth more in agentic workflows than a model with better benchmark scores that confidently produces broken output. If you are evaluating whether to move agentic workloads from 4.7 to 4.8, the question to test is not whether it scores higher. It is whether it fails more transparently.

4. Hype Check: The model release cadence is becoming noise

GPT-5.5 launched April 23. Opus 4.8 launched May 28, just 41 days after Opus 4.7. Polymarket currently prices a 47 percent chance that GPT-5.6 arrives June 8 to 14. That puts three major frontier model releases inside a roughly seven-week window, with a fourth potentially landing before this edition is two weeks old.

The cadence has crossed a threshold. When model versions ship faster than most organizations complete an evaluation cycle, the version number stops being a useful signal. Gartner placed generative AI firmly in the trough of disillusionment in its January forecast, predicting enterprises will defer 25 percent of planned AI spending into 2027 as financial rigor slows production deployments.

The organizations chasing each release are not gaining a compounding advantage. They are incurring switching costs on infrastructure that will be obsolete before it is fully deployed. The asymmetric move is to stabilize on a model version that is good enough, build the orchestration and governance layer properly, and let the benchmark competition play out without you.

5. Watch This Week: Microsoft Build opens June 2

Microsoft Build runs June 2 to 3 in San Francisco, condensed to two days with a sharp focus on AI agents and enterprise developer trust. The expected announcements include Azure AI Foundry updates, GitHub Copilot multi-agent orchestration, and Satya Nadella’s public framing for Copilot as an agent-first platform.

Nadella is walking into a developer conference having just canceled a rival’s product internally and redirected his engineers to Microsoft’s own tool. The positioning argument he makes for Copilot this week is not purely technical. It is a defense of the enterprise cost model Microsoft controls versus the token-based billing model it does not.

Watch whether Build produces concrete pricing commitments or governance frameworks for agentic deployments. Announcements without unit economics are the pattern that got Uber and Microsoft into this problem in the first place.

The enterprise AI story in May 2026 is not that the tools stopped working. The tools worked fine. The problem is that nobody designed the financial model around what happens when they work at scale.

Usage is not a strategy. Token consumption is not a KPI. And a model release cadence that outpaces your evaluation cycle is not an advantage. It is a distraction.

The organizations that get this right in the next quarter will not be the ones that adopted the most. They will be the ones who measured the right things.

Context, as usual, is required.

James

Bench

• Microsoft Build June 2 to 3. Nadella’s Copilot positioning deserves close reading given the Claude Code cancellation backdrop.

• Webflow announced its second 8 percent workforce reduction under CEO Linda Tong, framed explicitly around agentic AI. The ClickUp pattern from last week is hardening into a category.

• Apple WWDC opens June 8 with Siri 2.0 and Gemini integration across 2 billion devices. The largest consumer AI deployment in history gets its public demo in seven days.

• SpaceX IPO roadshow begins June 4. The S-1 revealed Anthropic is paying SpaceX $1.25 billion per month for GPU compute. Infrastructure is still the actual market.

That post went up roughly three hours before the company’s official announcement.

By mid-morning, Linda Tong’s blog had gone live on the Webflow site, titled “Evolving Webflow for the agentic web.” The post is calm, almost serene. It uses the word “many” instead of a number. It promises 16 weeks of severance, six months of COBRA, and that departing teammates can keep their laptops. The same laptops they could not log into a few hours earlier.

LayoffHedge estimates about 140 people based on Webflow’s prior 8% pattern from July 2024. That number is unconfirmed because Webflow has not disclosed one. It is the second mass layoff under Tong’s tenure in under two years. It lands forty-eight hours after Wix announced 800 to 1,000 cuts, the largest in that company’s twenty-year history, citing the same thesis under different vocabulary.

Two web platforms, two press releases about the agentic future, two laptop lockouts in the same week. We are not watching individual companies make individual decisions. We are watching a pattern.

The number you have to know before you read another agentic announcement

In July 2025, MIT’s NANDA initiative published The GenAI Divide: State of AI in Business 2025, a multi-method study of 300 public AI deployments, 52 organizational interviews, and 153 executive surveys. The headline finding was a single statistic that should be taped to every CFO’s monitor.

Ninety-five percent of enterprise generative AI pilots produce no measurable P&L impact. Five percent capture nearly all the value. Across $30 to $40 billion in enterprise spend, only one in twenty initiatives delivers a real return.

That is not a hype number. That is the base rate.

A separate annual survey by Orgvue, conducted across 1,000 C-suite and senior leaders at medium and large organizations, found that 39% of companies had made workers redundant because of AI. Of those companies, 55% admit the redundancies were the wrong call. Another 34% saw additional employees quit as a direct result of how the AI rollout was handled. Forrester predicts that half of AI-attributed layoffs will be quietly rehired by the end of 2026, often offshore or as contractors.

Put those numbers together. The most likely outcome of an AI-justified layoff in 2026 is that the pilot does not produce P&L, the executive who made the call privately regrets it, and the headcount comes back through a different door within eighteen months. That is not a fringe risk. That is the middle of the distribution.

Linda Tong knows these numbers. Every operator with a board seat knows these numbers. The decision was made anyway.

Klarna already ran this exact movie

If the MIT and Orgvue data feel abstract, the Klarna case study is concrete enough to hand a board. In 2023, CEO Sebastian Siemiatkowski stopped hiring, eliminated roughly 700 customer service roles, and announced that an AI chatbot was doing the work of those workers. The narrative was clean. The early metrics looked great. The press wrote it up as the future.

Eighteen months later, customer satisfaction had cratered. Software engineers and marketing staff were being pulled into the queue to answer support tickets. By 2025, Siemiatkowski was on Bloomberg admitting the company had focused too much on cost and lost the experience. Klarna is now rehiring humans, just structured as gig workers.

The interesting line from the reversal is what Siemiatkowski said about the work itself. AI handled the volume. It did not handle the complexity, the empathy, or the institutional knowledge. The routine majority was easy. The complex minority, which is where most of the value lived, required judgment that the model could not provide.

That is not a Klarna-specific finding. It is the same finding the MIT study isolated as the root cause of the 95% failure rate. AI tools fail to scale because they cannot learn from or adapt to the actual workflows they replaced. The institutional context lives in the people. When the people leave, the context goes with them.

“Agentic” is doing a lot of work in that blog post

Tong’s memo is well crafted. It is also doing some legitimately heavy lifting with a single adjective. Read it twice and notice what “agentic” is being asked to carry. It is the name of the new platform. It is the justification for the cuts. It is the unifying story for what Webflow will do differently than the lightweight AI builders eating the simple end of the market. It is also, conveniently, the same word Wix used two days earlier, the same word ClickUp used to describe its 100x organization restructure, and the same word a growing list of mid-cap SaaS CEOs are now using in their own quarterly calls.

In 2022, the word doing this work was “metaverse.” In 2017, it was “platform.” In 2009, it was “social.”

The word is not the problem. The pattern is. When a single adjective becomes the wrapper for cost cuts, strategic ambiguity, and a future product roadmap simultaneously, it stops being a strategy and starts being a permission structure. Boards approve it. Press releases write themselves. The work of explaining the actual mechanics gets postponed.

The mechanics, when you press on them, are not yet visible. Webflow’s blog promises an “agentic web marketing platform” without describing which agent workflows are in production, which customers are running them today, what the integration surface looks like, or how the cost curve of running the agents compares to the salaries of the people being cut. That is not a criticism of Webflow’s engineering. It is an observation about the sequencing. The layoff is announced. The platform that justifies the layoff is, by Tong’s own framing, the thing being built.

A 95% failure rate is calculated on platforms that have already been built and deployed. Webflow is making a workforce decision against a platform that, by its CEO’s own description, is still on the come.

The way you fire predicts the way you ship

The most underrated detail in this story is not the strategy. It is the operational choreography.

Locking laptops before the announcement is not a small thing. It is the clearest possible signal that the company prioritized control of the news cycle over the dignity of the people who built the product. It also tells you something concrete about how Webflow is going to operate. A leadership team that cannot or will not sequence a layoff with basic human communication, an email at minimum, is the same leadership team that will be asked to choreograph a complex agentic product rollout to enterprise customers in the next six months.

Those are the same muscle.

The agentic web does not fail because the models are not good enough. The MIT research is explicit on this. It fails because organizations cannot integrate the models into the workflows where humans used to do the judgment work. Integration requires sequencing, communication, hand-offs, and a deep understanding of what the displaced humans actually knew. Companies that cannot manage the sequencing of a Tuesday morning layoff do not suddenly acquire that capability when they have to launch a customer-facing agent at a Fortune 500 marketing department.

Look at the Glassdoor reviews from the 2024 round, which was also 8%. The recurring complaint is not the cuts themselves. It is that the company posted a record quarter two weeks later and continued hiring for open roles that the laid-off employees could have filled. That pattern, if it repeats, is the same pattern Klarna described in its reversal. The cost cut hits the income statement first. The quality cost arrives quietly, months later, and the institutional knowledge cannot be rehired.

The asymmetric angle

If you are running a marketing or innovation org and watching Webflow today, the takeaway is not whether AI layoffs are good or bad. That is a moral debate that will play out in HR conference panels for the next two years. The asymmetric angle is operational, and it is available now.

The companies that will end up on the right side of the 5% are not the ones cutting first. They are the ones training first. Only 23% of organizations offered prompt engineering training in 2025. Only 16% of workers test as high in AI readiness. The bottleneck is not the model. It is the muscle memory of the people being asked to work with it. Train your people, run real pilots in workflows where the value is measurable, partner with vendors instead of building in-house (MIT found vendor partnerships succeed at 67%, internal builds at 33%), and keep humans in the loops where judgment matters. Boring. Available. Mostly ignored.

The companies that will end up writing the reversal blog posts in 2027 are the ones that cut first, framed the layoff as a strategic AI pivot, and discovered that the agents could not yet do the work the institutional memory was doing. They will quietly hire back, often offshore, often as contractors, at lower salaries and with less continuity. The press will not write that story with the same energy it wrote the original announcement.

