The RACI chart your AI feature actually needs

A delivery lead at a services firm showed me the RACI for an AI project she had inherited. It was thorough in the usual way. Responsible for the model. Accountable for the budget. Consulted on security. Informed at the steering meeting. She had run dozens of these on normal software and they had served her well. Then she asked the question that had been bothering her. When the model gives a customer a confidently wrong answer, which row on this chart owns that? There was no row. There never had been.

That gap is not a detail. On a normal software feature the failure modes are mostly known, so a standard RACI covers them. An AI feature fails in ways the chart was never built to name. It is wrong sometimes and right most of the time, which is a completely different accountability problem than a bug that either exists or does not. If your RACI has no row for the non-deterministic failure, you have a chart for the wrong kind of feature.

The rows a normal RACI leaves out

Roughly 95% of enterprise AI pilots deliver no measurable return, and a recurring reason is that the org never assigned the roles this kind of system actually needs. A model keeps behaving after you ship it, and you have to keep answering for what it does. That creates jobs a standard chart does not list.

Someone has to own the eval set, meaning the definition of good enough and the tests that check it over time. Someone has to own the escalation, meaning what happens the first time the output is wrong in front of a customer. Someone has to own the cost, because inference bills move with usage in a way a fixed feature never does. And someone has to own the decision to keep it running or turn it off. On most charts I see, all four of those rows are blank, and blank rows are where AI features go to die.

The rows to add before you build

Owner of good enough. One person accountable for the eval criteria and for re-checking them as the feature and the world change. Quality with no owner drifts.
Owner of the wrong answer. A named human in the escalation path who answers when the model misleads a real user, decided before launch, not during the incident.
Owner of the running cost. Someone who watches the inference bill against usage and is expected to raise a hand before it triples, not after.
Owner of the kill decision. The person with the authority to say this is not working and we should stop. If no one holds it, a failing feature runs on inertia.

How we approach it at Density Labs

In our AI Opportunity Assessment ($2,500), we treat these four rows as required, not optional. We walk your existing RACI and find the accountability the chart assumes but never assigned: who answers for quality over time, who catches the confident wrong answer, who owns the bill, who can pull the plug. Naming those people on paper costs a conversation. Discovering they do not exist during a live incident costs a lot more.

A RACI that only covers who builds and who pays is a RACI for software that cannot surprise you. AI can. Chart the surprises before you ship them.