An agent only runs the company you already have

MIT FutureTech went through the annual filings of every S&P 500 company and found that only 11% had AI deeply integrated into their business processes in 2025. Up from 5% in 2022. It is moving, but slowly. I have been thinking about why that number is so low, and I do not believe the answer has much to do with the models.

A few of the things I read last week:

• MIT FutureTech built a measure from SEC 10-K filings to separate deep AI adoption from hype and found only 11% of S&P 500 firms had AI deeply integrated into business processes in 2025, up from 5% in 2022, with a J-curve in profitability but no observed difference in capex or productivity. (MIT FutureTech, arXiv)
• A new HBR piece argues firms spent two decades learning to break themselves into modular pieces, can decompose work far more easily than they can recombine it, and should push integration authority closer to the front line instead of leaving it in central governance. (HBR)
• Josh Tyrangiel, who studied real AI deployments for his book, says successful ones start from a defined business problem, not a licence: "the software is really, really good, but it requires a scalpel," and in the cases he studied, clinicians' domain expertise beat technical expertise. (HBR IdeaCast)
• Simon Willison suggests routing lower-stakes work to cheaper models in subagents, but insists that "judgment, review, and synthesis stay with the main loop" — the human keeps the judgement. (Simon Willison)
• Forrester finds most agent projects stuck in "proof-of-concept purgatory" until companies tie agent autonomy to measurable changes in how work gets done, and Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. (Forrester, via The Register)

My reading is that all five point at the same gap, and it is not a technical one.

The agent runs what it finds

An agent does not arrive with a better version of your company. It arrives with your company. It picks up the workflow as it actually is: the approval step nobody can explain any more, the handoff that only works because one colleague quietly chases it every Thursday, the process that exists in three versions depending on who you ask. If the workflow is undocumented, political or half broken, the agent does not repair any of that. It runs it. Faster, at volume, and without the small human corrections that used to hold it together.

I think this is what the proof-of-concept purgatory numbers are describing. Forrester's condition for getting out is worth sitting with: agent autonomy has to be tied to measurable changes in how work gets done. Not model quality. Changes in the work.

When Simon Willison writes that judgement, review and synthesis stay with the main loop, he means it as a way to build software. I read it as a question about organisations too. In most firms deploying agents right now, who is the main loop? Where does judgement about the whole piece of work actually live? I am not sure enough of us can answer.

Pulling work apart is easy, putting it back together is not

The HBR piece on modular firms gave me language for something I see in almost every organisation I work with. We spent twenty years learning to decompose: outsourcing, shared services, specialised roles, systems that each own a slice. What we never built is the reverse skill, recombining those slices into something coherent, and AI punishes exactly that weakness, because agents cut across the slices. What I find hopeful in the article is that it names a job. Push integration authority to the front line, they say. In plain words: someone close to the work has to own putting it back together.

That is also how I hear Tyrangiel. Start from a defined business problem, use the software like a scalpel, and put your best people on it, because in his deployments the clinicians beat the technologists. The people who own the work, then, not the people who own the tool.

There is a fair objection to all of this. Ethan Mollick has argued that waiting until the organisation is redesigned is itself a risk, and that you should get the tools into people's hands now. I agree with him, and I do not think anyone deploying today is being careless. My answer is not to wait. It is to deploy and redesign at the same time, and to name, in writing, the person who owns recombining the work each agent touches. Without that name, deployment runs the old mess faster. With it, every pilot becomes a redesign. It also fits what KPMG found, which I posted about last Wednesday: value shows up where a named person is accountable, 57% against 21% where nobody is.

If you work in L&D, HR, or transformation

If this lands anywhere, it lands with us. We sit closest to how work actually moves between people, and we are rarely in the room when an agent gets deployed. Two things we could do this week. Pick one workflow an agent already touches, or is about to, and ask who owns its redesign — not the tool, the workflow. If the answer is nobody, propose a name, even our own. And take one handoff everyone knows is half broken and redesign it before the next pilot starts, so the next agent has something worth running. Neither needs budget. Both need someone willing to own the recombining.

The provocation

For one AI project in your organisation, ask this week: if this agent works perfectly, who is responsible for changing the work around it? If you cannot get a name by Friday, you have found your constraint, and it was never the model.