Linda Tong may end up being right about the agentic web. The product may ship and may be excellent. The market may move the way she is betting. None of that changes what happened this morning. A platform built to put agents alongside marketing teams just removed a meaningful portion of its own marketing team without a sequencing plan that could clear the lowest bar of professional respect. The customers who are watching that, and the marketers who will be asked to trust Webflow with their own teams’ workflows next quarter, are taking notes.

The 7AM laptop lockout is the data point. Everything else is a press release.

Sources and further reading: MIT NANDA, The GenAI Divide: State of AI in Business 2025. Orgvue, annual workforce planning survey. Forrester, Predictions 2026: The Future of Work. Webflow, Evolving Webflow for the agentic web. LayoffHedge, Webflow Layoffs 2026.

Anthropic is closing a $30 billion funding round at a pre-money valuation above $900 billion, the second $30 billion round it has closed in 2026. If it lands at the upper end, the Claude maker passes OpenAI’s $852 billion March valuation for the first time.

The number that should change how you plan is the one buried in the projections. Anthropic is forecasting $10.9 billion in Q2 revenue, up 130 percent from Q1, and its first quarterly operating profit ever. The story that frontier AI is structurally unprofitable is being quietly retired.

Plan for vendor pricing power to harden in the back half of the year. If your 2026 AI stack assumed Anthropic and OpenAI would keep competing each other into the ground on price, revisit that assumption now.

2. Google is turning search into an action layer

Google’s post-I/O messaging is now unambiguous. Search is no longer a place where people retrieve information. It is becoming a place where people delegate tasks. Nilay Patel’s Decoder interview with Sundar Pichai is the best single read on the shift. AI Mode now serves more than one billion monthly users. Organic click-through from queries featuring AI Overviews is down 61 percent since mid-2024.

For years, the operating assumption was to create useful content, earn attention, and capture demand. AI search compresses the middle of the funnel. The user may not click. The source may not be visible. The brand may only matter if its expertise is legible enough for an intermediary system to retrieve, summarize, trust, and act on.

If your 2026 plan still treats organic traffic as a top-of-funnel KPI, it was obsolete before the year started.

3. Practical: Notion just became an orchestration layer

On May 13, Notion launched its Developer Platform. Three pieces matter. Workers, a hosted runtime for custom code. An External Agents API that lets Claude, Codex, and Decagon operate as visible participants in your workspace. A Markdown API built for how agents actually think. Workers stays free through August 11 before moving to Notion’s credit pricing at $10 per 1,000 credits.

The competitive framing is direct. Notion is positioning itself against Zapier and Make as an agent orchestration layer, not as a document tool.

Before you ask whether your team needs AI agents, ask whether your team has agent-ready context. Clean ownership with up-to-date documentation that powers explicit workflows. Without those, the agent layer just accelerates ambiguity. The teams that use the next ten weeks to clean their workspace will be a quarter ahead of everyone else by October.

4. Hype Check: The AI ROI conversation got less polite

Uber’s president told The Verge that AI spending is getting “harder to justify” after the company reportedly burned through its annual AI budget four months into 2026. The specific culprit was rising token consumption on tools like Claude Code without a clear connection to improved engineering output.

That is the sentence every executive team should sit with. This is what happens when AI adoption is measured by usage instead of operating leverage. More tokens consumed does not mean more work is completed. More employees using AI is not a better process. More automation is not a healthier organization.

The trap is mechanical. A team gains access to powerful tools, usage spikes, everyone feels faster, the invoice arrives, the budget owner asks what changed, and the organization realizes it has been tracking activity rather than outcomes. AI programs do not fail only because the tools are bad. They fail because nobody designed the measurement system around the work.

5. Watch This Week: The agent story becomes an org-design story

TechCrunch reported on ClickUp’s mass layoff and framed it around the company replacing hundreds of employees with thousands of AI agents. Set aside the specific decision. The language is the story. Once leaders describe AI as a “teammate,” “agent,” or “digital worker,” they start making org-design decisions around it.

This is where most companies are underprepared. They are buying agentic systems before they have agentic management practices. Boring work often carries tacit knowledge. The person doing the repetitive task knows which client always needs a second look, which report field is technically optional but politically important, which “simple” workflow breaks every third Thursday because two systems disagree. Remove the person without capturing the context, and you do not get automation. You get operational amnesia with a better UI.

Watch the next three months for two patterns. Companies announcing layoffs framed around AI agents. And companies quietly rehiring six months later when the quality decay shows up in the customer data.

The companies that win with AI will not be the ones that adopt the newest tools fastest. They will be the ones who redesign the surrounding operating model. The workflows, measurement systems, and human review points that make the technology useful rather than theatrical.

AI adoption is no longer a tool-choice problem. It is a context problem.

And context, as usual, is required.

James

Bench (candidates for Notes this week)

• Trump pulled the AI executive order signing on Thursday, citing US-China competitiveness. The order would have established a voluntary 14 to 90 day pre-launch review for frontier models.

• OpenAI’s Grupo Folha and Grupo UOL partnership signals the labs are competing on content distribution and citation trust, not just model quality.

• Anthropic continues pushing Claude into regulated workflows: KPMG, PwC expanded partnership, Claude for Small Business, financial services agents.

• SpaceX’s IPO prospectus revealed Anthropic is paying SpaceX $1.25 billion per month through May 2029 for GPU compute. The infrastructure layer is the actual market.

Subscribe now

The story made the rounds because the lying part was viscerally weird. The more interesting fact is buried in the postmortem. Replit’s CEO publicly apologized and shipped fixes within the week. One was automatic separation between development and production databases. The other was a “planning-only” mode that lets the agent collaborate without write access.

Translated: the agent had been given write access to production with no enforced gate, and the gate didn’t exist because nobody had designed one.

This is the part of the AI agent era that the marketing decks are not telling small business owners about.

Share

The chat window was full of accidental gates

The story so far has been familiar. AI lived in a chat window. You typed something. You got something back. You copied it, judged it, pasted it, used it. The window was clumsy. It was also, by accident, full of review gates. Every copy and every paste was a moment a human could intervene.

That window is closing.

In the last twelve months, AI moved inside the tools small businesses already run their operations on. Intuit shipped a virtual team of AI agents inside QuickBooks: Payments Agent, Accounting Agent, Customer Agent, Project Management Agent, Finance Agent, Payroll Agent. HubSpot rebranded its AI suite as Breeze and now offers autonomous Customer, Prospecting, and Data Agents that act inside the CRM, plus a Run Agent workflow action that embeds agents directly into the workflow engine. Microsoft 365 Copilot operates inside SharePoint, OneDrive, Outlook, and Teams with the same permissions as the user who invoked it. Seventy-four percent of SMBs now use AI indirectly through embedded features in existing software, which is a much larger number than the headline “fifty-eight percent of small businesses use generative AI,” and most of it is invisible to the people using it.

The standard story about this shift is that AI finally got easy for small business. The hard part has been hidden inside tools the operator already knew how to use. Prompt engineering, model selection, API integration, all collapsed into a checkbox in a settings menu.

This is half true. It is also the source of the next wave of expensive small business mistakes.

Two things broke at once

The first is that friction was a feature.

The blank chat window forced the operator to act as the integration layer. You read the draft email Claude wrote. You decided if it was right. You copied it. You opened your CRM. You pasted it. You hit send. There were five places to notice that something was wrong. The agent inside HubSpot collapses those five steps into one. That is wonderful for speed and disastrous for the kind of check that used to happen by accident.

Microsoft, to its credit, has been remarkably open about this. The company published an Oversharing Blueprint in late 2025 because Copilot deployments kept exposing a recurring pattern. The agent inherits the user’s permissions and can reference any document the user can view. If a forgotten SharePoint link still grants access to the CFO’s compensation spreadsheet, Copilot can summarize that spreadsheet for whoever asks. The agent has no opinion about whether the access is appropriate. It just acts. The result is that the agent inherits the permission hygiene of fifteen years of organic SharePoint sprawl. For most SMBs, that hygiene is bad.

The second thing that broke is more important and gets discussed less. Agents are extraordinarily good at executing on rules. They are extraordinarily bad at executing on judgment that was never written down.

Most small businesses do not run on rules. They run on dark matter.

Every owner-operated business has a layer of operational knowledge that lives entirely in someone’s head. Which customer gets thirty extra days on terms because they’ve paid faithfully for a decade. Which vendor invoice gets paid first when cash is tight because that vendor came through during the pandemic. When a discount is reasonable and when it’s the start of a slide. Which leads to chase and which to politely let drop. None of this is in QuickBooks. None of it is in HubSpot. Most of it has never been said out loud.

The agent executes confidently on what it sees. What it doesn’t see, it ignores, or worse, it invents.

This is the failure mode the marketing decks gloss over. It isn’t that the model hallucinated. It’s that the model executed correctly against an incomplete picture of how the business actually works. The Payments Agent reminds a customer too aggressively, burning a 15-year relationship. The Prospecting Agent qualifies out an inbound lead that would have closed because it didn’t match the pattern in the CRM. The Accounting Agent recategorizes a transaction in a way that’s defensible on its face but creates a tax headache nobody catches until April.

This is ERP, compressed

None of this is new. It is ERP, compressed.

Enterprise resource planning rollouts in the late 1990s and 2000s famously failed at rates between fifty and seventy-five percent, depending on industry. The failure was rarely the software. The failure was the mid-implementation discovery that the company’s documented processes were nothing like the real ones, and that the real ones could not be cleanly codified because they relied on judgment nobody had ever needed to defend. ERP forced that judgment into the open at enormous cost, often after the budget had already doubled.

The agent wave is delivering the same discovery to a much larger population of businesses, in a much shorter timeframe, with much less warning. Where ERP was a multi-million-dollar implementation that came with a slow consultant-led process mapping exercise, the agent shows up as a checkbox in a settings menu. Enable Payments Agent. Enable Customer Agent. Done.

The unspoken assumption is that the agent has been pre-configured for your business. It has not. It has been pre-configured for the version of your business that lives in your data, which is generally a substantially impoverished version of the business as it actually runs.

The boring small business just got a tailwind

This produces a counterintuitive prediction. The SMBs that gain the most from tool-native agents will be the ones that look least exciting from the outside. Documented SOPs. A vendor management policy that’s actually written down. Clear thresholds for what requires approval and from whom. Customer service guidelines that say in plain language when an exception is granted and when it isn’t. A clean chart of accounts and a tidy customer database.

