Nadia Osei had spent a week rereading the minutes of a meeting that lasted nine minutes. Five people had reviewed an AI-drafted evaluation, accepted its recommendation, and signed before the end of the day. The record preserved the decision and every approval. Fourteen months later, a regulator asked Solenne Life Sciences why the group had agreed.
“Our record shows that our process ran as designed. It does not show why the review board agreed with the evaluation, because we preserved approvals and did not preserve reasons.”
Solenne is a fictionalized composite in Chapter 10 of AI in the Org Chart, not a report about a specific company or incident. Its regulatory request followed a hospital report about an infusion set. A child had received less medication than the pump display showed. An investigation found that softened tubing could intermittently collapse at low flow. The engineering explanation became clear after the incident. The harder question was why earlier warning signs had been reviewed and left under routine monitoring.
What did the record preserve?
The company had not lost its paperwork. Its quality system tracked complaints, versions, signatures, and who opened which documents. Mira, the company’s AI system, had read complaint narratives, grouped them into categories, compared rates with historical data, and drafted the evaluation for the review board.
The analysis was defensible. Returned samples had passed testing. Complaint rates remained within historical limits after seasonal volume was considered. Reports used different language: slow infusion, alarm event, kinked line. No single category appeared to be rising enough to trigger action.
Alba Reyes, a complaints analyst, saw a pattern in the narratives that the separate categories concealed. Reading the reports in sequence, she noticed pediatric cases, low flow rates, and one pump model. She raised a possible cluster through the triage system. The note followed the workflow upstream. Mira processed the same complaint set in its next evaluation and again reported that the categories remained within band.
No one had formally rejected Alba’s concern. The workflow had absorbed it. At the review meeting, someone apparently raised a similar question, but the minutes recorded only that discussion had been held. The signed record could show who had concurred. It could not show what the analyst noticed, what the board considered, or why the concern did not change its recommendation.
Why wasn’t a complete audit trail enough?
An audit trail can establish that a process happened. It can show who opened a file, which version they saw, and when they approved the result. That evidence matters, particularly in regulated work. But an approval answers a different question from the reason a person agreed.
When counsel interviewed the five signers, one recalled the evaluation’s conclusion but could not say why he had accepted it. He initially remembered clean returned samples, then recognized that he was remembering a different review. The other participants could recall the busy afternoon or the meeting’s pace. None could reconstruct the reasoning behind that specific choice.
Their approval was genuine in the procedural sense. They had attended, read, and signed. Their signatures did not supply a judgment that the record had never asked them to put into words. The missing information was not another timestamp or another copy of the recommendation. It was what a person believed after considering it.
That gap matters wherever a system prepares an evaluation that can shape a consequential decision. A neat record can create the appearance that the human part is settled: the review occurred, the required people agreed, and the file is complete. But later, when a customer, patient, regulator, or board asks why the organization acted, the signatures may carry less than the company expects.
What would make the reasons visible?
Solenne’s response was to change the decision record itself. For decisions whose consequences could reach a patient, customer, or regulator before another review, the chair had to write four short entries before signing: what the group considered, what remained uncertain, who challenged the conclusion and how the challenge was answered, and why the chair agreed.
The limit mattered. The company did not add this record to every routine act. A judgment record on every small choice would become another field people completed by habit. Instead, Solenne reserved it for choices that could travel beyond the room before someone had a chance to look again.
The first month was awkward. Some chairs filled the new field with phrases like “consistent with prior data.” One wrote hundreds of words about a routine labeling change. Meetings ran late without producing better explanations. Nadia narrowed the requirement, and refused to provide a template that would make the words easier to paste.
Then writing the reason began to change what the meeting could see. A chair asked to explain agreement sometimes discovered that the agreement had not yet formed. The question surfaced while the people and information were still present, while the board could examine the analyst’s concern or decide why it did not alter the recommendation.
This is the difference between recording a signature and preserving a Judgment Trail. A signature establishes assent. A Judgment Trail gives the organization a way to reconstruct what informed that assent, including unresolved doubt and the treatment of a challenge. The point is not to make every decision defensible after the fact. It is to give people a reason to examine the choice before it becomes an answer the organization has to defend.
When Elias Brandt chaired a later review, he read the complaint narratives himself and called Alba before the meeting. Her view that the pattern was probably seasonal, with two hospital systems worth watching, went into the record alongside the uncertainty. The review ran late. Writing his reason took six minutes while the next agenda item waited.
The old meeting had fit inside its nine minutes. Its record was complete, right up to the point where the judgment should have been. A signature still ended the later meeting. This time, the file held a sentence its author was willing to explain a year afterward.
This essay draws on Chapter 10, “The Record Showed Approval, Not Judgment,” of AI in the Org Chart.