The team AI verification flow: three tiers, one owner, zero heroics

The figure was wrong in the client deck, and the retrospective went in circles. The analyst assumed the reviewer would catch it. The reviewer assumed the analyst had checked it. The AI that produced it assumed nothing, because it does not do that. Three careful people, zero checks.

This is the shape of most team AI incidents: not one person being careless, but a check that everyone reasonably believed was someone else's. Individual habits, like the 90-second check, protect individual work. The moment work flows between people, verification needs to be assigned, not assumed.

A flow that demands deep checking of everything dies in its first crunch. Tier the outputs instead.

Tier one, free: brainstorms, internal drafts, summaries for your own reading. No check required, and saying so out loud matters, because a flow with a free tier is a flow people can follow honestly.

Tier two, checked: anything leaving the team or informing a decision. Every consequential claim gets the source check by the person shipping it.

Tier three, double-checked: anything signed, filed, published, or sent to a client. The shipper checks, and a second person verifies the load-bearing claims independently. Tier three should be small; if everything is tier three, nothing is.

The retrospective failure above had a single root cause: verification was a shared value instead of an assigned job. The fix is one sentence in your flow: every deliverable names the person who owns its verification, and that person is whoever ships it, not whoever generated it.

Ownership by the shipper matters because they hold the context: they know which claims are load-bearing, which numbers a client will quote back, which paragraph came from the AI at 23:40. Assigned ownership also removes the social awkwardness of checking a colleague's AI output, because it is no longer an accusation. It is the job.

KPMG's 58% were not defying a policy; most teams have no policy, and where one exists, checking an unanchored AI claim costs minutes of source-hunting per claim. Under deadline, economics beat intentions every time.

So the flow's most important clause is about tooling: consequential AI work happens in tools where claims arrive with their sources attached. When every sentence links to the exact passage behind it, tier two costs seconds per claim and tier three minutes per document, and the flow stops being a tax. This is where Tatsulok slots into a team's stack: a shared library, answers only from your documents and collections, every claim one click from its passage, and an honest gap statement instead of improvisation.

The flow fits on one page and rolls out in one meeting. Agree on the three tiers with concrete examples from your own work. Name the ownership rule: the shipper verifies. Pick where tier two and three work happens, so the checking is cheap. And add the one cultural clause that makes it stick: finding an AI error before it ships is celebrated out loud, because every catch is a client conversation that never happened.

This week's drill for your team, no signup needed: at your next stand-up, ask everyone to classify the last AI output they used into tier one, two, or three, and whether it got the check its tier demanded. The gaps you hear are your rollout agenda.