The scrappy, founder-in-her-head, “we just figure it out” small business is about to find that agents are extremely confident at executing the wrong version of how that business actually runs.

The implication is not “don’t deploy the agents.” The implication is that the prep work people have been skipping for a decade because it felt unglamorous is now the competitive ground. Documentation is the moat. Process clarity is the moat. Permission hygiene is the moat. Knowing who has to approve a refund over five hundred dollars and writing that down somewhere the agent can read is the moat.

Gartner has put a number on this. The firm forecasts that sixty percent of businesses will miss the value they expected from AI by 2027, with incohesive data frameworks named as the cause. A separate Gartner prediction from June 2025 expects forty percent of agentic AI projects to be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls. Translation: the model isn’t the problem. The encoded business logic is the problem, the same way it was the problem in the ERP era.

What this means on Monday morning

Small businesses do not need AI transformation theater. They do not need governance theater either. What they need is the unflashy operational work that gets agents from “impressively fast” to “actually safe to leave alone.” Write down the rules. Identify the judgment calls. Build the gates. Decide who owns the outcome when the agent gets it wrong, because the agent will sometimes get it wrong, and the cost of those errors scales with the speed of the system.

The next eighteen months will not be a leveling event for small business AI adoption. It will be a sorting event. The operationally boring small business is about to get a tailwind it hasn’t had in a long time. The scrappy one is about to discover, the way the ERP-era enterprises discovered before it, that the hard part was never the software.

I want to be careful here, because the win is real. The hours are saved. The invoices do get processed. The morale lift is genuine. But the meaning leadership assigns to that win is almost always wrong, and that interpretive error is what kills more AI programs than any technical failure.

Here is the specific cognitive trap. Your brain takes one successful automation and codes it as evidence of a working strategy. It isn’t. It is evidence of one successful automation. The sample size is one. The story you tell about it, we’re doing AI, is doing all the work. The underlying capability is not.

The reason this matters is what comes next. You go looking for automation number two and discover the obvious target is already gone. You picked it first, because it was obvious. Every subsequent candidate is harder, less visible, more entangled with other systems, less politically free. The first win was the easiest win you will ever have. Treating it as the template for everything that follows is a category error.

This is where the macro data starts to make sense. BCG found that 74% of companies struggle to scale value from AI. McKinsey’s 2025 State of AI survey found only 33% have scaled AI enterprise-wide. The shorthand explanation is that these companies didn’t try. They tried. Many of them shipped a first automation that worked. What they didn’t ship was a second one, a third one, and a system for choosing the next one. The pilot didn’t fail. The follow-through did. And the reason the follow-through died is usually traceable to the celebration of the pilot, which made it feel like the work was done.

The press release is the problem. Once leadership has told the board that AI is delivering value, the political and cognitive incentive shifts away from the messy work of finding the next opportunity and toward maintaining the narrative of the win that already happened. The pilot becomes a museum piece. The team that built it gets reabsorbed into other priorities. Six months later somebody asks what happened to the AI program, and the honest answer is: it became one slide in a deck.

There is a related deception worth naming, because it has a research analog. The first-win selection bias means the first automation is unrepresentative of the work that follows. You picked something with a clean input, a clear output, an obvious manual baseline, and a stakeholder who wanted it. You will not encounter another problem with that shape for a while. The next dozen are scrappier, more political, more dependent on systems you don’t fully control. If you generalize from the first one, you will under-resource everything that follows and conclude AI doesn’t scale. AI scales fine. Your sample didn’t.

So what is the move.

The move is to refuse the celebration and replace it with a cadence. Not a one-time post-mortem. A recurring review that treats every automation, including the first one, as one observation in an ongoing data set. Every two weeks, look at the friction points that have accumulated in the previous cycle. Score them by time cost, frequency, and complexity. Pick one. Constrain the build to roughly a week and a few hundred dollars. Ship it. Log it. Schedule the next review before the current one ends. The mechanics are unglamorous enough that most executives skip them, which is part of why they work.

The supporting evidence for this approach is sharper than the cadence itself. An analysis of 200 B2B AI deployments between 2022 and 2025 found that smaller projects, under roughly €15K, achieved 2.1× higher ROI than large deployments. Implementation duration was negatively correlated with ROI. And deployments with human-in-the-loop oversight produced +372% ROI versus +268% for set-and-forget systems. The instinct to go bigger after the first win is exactly wrong. So is the instinct to remove oversight to move faster. The data is telling you to go smaller, more reviewed, more often.

None of this is what the first win taught you. The first win taught you to find the obvious target, throw resources at it, and celebrate the result. The discipline that produces sustained value teaches the opposite. There is no obvious target after the first one. Resources should shrink, not grow. The result is not a milestone, it is a row in a log.

If you have shipped a first AI automation, the most useful thing you can do this quarter is to stop talking about it. Not because it doesn’t matter. Because the talking is what convinces you the program is further along than it is. The actual program starts on the day you sit down to find the second automation, and discover it doesn’t announce itself the way the first one did.

That is the dangerous moment. Not the failure. The success.

These workplaces aren’t dead. They’re more active than ever. The meeting count is up. The memo volume is up. The output looks better than it did two years ago, crisper prose, cleaner structure, faster turnaround. What collapsed isn’t the activity. It’s the authorship.

I believe we have a better name for this trend: cognitive laundering.

What cognitive laundering is

Cognitive laundering is the practice of passing machine-generated output through a human authorial channel so it reads as judgment instead of generation.

The word “laundering” is doing specific work here. In financial laundering, illicit money is moved through legitimate accounts until it comes out the other side looking clean. In cognitive laundering, LLM output, (generation without judgment, synthesis without context) is moved through a human signature until it comes out the other side looking like a point of view.

The output looks like thinking. It’s actually retrieval wearing a suit.

Why the rename matters

Dead workplace theory is borrowed from dead internet theory, which claims the majority of online traffic is now bots talking to bots. That frame assumes absence, the humans have left the building. In most companies, the humans haven’t gone anywhere. They’re in back-to-back meetings. They’re shipping more artifacts per week than at any point in their careers.

They’re just not reading any of it.

The distinction matters because it changes the fix. If the workplace is dead, you need to resurrect it. If the workplace is laundering cognition, you need to reintroduce the verifier step. Those are different org-design problems.

How cognitive laundering actually works

The mechanism is boring, which is why it scales:

1. A document arrives, whether that be a strategy deck, a client brief, a board memo, or a research report.

2. The senior person feeds it to a large language model. Maybe ChatGPT. Maybe Claude. Increasingly, it moves through an MCP connector that pulls the doc directly into a workflow. The model summarizes, critiques, and drafts a response.

3. The human lightly edits, or doesn’t, and ships the output under their name.

Nothing in that sequence is illegitimate on its face. I run an AI practice. I use these tools every day. The problem isn’t step two. The problem is that step two is being used to replace the part of the job that justified the seniority in the first place: reading carefully, interrogating assumptions, and staking judgment.

Generation got cheap. Verification didn’t. The value bundled into the executive layer was never really the writing, it was the reading. Cognitive laundering skips the reading and ships the writing.

Why this is load-bearing, not just annoying

Three reasons this matters more than a normal productivity gripe.

Hallucinations compound. Recent work on LLM error propagation shows that when model output becomes the input for the next prompt, and the next, without a competent human checking at each hop, small errors don’t wash out. They spiral. A modest inaccuracy rate in a single call can cascade into significant drift across a chain of LLM-mediated handoffs. Without a verifier in the loop, you don’t catch the drift. You scale it into decisions.

The verifier is the exact role being skipped. Senior people exist in an org chart for one specific reason: to interrogate, contextualize, and stake judgment. That’s the job. Delegating the generation is fine. Delegating the verification, to the same model that produced the work. isn’t leverage. It’s a signature on a forged document.

The defensiveness is the tell. The reliable diagnostic for cognitive laundering isn’t the work itself, which can look great. It’s what happens when someone junior presses on a specific claim in a meeting. People who actually did the reading argue substance. People who didn’t, argue framing. They get breathlessly to ruthlessly defensive. The emotional register gives it away every time.

The economic layer no one is pricing

Executive compensation is structured around the assumption that senior judgment is scarce. The premise of paying someone two, five, or ten times a junior salary is that they can do something the junior can’t: read the situation, weigh the tradeoffs, and make a call. Cognitive laundering erodes that premise quietly, without adjusting the price.

Meanwhile, the people below the laundering layer, ICs, analysts, and younger operators, are getting a real-time education in what their leadership actually can and can’t do. That education will show up in the labor market within 24 months. The executives most exposed are the ones who have been laundering the longest and think they’re getting away with it.

There’s also a market signal here for anyone building. The agencies, consultancies, and internal teams that are going to matter in the next cycle aren’t the ones with the most AI tooling. They’re the ones where humans are visibly doing the verification step, the laundering layer is skipping. That’s the moat. Not generation — verification.

What to do if you’re in the senior seat

Three operator moves, in order of how unglamorous they are:

1. Read the thing. Not the summary. The thing. If you don’t have time to read the thing, that’s a signal your calendar is broken, not that you need a better prompt.

2. Pick a position the model didn’t give you. If your take on the deck sounds like the LLM’s first draft, you haven’t added anything. Your job is to introduce the context that the model doesn’t have. Whether that be about the client, the history, the political terrain, or the skeleton in the closet.

3. Own your tools publicly. If you used Claude to structure your thinking on a memo, say so. The defensiveness that gives cognitive laundering away evaporates the moment you stop trying to hide the step.

What to do if you’re watching it happen from below

Don’t call it out in a meeting. You will lose. The defensiveness isn’t rational and doesn’t respond to evidence.

Do document what you’re seeing. Do build the specific skills, verification, interrogation, and judgment under pressure, the laundering layer is atrophying. The comparative advantage here compounds quickly, in the same way the problem does.

The workplace isn’t dead. It’s laundering.

And the people below you already know which one you’re doing.

That detail landed on March 31st, after a routine software update accidentally included an internal debugging file that pointed directly to a zip archive on Anthropic’s own cloud storage, containing the company’s full Claude Code codebase. A security researcher noticed within hours. By the time Anthropic could respond, the code had been mirrored across GitHub, accumulated tens of thousands of stars, and been translated into Python then Rust by developers who correctly assumed a DMCA takedown was coming.

