When the Model Recommends It, Who Owns the Decision?

The sentence is becoming familiar in executive committees and boardrooms: the analysis recommends it. Often the analysis is good. The difficulty is rarely the quality of the recommendation. It is what happens to the decision around it, because three things that ought to stay distinct begin to blur.

When those three collapse into one, the recommendation quietly starts doing the work of the decision. Nobody chose that. It simply becomes easier to approve than to examine.

Collective accountability is not anonymous

A board decides collectively and is accountable collectively. That does not remove the need for a name. The board owns the decision it takes; a named executive owns the recommendation brought to it and the execution that follows. Collective accountability is a reason to ask harder questions in the room — not a place where individual responsibility can disappear.

Four moments where judgement stays human

In practice the risk gathers in four places. None of them needs a new committee. Each needs someone to pause long enough to ask a different question.

The first is the framing of the scenario. A model answers the question it was given. If the brief was to reduce cost in the service organisation by twelve per cent, the analysis will find the most efficient route to twelve per cent — not whether the service organisation is where the cost belongs. Before anyone reads the output, someone senior should ask who framed the question, and which options were removed before the analysis began.

The second is testing what was excluded, and what the data carries with it. Historical performance ratings, past appointments and promotion records hold the preferences of whoever made them, and a model trained on them can reproduce those preferences fluently. In succession and search this matters acutely, because the candidate who does not resemble yesterday’s appointments is precisely the one such a system is least equipped to see. The useful question is not whether the model is biased in general, but where this particular dataset could be carrying the past into the decision.

The third is the human consequence. Before approval, name the people who will live with the decision and describe, concretely, what they will experience and when. This is not a sentimental exercise. It often changes the sequence of a restructuring, the timing of an announcement and who hears first — the practical details that decide whether a sound decision is also a survivable one.

The fourth is the named accountable executive. One name per decision, recorded before it is taken rather than discovered after it goes wrong. The NIST AI Risk Management Framework makes the same point in the language of governance: roles and responsibilities for AI-supported work should be documented and clear, not assumed.

Decision hygiene

None of this requires suspicion of the tools. It requires hygiene — small, repeatable habits that make it visible when a recommendation is doing the work a decision should do. The research suggests why habits matter more than good intentions. In a 2025 survey of 319 knowledge workers, greater confidence in AI was associated with less self-reported critical thinking. The study does not establish cause, but the association is worth designing for.

In a separate 2025 experiment, a warning on screen improved how people handled faulty AI output without bringing them back to the level of those who worked unaided, and self-reported familiarity with AI offered no reliable protection. A reminder is not a process. The questions have to be part of how the decision is made.

The Decision Ownership Check

Five questions, asked before signature by whoever holds the decision. Each carries a yellow flag: a sign to slow down and find out before signing, not a veto.

For a well-prepared decision the check takes twenty to thirty minutes. It is a discipline for the room, not a substitute for due diligence, legal review or the risk processes an organisation already has. On a complex or high-stakes decision it is where the questions begin, not where they end.

An illustration

The following case is illustrative, not drawn from a client. A chief information officer brings the executive committee a recommendation from an AI-supported analysis: consolidate three platforms into one, with savings over three years and the risk scored as low. The Decision Ownership Check raises two yellow flags. The analysis was asked to minimise licence cost, not to weigh the dependency on a single supplier. And nobody has spoken to the two teams whose ways of working the change removes.

Neither flag stops the decision. Both change it. The committee asks for a scenario on supplier dependency and a conversation with the two teams before signature, and the decision is recorded in the CIO’s name. The recommendation survives; the decision now has an owner.

If the answers are clear, the recommendation has done its job and the decision can be taken with confidence. If they are not, the gap deserves more attention than the recommendation itself. For a board, the simplest test is this: if the recommendation proves wrong, would anyone in the room still be able to say why they agreed with it?

This note has a companion essay in CURRENT, on why intelligence is not judgement, and judgement is not accountability.