Everyone owned AI. No decision did.
Meta description: An AI operating model stands or falls on one named owner per consequential decision — not on a complete diagram of tools. More on his work: his consulting practice.
The company has done what governance consultants advise. There is a head of AI. There is a steering committee. There is a RACI — I have seen it, and it is thorough: every activity mapped, every role assigned, responsible and accountable and consulted and informed, all present and correct. Then I ask the question the RACI cannot answer: a model in production is causing harm, it is Tuesday afternoon, who may stop it? The room goes quiet. The RACI describes ownership of activities. Nobody owns the decision, and on Tuesday afternoon, the decision is the thing.
[AWAITING MIKLOS — scene: a RACI that is complete and still cannot say who may stop a model in production]
What do current titles actually decide?
The titles in these companies are recent and impressive: head of AI, VP of AI strategy, director of AI enablement. I have learned to ask what each title can decide unilaterally. The answers are revealing: budget within a threshold, hiring within a headcount plan, priorities within a portfolio. The decisions that matter in an AI incident — halt, override, accept liability — belong to no title, because they were never assigned. They were assumed: surely the head of AI, or surely the business owner, or surely IT. Assumptions are the gap wearing everyone’s confidence.
Why does a full RACI still leave a gap?
Miklos Roth
The decision I check first is ownership of the consequential choice. AI can do substantial preparatory work — gathering, comparing, drafting, testing variants. But qualified people must interpret results and own consequential business choices. A business owner or specialist develops an integrated understanding through years of work; an AI assistant can provide valuable information about components of that picture, but it does not possess the company’s tacit knowledge, its relationships, its constraints, or accountability for the outcome.
A RACI maps activities, and activities are continuous: develop, deploy, monitor, review. The gap lives in the moments — the discrete decisions that punctuate the activities. A RACI can assign “monitor the model” to a role and still leave unowned the decision that monitoring exists to inform. I have read RACIs where monitoring had an owner, escalation had an owner, communication had an owner, and stopping had none, because stopping is not an activity with a steady state. It is a decision someone must personally take, and the matrix has no cell for personal decisions taken under time pressure.
What hire is used as a substitute for a role?
The false explanation is the missing executive: hire a chief AI officer and ownership will follow. I have watched the hire arrive and the gap persist, because a person was added where a decision right was missing. The new executive inherits the ambiguity — they are accountable in the slide sense and powerless in the Tuesday sense, unable to halt a system owned operationally by a function that reports elsewhere. The hire satisfies the board’s question — who owns AI? — with a name. The name does not come with the rights, and the gap is now harder to see because it has a face.
What pattern marks role fog?
Role fog has a linguistic marker I listen for: verbs without subjects. In meetings about AI risk, things “should be reviewed,” incidents “need to be escalated,” outputs “ought to be validated.” The passive constructions are not sloppy speech. They are accurate speech — the passive voice is the truthful grammar of a decision with no subject. When I hear a leadership team discuss AI accountability for an hour without a single personal verb attached to a named person, I know the RACI on the wall is decoration.
There is a difference between an expert examining a consequential assertion and five people approving the same document without adding independent information. A sign-off is not meaningful merely because a human name appears next to it. The quality of control is more important than the ceremonial number of approvers. If the model cannot tell anyone who decides, the operating model is not finished — no matter how complete the diagram looks.
[AWAITING MIKLOS — scene: second scene]
What question finds the missing owner?
Role conversations in these companies start from the org chart: who should AI report to, which function should lead. These are structural questions, and they produce reorgs. The question that finds the gap is narrower and more uncomfortable: it names a specific decision and asks for a specific person. Asked of the room, it produces the silence I described — and the silence is the finding. Everything after the silence is the beginning of a real role.
The question that finishes or fails the operating model: for each consequential AI-assisted outcome, who — by name — interpreted the result, accepted it, and owns what happens next?
PR excerpt: [PR — awaiting Miklos material; cut 80–120 words from his paragraphs once available.]