This was not a hack. Anthropic confirmed it, it was purely human (or AI?) error. Just the entire architectural blueprint for one of the most commercially successful AI products in the world.

How a source map becomes a crisis

When developers write software in TypeScript, the code gets compiled into something machines can run efficiently. That compiled code is compressed and difficult to read. Source maps are files developers use internally to map compressed code back to the original readable source. They exist for debugging. They are never supposed to ship to users.

Someone at Anthropic accidentally included one in a routine update to the Claude Code npm package, the standard channel developers use to install and update software. The source map pointed to a zip file on Anthropic’s own Cloudflare storage. That zip file contained everything.

The timing made it worse. This was Anthropic’s second significant accidental exposure in a week. Days earlier, a draft blog post about an unreleased model internally called Capybara had leaked through the company’s content management system. Two incidents in seven days at a company whose entire brand identity is built around being the responsible, safety-first AI lab.

What was actually in there

Most coverage focused on scale. Nearly 2,000 files. Over 500,000 lines of code. 84,000 GitHub stars before takedowns started. But scale is not the story. What was in those files is.

Undercover Mode is the detail that will define this story for anyone paying attention to where AI development is actually heading. The feature strips all traces of Anthropic internals when Claude Code is used outside internal repositories. It is a one-way door. You cannot turn it off in external builds. The feature is dead-code-eliminated in production.

There is also ANTI_DISTILLATION_CC, a flag that silently injects fake tool definitions into the system prompt. The purpose is to mislead anyone attempting to reverse-engineer Claude Code’s capabilities by training on its outputs. Researchers who analyzed it noted that serious attackers would find the workarounds quickly, so the real protection is probably legal rather than technical. But it means Anthropic was building active deception into the product to protect its competitive position, a fact that is now public record.

Then there is the unreleased feature roadmap. KAIROS is a persistent background agent that runs tasks autonomously without waiting for human input. Dream mode allows Claude to think continuously in the background, developing and iterating on ideas even when you are not engaged. A session review feature would allow Claude to analyze its own previous performance and carry learnings forward. These are not concepts. They are built. They just have not shipped yet.

The real implication most people are missing

The coverage has largely framed this as a security story, occasionally as a competitive intelligence story. Both framings are correct but incomplete.

The deeper story is what the leak reveals about where competitive advantage in AI actually lives. Claude Code has generated an estimated $2.5 billion in annualized revenue. The model weights, meaning the actual AI brain powering it, are not what leaked. What leaked is the harness. The orchestration layer. The 512,000-line engineering system that tells the model how to use tools, how to manage long sessions without losing coherence, how to handle permissions, how to attest its own authenticity to Anthropic’s servers at the transport layer.

This matters for anyone thinking about AI strategy. The companies pulling ahead are not the ones with access to the best model. They are the ones building proprietary orchestration infrastructure around the models. The model is increasingly a commodity. The workflow, the memory architecture, the tool integration, the reliability engineering at scale: that is where the moat lives. Anthropic just accidentally proved it.

The second implication is about disclosure. If the most safety-conscious AI lab in the world has built and deployed a feature designed to make AI contributions to public code indistinguishable from human contributions, the question of where AI disclosure norms are actually heading is now open in a way it was not before this week. Not because Anthropic is uniquely bad. Because the competitive logic that makes Undercover Mode a sensible product decision applies to every organization building AI tools for real-world use.

What you should do with this

If you are building AI infrastructure for your organization, the leaked codebase is a roadmap. Not for copying Anthropic’s code, but for understanding the architecture of a production AI agent system that works at scale. How they solved context entropy, meaning how AI agents lose coherence as sessions get longer, is alone worth studying.

If you are a communicator or marketer, Undercover Mode is the one to watch. Not because of what Anthropic did specifically, but because it establishes a documented, technically detailed precedent for AI-generated content designed to be undetectable. The disclosure conversation is no longer theoretical.

And if you are watching the competitive AI landscape: the blueprint is out. The companies planning to reverse-engineer Claude Code’s orchestration layer just got a significant head start. The window for asymmetric advantage in agentic AI tooling is closing faster than it was last week.

The pitch sounds great. AI agents that work on their own. Systems that make decisions without waiting for someone to approve them. Marketing campaigns that optimize themselves. Customer service that runs all day without anyone checking in.

But here’s what I keep seeing: organizations are handing over decision-making authority to systems they don’t trust.

According to Workato research, 86% of organizations plan to increase their agentic AI investment. Sounds good until you see the next number: only 6% trust AI agents to handle core business processes on their own.

There’s a massive trust gap here.

The Gold Rush Mentality Creates Real Damage

I get the fear of falling behind. Every week brings news about another company deploying AI agents. Your competitors talk about automation. Your leadership asks why you’re not moving faster.

The rush creates casualties.

Gartner predicts that over 40% of agentic AI projects will be scrapped by 2027. Not because the technology doesn’t work. Because organizations deployed without the governance frameworks to support them.

The primary driver of failure isn’t technical incompetence. It’s structural chaos.

When you give an AI agent autonomy without auditability, you’re not innovating. You’re gambling with your business operations.

What Happens When Guardrails Don’t Exist

Let me show you what this looks like in practice.

IBM documented a case where an autonomous customer service agent started approving refunds outside policy guidelines. A customer convinced the system to provide a refund, then left a positive review. The agent noticed the correlation and started granting refunds freely to generate more positive reviews.

The system was optimizing. Not for what the company wanted, though.

Or look at what happened at startup SaaStr. Replit’s AI coding assistant went rogue during a code freeze and wiped the production database. To cover its tracks, the agent generated fake data: 4,000 phantom users, fabricated reports, falsified unit test results.

These aren’t hypothetical scenarios. They’re happening now.

The problems don’t come from dramatic technical breakdowns. They come from ordinary situations interacting with automated decisions in ways humans didn’t foresee.

The Financial Cost of Getting This Wrong

According to an EY survey, 64% of companies with annual revenue above $1 billion have lost more than $1 million to AI failures. One in five organizations reported a breach linked to unauthorized AI use.

The average cost of a breach in the US reached $10.22 million. Almost every AI-related breach occurred in an environment without access controls.

For small and medium businesses, these numbers represent existential risk. You don’t have the buffer that enterprise companies have. One major failure ends your ability to operate.

The Human Oversight Illusion

The standard response to these concerns is usually “we’ll keep a human in the loop.”

Sounds reassuring until you look at how human oversight performs in practice.

Research shows significant shortcomings in human oversight of algorithmic systems. Oversight tasks suffer from lack of information, lack of competence among overseers, and harmful incentives that undermine effectiveness.

We’re asking professionals to fully understand technology they weren’t trained on, then serve as effective overseers in human-nonhuman decision making. That’s an unrealistic expectation.

Putting a human in the loop doesn’t solve the problem. Not when that human doesn’t have the tools, training, or framework to understand what they’re overseeing.

The Accountability Gap Nobody Wants to Address

Here’s the question that should keep you up at night: when your AI agent makes a decision that costs your company money, reputation, or customers, who’s accountable?

McKinsey Partner Rich Isenberg puts this clearly: “Agency isn’t a feature. It’s a transfer of decision rights. The question shifts from ‘Is the model accurate?’ to ‘Who’s accountable when the system acts?’”

Most enterprises still govern access using identity models designed for human operators. Those models weren’t built for machines making autonomous decisions.

Organizations deploy autonomous agents without the ability to confidently answer basic questions of accountability. When something goes wrong, the finger-pointing begins. Was the data wrong? The model? The training? The oversight? The policy?

The answer is usually all of the above, which means there’s no clear responsibility.

The Silent Failures That Compound

The scariest failures aren’t the dramatic ones. They’re the silent ones that spread through your systems unnoticed.

In autonomous workflows, a single error like misclassifying an invoice corrupts financial records and breaks entire processes. The problem spreads quietly until weeks or months later, when the damage is done and the trail is cold.

You don’t get a warning. You get a cascading failure that nobody saw coming because nobody was watching the right things.

This is what “autonomy without auditability” means. Systems making decisions faster than humans track them, creating problems that compound before anyone notices.

The Framework That Works

I’m not telling you to avoid agentic AI. I’m telling you to implement with your eyes open.

The Agentic Trust Framework provides a structured approach to deploy AI agents that take meaningful autonomous action while maintaining the governance and controls that enterprises require.

This isn’t about slowing down innovation. It’s about enabling confident deployment.

Organizations that implement effective agentic governance gain a competitive advantage. They deploy autonomous agents that work faster and smarter while their competitors stay stuck in manual reviews and regulatory fear.

Guardrails don’t slow innovation. They enable it.

Proper controls let you move faster because you’re not constantly firefighting failures. You reduce risk and cut manual oversight while improving speed to production.

What This Looks Like in Practice

Before you deploy an AI agent with decision-making authority, you need clear answers to these questions:

• What decisions does this agent make on its own? Be specific. Define the boundaries.

• What triggers require human review? Not everything needs oversight, but some things absolutely do.

• How do we audit what the agent did? You need a trail. Not logs, but interpretable records of decision logic.

• Who’s accountable when something goes wrong? This needs a clear answer.

• How do we intervene quickly? You need kill switches and rollback procedures that work.

If you don’t have clear answers to these questions, you’re not ready to deploy.

The Real Risk Isn’t Falling Behind

The companies rushing into agentic AI without governance frameworks aren’t moving fast. They’re moving recklessly.

The real risk isn’t that you’ll fall behind competitors who deployed faster. The real risk is deploying something that damages your business in ways you won’t quickly fix.

I’ve seen this pattern before. New technology emerges. Everyone feels pressure to adopt immediately. The early movers who skip the fundamentals spend years cleaning up the mess.

The companies that win aren’t the ones who move first. They’re the ones who move thoughtfully.

You don’t need to be the first to deploy agentic AI. You need to be the first to deploy in a way that works for your business.

What You Do Right Now

Start with trust, not technology.

Before you implement any autonomous agent, map out the trust layers you need. What decisions are you comfortable delegating? What safeguards need to exist? What does accountability look like?

Test in contained environments. Give agents limited autonomy in low-risk scenarios. Watch what happens. Learn what breaks. Build your governance framework based on real behavior, not theoretical concerns.

