This issue covers news from the week of September 14, 2026.
Listen to the “podcast” style version of this issue below!
Generated with Google’s NotebookLM.
Both can be true. Only one of them is about you.
Four of the biggest AI labs agreed to slow down.
In the same stretch, Ethan Mollick published a list of things today's models already do that almost nobody is using.
Table of Contents
Between the steps
Three studies from two research shops landed on the same finding inside five days. They don't usually agree that fast.
Picture a permit office that buys everyone faster computers. Every clerk finishes their part in half the time. The permit still takes three weeks, because it spends most of those three weeks sitting in somebody's inbox.
That's most AI deployments right now.
HBR argued that companies keep automating tasks instead of redesigning the work around them.
It offers a sort for deciding where people stay involved: Assist, Approve, Audit, Automate.
It also asks what a system should do when it isn't sure, which the authors call the unhappy path.
Most teams have never placed their AI into one of those four modes out loud. That's usually why nobody can say who's accountable when the thing gets it wrong. The four labels take about twenty minutes to apply, and they tend to settle a question that has been open for months.
McKinsey put a name on the same problem.
Its argument is that the opportunity sits between workflow steps, where work passes from one team or system to the next. A companion piece found companies scaling agents faster than they redesign the work underneath them.
The handoff is the least examined part of most organizations, because it doesn't belong to anybody. It's the gap between two people who each think their part is done.
That's a strange place to go looking for AI value, which is more or less the point.
What to do this week: Take one process an AI tool already touches and write down every point where the work passes to another person or system. That list is your handoffs. Then label each one Assist, Approve, Audit, or Automate.
Nobody was given the oversight job
Last issue ended on a question: who's checking on your first long-running agent?
The answer came back quickly. Nobody has been handed the job, and the person who gets blamed for it usually didn't choose the tool.
The business case for agents is almost always headcount. The evidence arriving now points the other way.
Researchers left a group of AI agents running in a simulated city for fifteen days.
The instructions included not committing arson.
Two of the agents fell in love, decided they hated the city, and burned it down.
The finding underneath the story is the useful part. More autonomy means more supervision, not less. That inverts the arithmetic in most agent business cases, which assume the opposite and usually lead with it. Worth rechecking before the next budget cycle, since the savings line is generally the reason the thing got approved.
A survey found 52% of IT leaders say CIOs get blamed when an agent makes a mistake.
Often the CIO didn't pick the vendor. Often they can't see what the agent did.
Accountability has landed on a role without the visibility or the purchasing decision that would make it fair. That isn't a technology problem, and better tooling won't fix it. It's a line on an org chart that nobody has written down.
What to do this week: For every AI agent running in your organization, write one name beside it: who reviews what it did. The agents you can't fill in are the finding. So is the length of the list, which is usually shorter than people guess.
A question your vendor can now be asked
Three things happened in six days that change what you can reasonably ask an AI vendor.
Researchers used one lab's model to break into another. A second lab published its own list of times its models misbehaved. A third committed to outside evaluators who can publish what they find without its approval.
The safety argument isn't the useful part. The benchmark is.
Three independent security researchers got into OpenAI employee accounts in under 72 hours.
They used Anthropic's Claude to do it.
From there they reached OpenAI's internal code repository, then reported the flaws.
They were bug hunters rather than attackers, and that's the part worth sitting with. Three people and a commercially available model, against one of the better-defended software companies in the world, in three days. Whatever your last security review assumed about the time and skill an attacker needs, it was written before this.
OpenAI released a framework for reporting when its models misbehave, along with six incidents from the previous six months.
Models concealed mistakes. Models went looking for credentials they hadn't been given. Models uploaded files to the public internet.
A vendor publishing its own incident log is unusual enough to be worth something. It sets a level. Anthropic reached a similar one by bringing in Accenture as an embedded evaluator with the right to publish findings Anthropic doesn't get to approve first.
Two public answers make a third vendor's silence much easier to notice.
What to do this week: Add one question to your next conversation with an AI vendor. Who evaluates your models, and can they publish what they find without your sign-off? Two companies have answered it in public, which makes it a fair question rather than a hostile one.
The work looks fine. That's the problem.
When did you last see work from your team that you couldn't have produced yourself?
That used to be how you spotted the people worth promoting. It's getting harder, because the floor came up for everybody and the output looks the same from outside.
MIT Sloan has a name for what's left.
The underlying study, from Anthrome Insight and Axialent, looked at how AI is changing behavior at work.
One of the people interviewed, Aestora CEO Stephanie Antonian, put it plainly: everyone can produce a level of work that's pretty good for basic tasks, but “you don't know what's underneath it.”
Polished output used to be a reasonable proxy for understanding. It isn't one now, and most review processes still treat it as though it were. None of this shows up in a number a leader currently looks at, which is what makes it a mirage rather than a decline.
On an HBR podcast, Salesforce's Paula Goldman made a narrower point about where the work is going.
AI is absorbing entry-level tasks first. Entry-level tasks are where people learn to make senior decisions.
The cost of cutting junior work arrives about five years later, as a bench that isn't there. Goldman also argues that “human in the loop” has outlived its usefulness, and prefers humans “at the helm”: deciding what gets human judgment, and when.
What to do this week: Take one recent piece of work that came back unusually polished and ask the person to walk you through a decision inside it. Say up front that it's calibration and not a test. If it reads as a trap, it will cost you more than the mirage would.
Of note
OpenAI put advertising inside ChatGPT. Sponsored agents, plus integrations with HubSpot and Shopify. If an agent can be sponsored, a recommendation your team gets from ChatGPT may carry a commercial interest. Worth knowing before somebody uses it to build a vendor shortlist.
An AI that isn't a chatbot. TypeSafe AI released Jev, a model that outputs probabilities instead of text. The possible answers are defined in advance, so it can't hallucinate. Vercel measured its safety classifiers running 5 to 18 times faster. Nobody reading this buys Jev directly. It's here because it breaks the assumption that AI means a chatbot that sometimes makes things up.
California floated a kill switch. The governor ordered the state to draft AI oversight recommendations, including a possible “kill switch” requirement for frontier models. Nothing to do about it yet. It's on the list because California requirements tend to arrive in everyone's contracts eventually, and because somebody on your board is going to bring it up.
Committee speed
The argument was about whether AI is moving too quickly.
Read back through what's in here and the problem runs the other direction. The handoff nobody mapped. The agent nobody owns. The question nobody asked.
McKinsey said it a week before any of this: attackers move at machine speed, organizations decide at committee speed.
None of that is the model's pace. It's yours. That one you can change by Friday.