The constraint nobody will name

Twelve reports, same five things. The one constraint they all walk around is the one that matters.

I read twelve reports this week. They arrived the way these things always arrive now, as a numbered list on LinkedIn, with an instruction at the top to save it and send it to your leadership team. McKinsey, Bain, BCG, Deloitte, PwC, Microsoft, and a few others. Every major firm, all aimed at the same subject, which is how to run AI now that the agents have turned up.

Pull them apart and something odd happens. The branding falls away and the arguments converge. Ten of the twelve say the same five things. This is an operating-model shift, not a tooling rollout. Data and governance are the gating constraint. Pick a few workflows and go deep rather than wide. People move from doing the work to supervising it. And most companies are stuck in pilots, with the return on investment still somewhere over the horizon.

That convergence is not a complaint. It is a signal. The consensus has settled, which means the thesis is no longer the differentiator. Everyone owns the thesis now. You can buy it by the dozen.

Which left me with a more useful question. If the arguments are now the same, how do you tell a report with substance from a report with none?

The tell

A report with substance names a constraint, you can fail at.

Bain's piece does this. It talks about identity and access for agents, about prompt injection, about designing systems for failure rather than only for performance. Those are things that can go wrong in a specific, checkable way. Deloitte does it too, with the unglamorous numbers, the finding that only around one organisation in eight has agents genuinely in production rather than in a pilot. You can argue with a number. You can act on it. You cannot do either with a vision.

The reports that read as hype do the opposite. They reframe the firm. They promise net-new advantage and a new agentic operating model and a maturity ladder to climb. None of it is wrong, exactly. It is simply unfalsifiable. You cannot fail at becoming an agentic organisation, because no one can ever point to the morning it did not happen.

So I went looking for the constraint that almost none of the twelve will name plainly. It is not compute, which they cover. It is not data, which they cover at length. It is not even governance, which has quietly become the safe word of the entire genre, the thing you say to sound responsible without committing to anything you could be measured against.

The constraint they walk around is accountability.

Product versus production system

Here is the distinction the whole argument turns on.

A product is something you ship. You build it, you release it, you move on. If it breaks, the user files a ticket. A production system is something you run. It produces units of output, continuously, and someone is answerable for every single one of them.

Most organisations have bought AI as a product. They have not yet built it as a production system. And the seam where that gap shows is right at the top, where the work meets the board, the regulator, and the customer who was harmed. When an agent makes a decision that turns out to be wrong, the question is brutally simple, and most firms cannot answer it. Who owns this output?

I sat in a room not long ago where a capable team had automated a chunk of a regulated process. It worked. It was faster, it was cheaper, and the slides were genuinely good. I asked one question. If this produces a bad result, and it reaches a patient, or an inspector, whose name is on it? There was some shuffling, and then the honest answer, which was that the model did it.

The model is not a name. The model cannot be called in front of a regulator. The model does not sit through the review, does not lose sleep, does not lose its job. When the answer to who owns this is the model, what you actually have is no owner at all.

Why the reports go quiet here

That is the accountability gap. In the production system it is the failure mode that sits above the human operating layer, at the boundary where the work has to be defended to someone with the power to shut it down. It is the seam the consultancy reports tend to walk around rather than through.

They are not being dishonest. They are being commercial. The advice that sells is the advice that adds, a new capability, a new operating model, a new layer of agents. The accountability gap cannot be sold that way, because closing it does not add anything. It takes something away. It removes the comfortable ambiguity about who is answerable when an automated decision goes wrong.

Watch the language and you can see the avoidance. Embed governance. Make oversight everyone's role. Read that last one slowly. When oversight is everyone's role, it is no one's accountability. Shared responsibility is the polite name for an empty chair.

The ground is moving in a way that makes this sharper, not softer. A few days ago a single government directive forced a frontier lab to switch off two of its most capable models overnight, worldwide, three days after launch. Teams that had wired those models into live work woke up to find the substrate simply gone. The lesson there is the same lesson as here. You can outsource the model. You cannot outsource the accountability for what the model did while you were using it. The vendor's problem became your problem the moment your name was on the output.

What closing it actually requires

The fix is not another governance framework. It is older and duller than that, and it is the part nobody puts on a slide.

Every unit of output a system produces needs a named human owner, decided before it produces anything. Not a committee. Not a policy document. A person, who can explain the decision, who saw the evidence behind it, and whose name is attached when it reaches the board or the regulator. That is precisely what a production system has and a product does not. It is the three Ps doing their quiet work. Permission, the licence to deploy in the first place. Practice, the operating procedure that holds when things are busy and tired and going wrong. And proof, the evidence trail you can put on the table afterwards, when someone asks you to account for what you did.

The twelve reports are right that the consensus has arrived, and right that this is an operating-model problem rather than a technology one. They are simply quiet about the part that is hard to sell, which is that running AI as a production system means someone, by name, is now accountable for output they did not personally produce. That is uncomfortable. It is also the entire job.

A product asks who built it. A production system asks who is answerable for it. Until your AI has a clear answer to the second question, you do not have a production system. You have a pilot with good slides.