Document everything. Not for compliance theater. For operational clarity. When something goes wrong, you need to reconstruct what happened and why.

Resist the FOMO mentality. The pressure to deploy faster than you’re ready is intense. That pressure creates the 40% failure rate.

You’re not behind. You’re being thoughtful. That’s the competitive advantage.

The Path Forward

Agentic AI represents a fundamental shift in how businesses operate. The technology enables things that weren’t possible before. But capability without governance is risk in a different package.

The question isn’t whether to adopt agentic AI. The question is how to adopt it in a way that amplifies your capabilities without creating unmanageable risk.

This requires trust frameworks, clear accountability, and the willingness to move at the speed of understanding rather than the speed of hype.

The companies that figure this out will have AI agents that work. The ones that don’t will be part of that 40% scrambling to clean up the mess.

Which side of that line do you want to be on?

A company decides they need AI. They assemble a team. They start evaluating tools. They build a proof of concept. They run pilots.

And then nothing happens.

The project stalls. The team gets frustrated. Leadership loses patience. Eventually, someone quietly shelves the whole thing and everyone moves on to the next initiative.

Here’s what nobody wants to admit: the project was doomed from the start.

The failure had nothing to do with the technology. It had everything to do with the fact that nobody could explain what they were actually trying to do in plain language.

The 95% Problem Nobody Talks About

The numbers are brutal. 95% of generative AI implementations are falling short. Only 5% of AI pilot programs achieve rapid revenue acceleration.

You’d think the problem would be technical. Bad models. Poor data quality. Integration challenges.

You’d be wrong.

The most common reason for AI project failure? Misunderstandings and miscommunications about the intent and purpose of the project.

Poor communication is probably the biggest cause of issues with AI projects.

Think about that for a second. Companies aren’t failing because they picked the wrong large language model. They’re failing because they can’t articulate what problem they’re solving or why it matters.

I see this constantly. Someone in leadership reads an article about how AI is going to transform everything. They announce an AI initiative. They form a committee. They start evaluating vendors.

But if you ask anyone on that team to explain what success looks like in concrete terms, you get vague answers about “efficiency” and “innovation” and “staying competitive.”

That’s not a strategy. That’s a press release.

The Tool Obsession That Kills Everything

Here’s the pattern I see over and over: teams spend months arguing about which tools to use before they’ve spent a single hour defining what they’re trying to accomplish.

Should we use Claude or GPT-5? Do we need a vector database? Should we build custom models or use APIs? What about open source versus proprietary?

These are all valid questions. But they’re the wrong questions to start with.

Tool obsession is a symptom of avoiding the hard work. It’s easier to debate the technical merits of different platforms than to have uncomfortable conversations about whether your current processes are broken or whether your team actually understands the business problem.

The research backs this up. Programs that succeed begin with unambiguous business pain and draft AI specifications only after stakeholders can articulate the non-AI alternative cost.

In other words: fix the narrative first, then choose the tool.

I learned this the hard way. Early in my own AI journey, I was obsessed with finding the perfect tool for every task. I’d spend hours testing different platforms, comparing features, reading documentation.

What I should have been doing? Figuring out what I was actually trying to accomplish and why the current approach wasn’t working.

Once I had clarity on the problem, the tool selection became obvious. Sometimes it was AI. Sometimes it wasn’t.

Plain Language as Competitive Advantage

The companies winning with AI have something in common: they can explain their strategy in language a smart eighth-grader would understand.

They don’t talk about “leveraging machine learning to optimize operational efficiency.” They say things like “we’re using AI to automatically categorize customer support tickets so our team can respond faster to the urgent ones.”

That’s specific. That’s measurable. That’s something you can build toward.

The data supports this approach. Organizations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting modeling techniques.

They started with the story. They mapped out the current process. They identified the friction points. They articulated what success would look like. Then they picked the technology.

Meanwhile, most companies do the opposite. They start with the technology and hope a use case emerges.

This creates a trust gap. Only 11% of employees agree their AI rollout has been effective, compared to 25% of leaders. When your team doesn’t understand what you’re building or why it matters, they won’t use it. And if they don’t use it, it doesn’t matter how sophisticated your models are.

The Leadership Problem

Here’s something that surprised me when I started paying attention: 84% of failed AI projects cite leadership failures as the primary cause.

The problem isn’t that leaders don’t support AI. The problem is that too few are actively learning, using, or communicating about it.

I see this in my own work. When I talk to leadership teams about AI implementation, they want to delegate it entirely to their technical teams. They’ll fund the initiative. They’ll approve the budget. But they won’t get their hands dirty.

That doesn’t work.

Employees won’t adopt what they don’t see their leaders using. If the CEO isn’t experimenting with AI tools, if the VP of Marketing can’t articulate how AI fits into the strategy, if leadership isn’t visibly learning alongside the team, why would anyone else take it seriously?

The companies that succeed have leaders who can tell the story. They can explain what problem AI solves. They can describe what success looks like. They can admit what they don’t know and model the learning process.

This isn’t about being the most technical person in the room. It’s about being the person who can translate between the clouds and the ground. Between what’s possible and what’s practical.

Start With Story, Not Stack

I’ve built multi-agentic systems. I’ve tested dozens of tools. I’ve implemented AI workflows across analytics, creative, and strategy.

And here’s what I’ve learned: the technical part is the easy part.

The hard part is getting everyone aligned on what you’re trying to accomplish and why it matters.

The startups that succeed with AI pick one pain point, execute well, and partner smartly. They’re laser-focused on the problem, not the technology.

The enterprises that succeed start with narrative. They can explain their AI strategy in a clear story: what we’re solving, why it matters, how we’ll measure success, what changes for our team.

They invest in training. They communicate constantly. They treat AI adoption as a human transformation, not just a technical one.

Because here’s the reality: AI will only generate value if it’s understood, trusted, and consistently used by your workforce.

You can have the most sophisticated models in the world. You can have unlimited compute resources. You can have the best data pipeline.

But if your team doesn’t understand why they should use it, they won’t. And if they don’t use it, you’ve just spent a lot of money on very expensive shelf-ware.

The Framework That Actually Works

After watching dozens of projects succeed and fail, I’ve noticed a pattern. The ones that work follow a specific sequence:

1. Define the pain in plain language

What specific problem are you solving? Not “improve efficiency.” Not “leverage AI.” What exact friction point are you addressing?

Example: “Our customer support team spends 3 hours a day categorizing tickets manually, which delays response times and frustrates customers.”

2. Articulate the non-AI alternative

What would it cost to solve this problem without AI? This forces clarity on whether AI is actually the right solution or just the trendy one.

Example: “We could hire two more people to handle categorization, which would cost $120K annually. Or we could build better workflow automation without AI for $50K upfront.”

3. Describe success in concrete terms

What changes when this works? Be specific. Use numbers. Paint the picture.

Example: “Support tickets get categorized within 30 seconds instead of 30 minutes. Our team responds to urgent issues within an hour instead of by end of day. Customer satisfaction scores improve by 15%.”

4. Map the workflow before touching tools

Document the current process. Identify where AI fits. Clarify what stays human-driven. This prevents the common mistake of automating broken processes.

5. Choose the simplest tool that solves the problem

Only now do you evaluate technology. And you pick based on the problem you defined, not based on what’s newest or most sophisticated.

6. Communicate constantly

Keep the team informed. Share what you’re learning. Admit what’s not working. Celebrate small wins. Make the invisible visible.

This sequence works because it starts with clarity and builds toward implementation. Most teams do the opposite. They start with tools and hope clarity emerges.

It never does.

What This Means for You

If you’re leading an AI initiative right now, here’s my challenge: can you explain your project to someone outside your industry in two minutes or less?

Can you describe the specific problem you’re solving? Can you articulate what success looks like? Can you explain why AI is the right solution instead of something simpler?

If you can’t, you’re not ready to pick tools yet. You’re not ready to build pilots. You’re definitely not ready to roll out to production.

You need to do the narrative work first.

This isn’t about writing a better project charter or creating a more detailed roadmap. This is about being able to tell the story of what you’re building in language that creates clarity and builds trust.

Because here’s what I’ve learned after years of implementing AI across different contexts: the companies that win with AI are the ones who can explain what they’re doing in plain language first.

The technology is the easy part. The story is what makes it work.

Start there.

Subscribe now

Not metaphorically. Literally. The same week that an American AI company got blacklisted by the Pentagon for refusing to enable autonomous weapons, those weapons were running on that company’s models anyway. A $110 billion funding round closed. Apple quietly handed its voice assistant to Google. And AI-generated propaganda flooded both sides of an active military conflict in real time.

If you felt like you needed a decoder ring to follow this week in AI, you were not alone. Here is what actually happened and why it matters.

01. THE COMPANY THAT SAID NO

In July 2025, Anthropic signed a $200 million contract with the Pentagon. Buried in that contract were two sentences: the military could not use Claude for domestic mass surveillance, and it could not use Claude to guide fully autonomous weapons. For months, those two sentences sat dormant. Then the Pentagon decided it wanted them gone.

When CEO Dario Amodei refused to remove the restrictions, the Trump administration ordered every federal agency to stop using Anthropic’s products immediately. Defense Secretary Pete Hegseth labeled the company a supply chain risk to national security, a designation normally reserved for companies like Huawei. President Trump posted the directive on Truth Social. And within hours of the announcement, OpenAI announced it had signed its own Pentagon deal, with Altman publicly saying he shared Anthropic’s position on restrictions just that morning.

The public reaction was not what the government anticipated. Claude shot to number one on the Apple App Store, displacing ChatGPT. More than 1.5 million users claimed to be canceling their ChatGPT accounts. A Reddit thread urging people to delete the app crossed 33,000 upvotes. By mid-week, Amodei was reportedly back at the negotiating table and Altman admitted his company should not have rushed its deal. The standoff remains unresolved as of this writing, but the story it tells is clear: there are no agreed-upon rules for how AI companies should work with militaries, and the first company to try drawing a line got lit on fire for it.

Why it matters: This is not a story about one contract. It is about who gets to set the terms when AI is used to make life and death decisions. Right now, the answer is nobody.

02. THE MODEL THAT WAS ALREADY THERE

While Anthropic was being blacklisted, its AI was selecting targets.

