A steering committee I sat in recently opened with two executives disagreeing about whether a workstream was on track. Nothing unusual there — except neither of them was quoting the status report. Each was quoting an AI summary they'd run themselves that morning, against the same program record, and the summaries disagreed. The next twenty minutes weren't spent deciding anything. They were spent litigating which machine had briefed whom.
This is a new failure mode, and it's an ironic one, because it's caused by something we wanted. Making the record queryable was the point. I've argued on this site that AI should kill the status-report tax — the person-week a month teams burn manufacturing slides. And it can. But the report it replaces was doing a second job nobody itemized: it was the one picture everybody looked at. Kill the report without replacing that function and you don't get a faster shared truth. You get no shared truth at all — just a market of private ones.
The scale of this is easy to underestimate. Microsoft and LinkedIn's Work Trend Index found that 75% of knowledge workers already use AI at work, and 78% of those users bring their own tools. On a real program that means your stakeholders aren't even querying the record with the same model, let alone the same question. Every seat at the table has its own analyst now. None of those analysts have compared notes.
Why the same record yields different answers
The disagreement isn't usually a hallucination problem. Both summaries in that steering committee were defensible readings of the same messy record. They differed for ordinary reasons: the questions were different, and a question smuggles in a frame. "What's the status of the integration workstream?" and "Are there risks to the integration go-live?" pull different evidence out of the same corpus. One exec's tool could see the risk register; the other's could only see Jira. One asked yesterday, one asked after this morning's standup sync. Small input differences, compounded by a delivery voice that sounds equally certain either way, produce confident, polished, incompatible briefings.
A human analyst would flag the frame — "depends what you count as the workstream." The model answers the question as asked, for an audience of one. And because each answer is private, the divergence is invisible until it surfaces in the worst possible venue: mid-meeting, in front of the sponsor, with positions already taken.
The point of the map was never the map
The military has a term for what just broke: the common operating picture. One map, one set of graphics, every echelon fighting off the same version of it. What made the COP valuable was never that it was perfectly accurate — it often wasn't. It was that it was common. When the map was wrong, everyone was wrong the same way, corrections propagated to everyone at once, and any two leaders could argue about what to do rather than about what was true.
A program's status report — for all its sins, and I've catalogued them — was its common operating picture. The RAG colors got negotiated, the narrative got polished, but when the CFO and the delivery lead disagreed, they disagreed about the same page. That page also carried something a private AI answer never does: a name at the bottom. Someone stood behind the picture. Nobody signs a chat response. When a decision made on one gets audited later, there's no artifact to point to — just a prompt history nobody kept.
AI didn't fragment your data — it fragmented your audience. The value of a status report was never the information alone. It was that everyone was looking at the same version of it at the same time, with a name signed at the bottom. Automate the drafting all you want. The commonness is the part you have to rebuild on purpose.
Rebuild the single picture on purpose
The answer is not to ban the querying. It's genuinely useful, people will do it anyway, and a prohibition just drives it underground. The answer is to restore a canonical picture and demote everything else to what it actually is: private analysis. On my programs that looks like a small set of rules. There is one program picture — AI-drafted, human-verified, signed by a named owner, published on a fixed cadence — and it is the only artifact that can be cited as fact in a decision forum. Anyone can query anything, but a private AI answer arrives at the meeting as a question — "my summary flagged X, does the signed picture agree?" — not as a competing truth. When an answer does enter a decision, it comes with provenance: what was asked, what the tool could see, and when — the same chain of custody you'd demand for any number in a board pack.
And when two summaries diverge, treat the divergence as free intelligence. Two defensible readings of the record usually means the record is ambiguous at exactly that point — an undefined term, a stale field, two systems disagreeing about the same milestone. That's not an AI dispute to referee. That's a data-discipline defect the machines just found for you, and it goes on the backlog like any other defect.
- Name one canonical program picture — AI-drafted, human-signed, fixed cadence — and make it the only citable source of truth in decision forums.
- Let people query freely, but classify private AI answers as analysis, not status. They arrive at meetings as questions, not competing facts.
- Require provenance when an AI answer enters a decision: what was asked, what the tool could see, and when.
- Treat divergent summaries as a defect report on the record — find the ambiguity that produced them and fix it at the source.
- Keep a name at the bottom. A picture nobody signs is a picture nobody has to defend — and nobody will.
The bottom line
Every program I've rescued had, somewhere near the root cause, a period when leadership stopped sharing a picture of reality. It used to take months of drift to get there. A fleet of private analysts can get you there in a quarter, with better formatting. The fix costs almost nothing: one signed picture, one cadence, one rule about what counts as fact in the room. Machines can draft the map faster than any staff section in history. Making sure everyone is fighting off the same map — that's still, and stubbornly, a command decision.