AI governance has become one of those concepts that sounds unimpeachable in a board deck and falls apart completely in practice. Organizations create Centers of Excellence, convene AI Councils, draft responsible AI charters. It all looks very responsible. And it is — right up until the moment something goes wrong at 2am on a Saturday and there are four people who could theoretically be responsible but none who will actually pick up the phone.
The most dangerous governance structure in AI is the one that distributes responsibility so evenly that it evaporates. Not because anyone intended it, but because committee accountability is an oxymoron.
The accountability vacuum
AI gets assigned to a Center of Excellence or an AI Council that meets monthly, reviews dashboards, and has no actual decision-making authority. When something goes wrong, there are four people who could theoretically be responsible but none who will actually pick up the phone.
A major retailer’s AI pricing engine began recommending prices below cost during a holiday weekend. Engineering assumed business was monitoring. Business assumed engineering had guardrails. The AI team assumed pricing had override authority. Nobody acted for 72 hours. Estimated impact: $2M+ in margin erosion.
Then there’s the kill switch problem. Organizations invest months and millions into AI projects and create institutional momentum that makes it politically impossible to stop a failing initiative. IBM’s Watson for Oncology recommended cancer treatments worldwide. Internal documents revealed it sometimes recommended unsafe treatments, but organizational momentum behind “AI-powered cancer care” made it extraordinarily difficult for clinicians to push back. Accountability deferred to the technology’s reputation rather than clinical outcomes.
What actual accountability looks like
Airbnb assigns a DRI (Directly Responsible Individual) to every AI feature. A single named individual — not a team, not a committee — is personally accountable for each AI deployment’s outcomes. This person has the authority to pause or kill the deployment without committee approval. When pricing suggestions showed signs of discrimination, the DRI had pre-authorized authority to disable the feature within hours, investigate, and re-enable only after the issue was understood and fixed.
Netflix includes pre-registered failure criteria for every model deployment. Before launching an AI initiative, they define the specific conditions under which they will stop, pause, or pivot — documented before institutional momentum builds. If a recommendation model’s metrics drop below threshold for 48 hours, it automatically rolls back. The decision to stop is made before the emotional investment begins.
The diagnostic question
The assessment signal for accountability is deceptively simple: can you name the single individual accountable for AI outcomes? Not the committee. Not the department. The person whose phone rings. If the answer takes more than five seconds, you have a governance document, not governance behavior.
Has your organization ever killed an AI project mid-flight? Are there pre-defined criteria for pausing a deployment? When was the last time someone said “no” to an AI initiative — and what happened to them politically? These questions reveal whether accountability is structural or performative.