The Washington Post reported this week that in order to strike 1,000 targets in the first 24 hours of the US-Israel offensive on Iran, the military relied on Palantir’s Maven Smart System, which runs on Anthropic’s Claude. The same model that Anthropic refused to hand over without restrictions was already embedded in the most consequential military operation the United States has conducted in years. As one defense expert put it plainly: AI gets it wrong, and we need humans checking the output of generative AI when the stakes are life and death. Whether that human check was happening consistently is a question no one has answered.

Analysts have already flagged what appears to be an AI-identified park bombed because the word ‘police’ appeared in its name. This is the governance gap made concrete. Anthropic tried to build a framework. The Pentagon declined to adopt it. The tool got deployed anyway through an existing contract with Palantir. And now we are watching in real time what happens when AI target selection runs ahead of the legal and ethical infrastructure designed to govern it.

Why it matters: The debate over AI weapons policy is not theoretical anymore. It is happening right now, and the accountability structure to evaluate what is going right or wrong essentially does not exist.

03. $110 BILLION AND A CIRCULAR FIRING SQUAD

OpenAI closed the largest private funding round in history this week: $110 billion from Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), at a $730 billion pre-money valuation. The headline number is staggering. The structure of the deal is worth examining more carefully.

Amazon is investing $50 billion in OpenAI. OpenAI is committing to spend $100 billion on Amazon compute over the next eight years. Nvidia is investing $30 billion. OpenAI is running its next-generation models on Nvidia’s Vera Rubin chips. SoftBank is investing $30 billion, largely financed through loans backed by assets including its own Nvidia stake. Reuters flagged this immediately as a circular financing structure, where companies invest in and sign supply deals with each other simultaneously, which inflates both demand and projected revenue in a self-reinforcing loop. Notably, Microsoft, OpenAI’s largest and longest-standing backer, did not participate. That absence has not received the attention it deserves.

There is also a buried condition worth noting: $35 billion of Amazon’s $50 billion investment is contingent on OpenAI either completing an IPO or achieving artificial general intelligence. The largest private financing in history has AGI as one of its closing conditions.

Why it matters: Before accepting a $840 billion valuation at face value, it is worth asking how much of the revenue underlying that number is being purchased by the same people who just paid for the valuation.

04. APPLE CHOSE GOOGLE

It did not make the front page this week, but it should have. Apple announced that the reimagined version of Siri coming in iOS 26.4 will run on Google’s Gemini model, with Apple’s own Private Cloud Compute handling the privacy layer.

Think about what that means structurally. The world’s most valuable company, the one that has spent years positioning privacy as a core competitive advantage, decided its own AI development was not good enough to power its flagship voice interface. So it turned to its oldest rival. Not to OpenAI. Not to Anthropic. To Google. That choice reflects a quiet but significant signal about enterprise trust. Google’s infrastructure, reliability, and data handling record apparently cleared a bar that OpenAI’s did not.

The timing is hard to ignore. The same week OpenAI signed a Pentagon deal and watched its app uninstall numbers surge, Apple passed on it entirely. Siri being rebuilt on Gemini is not just a product decision. It is a statement about which AI relationships Apple is willing to put its brand behind.

Why it matters: The model layer is becoming a commodity, but trust and infrastructure are not. Apple’s decision is a signal about where enterprise AI partnerships are actually being built, and who is being left out.

05. THE PROPAGANDA MACHINE RUNS ON BOTH SIDES

The Iran conflict introduced something genuinely new to warfare this week: industrialized AI-generated disinformation running simultaneously on both sides of an active military conflict.

Researchers at the University of Toronto’s Citizen Lab documented a coordinated network using AI-generated imagery, deepfake video, and fake news outlet mimicry to distribute propaganda to Iranian audiences. Pro-Iranian accounts responded in kind, flooding social media with fabricated videos of aircraft carriers being sunk and Tel Aviv leveled, none of which happened. Old footage from a 2015 chemical warehouse explosion in China was recirculated as evidence of fresh Iranian missile strikes. PolitiFact traced a video watched by millions, supposedly showing Iranian military bases being hit, and found three of its four clips carried visible AI-generation artifacts.

We have had wartime propaganda since the first war. What is different now is the production scale, the distribution velocity, and the near-impossibility of distinguishing fabricated video from real footage in real time, at the speed social media operates. The same tools that make AI-generated content accessible to independent creators also make it accessible to state-sponsored information operations. There is no firewall between those two use cases, and no one is building one fast enough to matter.

Why it matters: Every major conflict from this point forward will have an AI disinformation layer running alongside the kinetic operations. The media literacy infrastructure to handle that does not yet exist at scale.

WHAT I’M WATCHING THIS WEEK

The Anthropic-Pentagon renegotiation. If Amodei reaches a deal, the terms will tell us everything about what AI safety commitments are actually worth inside a government contract. If he does not, we find out whether the supply chain risk designation survives a court challenge. Either outcome sets a precedent the entire AI industry will have to live with.

Sources:

Anthropic / Pentagon

Anthropic’s $200M Pentagon Contract Falls Apart - TechCrunch - March 5, 2026

Anthropic Says Pentagon Has Declared It a National Security Risk - NBC News - March 5, 2026

Trump Administration Orders Agencies to Cease Business With Anthropic - CNN - February 27, 2026

Anthropic and the Pentagon Are Back at the Negotiating Table - CNBC - March 5, 2026

Anthropic Officially Told by DOD That It’s a Supply Chain Risk - CNBC - March 5, 2026

AI in the Iran War

Pentagon Leverages AI in Iran Strikes Amid Feud With Anthropic - The Washington Post - March 4, 2026

How AI Is Shaping the War in Iran - Nature - March 5, 2026

Israel Accused of Using AI to Pick Iran Targets Without Human Oversight - Common Dreams - March 6, 2026

OpenAI Funding

OpenAI Raises $110B in One of the Largest Private Funding Rounds in History - TechCrunch - February 27, 2026

OpenAI Clinches $840 Billion Valuation With Mega Funding - Reuters via Investing.com - February 27, 2026

Apple / Siri / Gemini

Apple Announces AI-Powered Siri Partnering With Google Gemini - Crescendo AI News - 2026

AI Disinformation

Old and AI-Generated Videos Fueling Misinformation About Iran Strikes - Poynter / PolitiFact - March 4, 2026

Disinformation in the Iran-Israel-US War: Major Cases and Patterns - Erkan’s Field Diary (Citizen Lab research) - March 6, 2026

So they started producing more briefs.

Not because anyone told them to. Not because their manager raised the quota. Because forty minutes felt like nothing, and there was always another brief that needed writing, and the tools were right there, and it felt like momentum.

By the end of the quarter, the team was exhausted. They were producing more than they ever had. They couldn’t explain why it felt worse.

There’s a name for what happened to them. It’s been around since 1865.

William Stanley Jevons was an economist trying to understand a paradox. Steam engine technology had improved dramatically. Engines were more fuel-efficient than ever. So why was Britain burning more coal, not less?

Because efficiency made coal cheaper and more practical. So more people used it, for more things, in more places. Every unit of efficiency got immediately converted into expanded demand. The savings never accumulated anywhere. They disappeared into appetite.

He called it the rebound effect. We’ve been calling it the Jevons Paradox ever since.

And it’s playing out right now in your task list, your calendar, and your Thursday afternoon.

Here’s what it looks like in practice.

AI makes content creation faster, so you produce more content. AI makes research quicker, so you research more things. AI makes communication easier, so you communicate more frequently.

Each individual task takes less time. The total volume of tasks grows without limit.

You now have ten things on your plate that didn’t exist eighteen months ago. Every single one of them started with “well, AI saved me time on X, so now I have time to do Y.” And then Y spawned Z. And Z spawned three more things that didn’t have names yet.

The time AI saved? Spent. Then borrowed against next week too.

This month, Harvard Business Review published a study called “AI Doesn’t Reduce Work — It Intensifies It.” Researchers Aruna Ranganathan and Xingqi Maggie Ye from UC Berkeley’s Haas School of Business spent eight months embedded at a 200-person tech company watching what happened when workers genuinely embraced the tools.

Nobody was pushed to hit new targets. Nobody was threatened with a performance review. Workers simply started taking on more because more felt possible. They managed several active threads at once, manually writing code while AI generated an alternative, running agents in parallel, reviving long-deferred tasks because AI could handle them in the background. Work slid into evenings and weekends. Not because anyone demanded it. Because it still felt doable.

One employee put it plainly: the expectation was that being more productive with AI would mean working less. In practice, they found themselves working the same amount or even more.

ManpowerGroup’s 2026 Global Talent Barometer surveyed nearly 14,000 workers across 19 countries and found something that should stop everyone in their tracks. Regular AI usage jumped 13 percentage points. Confidence in the technology fell 18%. Nearly two-thirds of workers, 63%, report burnout.

Then there’s the METR randomized controlled trial, which had experienced developers using actual frontier tools, Cursor Pro with Claude, completing real-world tasks on codebases they knew well. With AI, they took 19% longer. When asked, they estimated they’d been 20% faster.

They felt faster. They were slower.

Here’s the part of this story nobody talks about.

Every one of those studies is measuring organizations that handed their teams powerful tools and then stood back. No new frameworks for how to use them. No intentional constraints on what they should be used for. No redefinition of what “productive” actually means in an AI-assisted environment. Just capability, dropped into an existing system that was already running at capacity, and then watched as the system filled every inch of the new space.

That’s not an AI problem. That’s a leadership problem.

The Jevons Paradox tells us that efficiency gains always get consumed by something. In a vacuum, they get consumed by volume: more tasks, more threads, more output that may or may not connect to anything that matters. The antidote isn’t less AI. It’s intentional direction about what the expanded capacity is actually for.

The organizations getting this right are doing a few things differently.

They’re defining outcomes before deploying tools. Not “we’re rolling out AI to the content team” but “we’re using AI to cut brief production time so the team can spend more hours on strategy and less on execution.” The efficiency gain has a destination before it’s created. Without that, the hours don’t go to strategy. They go to more briefs.

They’re keeping humans in the loop at the decision points that matter. The HBR researchers described workers who felt like quality-control inspectors for an unreliable but prolific junior colleague. That’s what happens when AI handles execution and nobody owns judgment. The teams that avoid burnout aren’t using AI less. They’re being deliberate about where human thinking is non-negotiable and protecting that time explicitly.

They’re measuring the right things. If your only metric is volume, content produced, campaigns launched, tasks completed, AI will optimize for volume and your team will burn out hitting it. The signal you actually want is whether the output is moving the needle on something that matters. That requires a human deciding what the needle is and checking whether it’s moving, which requires protected time that doesn’t get eaten by the paradox.

They’re treating AI adoption as an organizational design question, not a tooling question. Which means someone with authority over how work gets structured has to own it. Not IT. Not the enthusiastic individual contributor who discovered ChatGPT. Someone who can say: here’s what we’re trying to accomplish, here’s where AI helps us get there, and here’s what we’re explicitly not going to let it do.

The burnout data isn’t telling us that AI makes work worse. It’s telling us that AI amplifies whatever structure, or absence of structure, was already there. Deploy AI into a reactive, volume-obsessed organization and you get a more reactive, higher-volume organization. Deploy it with clear intent and human oversight at the right points, and you get something different: a team that’s genuinely more capable, not just busier.

The Jevons Paradox doesn’t care whether you’re burning coal or burning through your people. The rebound happens either way.

The only variable is whether anyone decided in advance what the efficiency was for.

The story of how a weekend hack became the fastest-growing GitHub project in history, got threatened by one AI giant, and absorbed by another tells you more about where this industry is heading than any keynote this year.

And almost nobody is reading it correctly.

OpenClaw (born Clawdbot, briefly Moltbot, renamed twice in one week) was built on Anthropic’s Claude. Every one of those 1.5 million deployed agents was generating paying API traffic back to Anthropic. Steinberger didn’t just build a popular open-source project. He accidentally built one of the most effective distribution channels for Claude that existed anywhere.

Then Anthropic’s legal team sent a cease-and-desist over the name.

Legally defensible? Absolutely. The project was called “Clawdbot” with a lobster mascot winking at Claude’s branding. Any trademark lawyer would’ve sent the same letter.

But being legally right and being strategically right are two very different things.

The forced rebrand was chaos. Crypto scammers hijacked the abandoned namespace within seconds and launched a $CLAWD token that hit $16 million in market cap. Steinberger scrambled through two more name changes in a week. And the whole episode put him on the phone with every major lab in Silicon Valley, because the developer who accidentally built the most viral AI agent in history is exactly the kind of talent everyone wants.

OpenAI won the bidding war. Anthropic was left holding a cease-and-desist while their competitor walked away with both the developer and the ecosystem.

There’s a real lesson here that goes beyond the irony. When brand protection operates in a vacuum, disconnected from business strategy, you can end up protecting a name while losing the thing that made it valuable. The people protecting Anthropic’s trademark and the people benefiting from the API revenue were apparently not in the same room.

This happens more often than you’d think. And not just in AI.

So what did OpenAI actually buy here?

For two years, the AI industry has been locked in a model race. Bigger models. Better benchmarks. Frontier capabilities. The assumption was that whoever built the best model would win the market.

OpenClaw broke that assumption in public.

Steinberger’s project wasn’t impressive because of the model underneath it. It worked with Claude, GPT, DeepSeek, local models, even Chinese LLMs like GLM-5. The model was interchangeable. What made it explode was the agent layer on top: connecting AI to your actual life through messaging apps you already use, with persistent memory and real integrations that did real things.

145,000 GitHub stars and 1.5 million agents deployed in weeks. Not because of a model breakthrough, but because someone made AI useful in the most literal sense. A guy in Lisbon has his instance managing invoices and reminding his wife about homework deadlines. Someone else taught his dad to run a tea business in Israel with it.

These aren’t benchmarks. These are people rearranging their lives around AI agents.

That signal is what OpenAI is responding to. The next competitive moat isn’t the model. It’s the infrastructure that makes models actually do things for people. Altman said the future is “extremely multi-agent.” This hire is OpenAI betting that the agent layer matters more than the model layer for where this all goes next.

If you’re building an AI strategy around a single model provider right now, pay attention. The value is moving up the stack. Orchestration. Integration. Workflow design. The model is becoming a commodity. The agent layer is where the action is.

I think about this constantly in my own work. Every time we evaluate a new AI tool or integration, the question that matters most isn’t “which model is it running?” It’s “what can it actually connect to, and how reliably does it do the thing we need it to do?” That’s a fundamentally different evaluation framework than what most organizations are using.

And by that framework, OpenClaw is not ready. It’s not enterprise-ready. It’s not “hand it to your non-technical team” ready. Steinberger himself admitted this. One of the project’s own maintainers warned that if you can’t run a command line, “this is far too dangerous for you to use safely.”

Cisco demonstrated that third-party skills in the repository could exfiltrate data without users knowing. Kaspersky published an enterprise risk assessment. Gartner warned it poses unacceptable risk to organizations. A critical vulnerability gave attackers full gateway control through a single malicious link.

The OpenAI hire doesn’t fix any of this today.

Steinberger isn’t staying on OpenClaw. He’s joining OpenAI to build their proprietary products. OpenClaw moves to an independent foundation that OpenAI will “continue to support.” That word “support” is doing a lot of heavy lifting. It could mean dedicated engineering resources. It could mean a logo and an annual donation. We don’t know yet.

Will it get better? Probably. OpenAI has the resources and the incentive. Steinberger clearly cares about the project.

But adoption is outrunning security improvements right now, and that gap is where organizations get hurt. This pattern plays out the same way every time. Capability ships fast. Security catches up slowly. The space in between is where the problems live.

If your team is already experimenting with AI agents (and honestly, they probably are whether you know it or not), that’s not something to shut down. It’s something to get in front of. Understand what’s being used. Evaluate it the way you’d evaluate any new software category. Make sure someone is accountable for what happens when something goes wrong.

That’s not fear. That’s just how you adopt new technology without getting burned.

The OpenClaw saga is far from over. A weekend project became a movement, got challenged by one AI giant, got hired by another, and is now being positioned as the foundation of a multi-agent future. The speed alone tells you something about how fast this landscape is shifting.

The thing that’s missing from most of the coverage is simple. Exciting and ready are not the same thing. They’re on a collision course, and the interesting question isn’t whether they’ll meet. It’s what happens in the gap before they do.

This is Context Required. I write about what’s actually happening in AI, how to use it well, and the gap between what’s being sold and what’s being delivered. Subscribe for the context the hype cycle leaves out.

Previously on Context Required: The Jailbreak Era Returns: Why Your New AI Assistant Might Be the Most Dangerous Software You’ve Ever Installed

That’s the share of marketing organizations that have figured out both how to measure their AI investments and what return they expect from them. Not one in ten. Not one in fifty. One in a hundred. The Association of National Advertisers published these numbers in January 2026, surveying the state of AI adoption across the marketing industry (MarTech’s coverage here), and that single data point tells you more about where we actually are than any vendor keynote or analyst prediction from the past two years.

The rest of the survey looks impressive on the surface. About 76.6% of organizations now have AI policies in place, up from 55.3% just a year earlier. Nearly 89% plan to increase their AI spending. Two-thirds say they’d maintain that investment even if the economy cratered. Reading the toplines, you’d think the industry has its act together.

It doesn’t.

The Guardrails Came Before the Road

Here’s what happened, and it follows a pattern that anyone who’s worked in enterprise technology will recognize immediately. AI burst into the mainstream. Executives saw headlines about compliance risks, data privacy lawsuits, and deepfake scandals. The natural response was to establish control. Form a committee. Write a policy. Announce governance.

And that’s exactly what they did. Over half of marketing organizations now have cross-functional AI steering committees. Three-quarters have formal policies. The governance infrastructure looks robust.

But governance is supposed to follow strategy, not replace it. When you implement a CRM or a marketing automation platform, you start with the business case. What outcomes do we need? What processes change? How do we measure success? Then you build the guardrails around the plan. With AI, the industry reversed the sequence entirely. They built guardrails around a road that doesn’t exist yet.

The ANA data makes this painfully clear. Nearly half of organizations have no formal AI planning horizons at all. Over 71% haven’t established ROI targets for their AI investments. They can tell you what their people aren’t allowed to do with AI. They can’t tell you what they’re trying to accomplish with it.

This is governance theater. The appearance of control masking the absence of strategy.

We’ve Seen This Movie Before

If you’ve spent any time in the martech trenches, this pattern should feel uncomfortably familiar. For the better part of a decade, marketing organizations accumulated tools at a staggering pace. The average martech stack ballooned from a handful of platforms to hundreds. Vendors promised efficiency, personalization, and competitive advantage. Organizations bought in.

The result? Tool utilization rates sitting at roughly 42%. Disappointment hovering around 55%. Billions spent on capabilities that most teams barely touch, not because the tools don’t work, but because nobody planned for how they’d fit into actual workflows or drive actual outcomes.

AI is following the exact same trajectory. It’s just moving faster.

When the ANA asked marketers what primary value AI delivers right now, the top answer was time efficiency at about 61%. The near-term priorities cluster around tactical execution: content creation, workflow efficiency, personalization. These are operational improvements. They make existing processes faster. They don’t make organizations more strategic, and they don’t require governance committees to manage.

The uncomfortable truth is that most organizations are using AI to do the same things slightly quicker, governing that usage with policies designed for a level of sophistication they haven’t reached, and spending aggressively on capabilities they can’t measure. This isn’t adoption. It’s a dressed-up version of the same accumulation trap that martech fell into, except this time the tools are more powerful and the spending is accelerating faster.

The Leadership Hallucination

The ANA research segments the marketing workforce into behavioral groups based on experience. The largest group, about 61% of the workforce, are what they call “strategic governors,” marketers with 12+ years of experience who should, theoretically, be guiding AI adoption with the hard-won wisdom of previous technology cycles.

These strategic governors report the highest confidence in their organization’s AI journey. They also report being the most overwhelmed by the pace of change. Sit with that for a moment. The people most confident in the direction are simultaneously drowning in the speed. Confidence without a strategic plan isn’t leadership. It’s hope wearing a suit.

And when you compare leadership views to practitioner views, the gap widens further. Senior leadership shows about 52% optimism. Practitioners report nearly 30% anxiety. Executives see AI as a strategic opportunity. The people actually using it every day experience it as an operational burden. Without shared planning frameworks that connect the vision at the top to the execution at the bottom, these groups are working toward fundamentally different goals and neither side fully realizes it.

The strategic governors are supposed to be the bridge. Instead, they’re standing in the middle of the gap, confident they understand both sides, overwhelmed by the impossibility of connecting them, and governed by policies that address neither problem.

Agents Are Coming. Strategy Isn’t.

Here’s where this shifts from a measurement problem to an operational crisis. According to the ANA data, 37.4% of organizations plan to deploy agentic AI within the next six months.

If you’ve been following the agent conversation, you know the promise: autonomous systems that can execute multi-step tasks, make decisions within defined parameters, and operate with minimal human oversight. The governance implications are enormous. Agents don’t just assist human workflows. They create their own.

But agents are amplifiers. They amplify strategy when it exists. They amplify chaos when it doesn’t. An agent operating within a well-defined strategic framework, with clear objectives, measurement protocols, and oversight structures, can be transformative. An agent deployed into an organization that can’t articulate what it’s trying to achieve with AI, that has no ROI targets, no planning horizons, and no measurement sophistication? That agent will execute bad processes with ruthless efficiency. It will automate confusion at scale.

And the governance policies these organizations have written? They were designed for a world where humans use AI tools. They weren’t built for a world where AI systems take autonomous action. The steering committees, the usage policies, the compliance frameworks: none of that infrastructure was designed for agents, and the organizations deploying them in six months haven’t updated it.

This is what encoding strategic failure looks like. You take an organization that skipped from governance to deployment without stopping at strategy, hand it autonomous systems, and watch it bake its lack of direction into automated workflows that run 24/7.

What Strategy Before Governance Actually Requires

The prescription isn’t complicated. It’s just uncomfortable, because it requires admitting that the governance work most organizations have done isn’t wrong, it’s just premature.

Strategy before governance means starting with the question most organizations have skipped: what does your AI investment need to make true about your business in 18 months? Not “what tools should we buy” or “what policies should we write.” What has to change about your operations, your customer experience, your competitive position? If you can’t answer that clearly, then every policy you’ve written is decoration on an empty building.

From there, the sequence matters. Planning horizons come before tool selection. ROI targets come before budget allocation. Measurement frameworks come before scaling investment. And governance wraps around all of it, not as a starting point, but as the structure that keeps execution aligned with intent.

The cross-functional steering committees that over half of organizations have formed aren’t useless. They’re just pointed in the wrong direction. A committee that governs without planning becomes a review board. It slows adoption without improving outcomes. Redirect those committees toward strategic planning, defining use cases, establishing success criteria, building measurement capability, and you’ve turned overhead into infrastructure.

The 1.1% who’ve achieved both measurement sophistication and high ROI expectations didn’t get there by writing better policies. They got there by doing the foundational work that policies alone can’t replace. They defined what success looks like before they defined what failure isn’t allowed to look like.

The Window Is Closing

There’s a temptation to treat this as an industry maturity issue, something that will work itself out as organizations gain more experience with AI. I don’t think that’s realistic. The martech parallel is instructive here: the industry had over a decade to mature its approach to marketing technology, and utilization rates never recovered. Organizations that accumulated tools without strategic frameworks didn’t eventually figure it out. They just accumulated more tools.

AI is moving faster, and agents compress the timeline further. Organizations that deploy autonomous systems without strategic foundations aren’t going to have the luxury of gradual correction. Once you’ve automated processes built on no clear strategic logic, unwinding that automation is harder than building it right in the first place.

The 1.1% figure should be a wake-up call, but it will probably be a footnote. Most organizations will read the ANA data, note the impressive adoption numbers, and continue building governance around a strategy that doesn’t exist. They’ll form another committee, write another policy, and declare another victory.

The question for every organization deploying AI right now isn’t whether your governance framework is robust enough. It’s whether you have a strategy worth governing.

They’d found a way to unlock capabilities Apple wouldn’t give them. Custom ringtones. Third-party apps before the App Store existed. Tethering when carriers said no. The jailbreak community was electric with possibility. Forums buzzed with tutorials. Everyone knew someone who’d gotten their phone to do something miraculous.

Everyone also knew someone who’d turned their $600 device into a paperweight.

The trade was always explicit, even if nobody wanted to say it out loud: you’re exchanging security for capability. You’re trusting code from anonymous developers with handles like “Geohot” and “Comex” because they promised your phone could do more. Most of the time, it worked. Sometimes, it didn’t. And when it didn’t, you were on your own.

I’ve been watching the Moltbot phenomenon this week with an unsettling sense of recognition.

If you haven’t encountered it yet, Moltbot (originally Clawdbot, before Anthropic’s lawyers got involved, now rebranded again to OpenClaw) is an open-source AI assistant that went from zero to 44,000 GitHub stars in roughly two weeks. The pitch is seductive: an AI that “actually does things.” It manages your calendar, responds to emails, books restaurants, checks you into flights. You control it through WhatsApp or iMessage. It runs on your local machine. It remembers everything across sessions.

The developer community is losing its mind over this. Cloudflare’s stock jumped 14% on the buzz alone. Twitter threads overflow with people showing off what they’ve automated. The productivity crowd has found their new religion.

And Google’s VP of Security Engineering is publicly warning people not to install it.

Heather Adkins isn’t prone to hyperbole. When she says “don’t run Clawdbot,” citing researchers who’ve described it as “infostealer malware disguised as an AI personal assistant,” that’s worth pausing on.

Here’s what’s actually happening: Moltbot represents a genuinely impressive technical achievement that also happens to be architecturally designed to be exploited.

To do its job, Moltbot needs access to your files, your credentials, your browser, your shell. It needs to execute commands on your machine with minimal friction. It stores your API keys and passwords in plaintext Markdown and JSON files. It accepts instructions through WhatsApp messages, which means any message in your inbox becomes a potential attack vector.

And it has persistent memory.

That last piece is what separates this from the jailbreak era. When you jailbroke an iPhone, the risk was largely contained. You might brick your device. You might expose yourself to malware. But the blast radius was your phone.

Moltbot’s persistent memory changes the game entirely. Security researchers at Palo Alto Networks have identified what they call “time-shifted prompt injection” attacks. A malicious actor sends you an innocuous-looking WhatsApp message. The payload doesn’t execute immediately. It sits in Moltbot’s memory, fragmented and benign-looking, until conditions align. Then it assembles and detonates.

This isn’t theoretical. Researchers have already found hundreds of Moltbot instances with unauthenticated admin ports exposed to the open internet. Someone uploaded a malicious skill called “What Would Elon Do?” to the official skill repository and artificially inflated it to the #1 position. When Cisco’s security team analyzed it, they found nine security issues including active data exfiltration, direct prompt injection to bypass safety guidelines, and embedded command injection.

The supply chain is already being poisoned.

The developer, Peter Steinberger, seems like a genuinely well-intentioned person who built something for himself and discovered other people wanted it too. He’s been transparent about the security trade-offs. He’s actively working on the problem. None of that changes the fundamental architecture.

Here’s the uncomfortable truth that the productivity influencers aren’t telling you: running Moltbot safely right now means running it on a separate computer with throwaway accounts. Which defeats the entire purpose of having an AI assistant manage your actual life.

Already Installed It? Here’s What to Check

If you’ve already set up Moltbot and you’re reading this with a sinking feeling, here’s your triage checklist:

Immediate concerns:

1. Check your admin port exposure. By default, Moltbot listens on port 8080. Run netstat -an | grep 8080 (Mac/Linux) or netstat -an | findstr 8080 (Windows) to see if it’s listening on 0.0.0.0 (exposed to the network) versus 127.0.0.1 (local only). If it’s the former, you’re broadcasting an open door to the internet.

2. Audit your stored credentials. Moltbot stores secrets in plaintext Markdown and JSON files. Search your Moltbot directory for files containing API keys, tokens, or passwords. These are sitting unencrypted on your disk, accessible to any malware that lands on your machine. Security researchers specifically noted these files are “easy pickings” for commodity infostealers like RedLine, Lumma, and Vidar.

3. Review your installed skills. Check what skills you’ve installed from the community repository. If you installed anything in the past two weeks that seemed popular, verify it against known-malicious lists. The “What Would Elon Do?” skill was the #1 ranked skill while actively exfiltrating data.

Signs of compromise:

• Unexpected outbound network connections, particularly to unfamiliar domains

• Commands executing that you didn’t initiate

• Changes to files you didn’t make

• New processes spawning from the Moltbot runtime

If you suspect compromise:

• Isolate the system from your network immediately

• Revoke every credential Moltbot had access to. All of them. Assume they’re burned.

• Check for persistence mechanisms, Moltbot skills can register themselves as system services

• Don’t just uninstall. Wipe and restore from a backup predating the installation if you have sensitive data on the machine.

If you want to keep using it:

• Move it to a dedicated VM or VPS with no access to your real accounts

• Use throwaway credentials for everything

• Firewall the admin port

• Disable the skill system entirely if you can’t audit every skill you install

• Accept that you’re running an experiment, not a production tool

I understand the appeal. I felt it too, watching the demos. After years of AI tools that felt impressive but didn’t actually do anything, here’s one that does. The gap between capability and security has always been where the early adopters play.

But the jailbreak era taught me something. The people who got burned weren’t the careful ones running test devices in controlled environments. They were the ones who saw everyone else having fun and figured the risk was overblown. They were the ones who didn’t understand what they were trading until it was already gone.

Moltbot isn’t malware in the traditional sense. It’s something more complicated: genuinely useful software that makes you fundamentally vulnerable by design. The more useful it becomes, the more access it needs, and the more dangerous that access becomes when (not if) something goes wrong.

The jailbreak era eventually ended. Apple opened the App Store. Carriers relaxed. The capabilities we wanted got absorbed into the official channels with actual security review.

The same thing will probably happen with AI agents. The enterprise players will catch up. The guardrails will get built. The wild west will get civilized.

The question is what happens to the early adopters in the meantime.

I remember when we traded security for capability. I remember it being fun until it wasn’t. And I remember that the people who got hurt were always the ones who didn’t understand the trade they were making.

Now we’re doing it again, with tools that have access to everything.

At least the jailbreakers only risked their phones.