AI governance committee minutes: what to record

Ask a health system for evidence that its AI governance operates and you will usually be handed the charter. The charter is the easy artifact. It proves that somebody wrote a charter.

The minutes are the hard artifact, and they are the one that carries the claim. A committee is a thing that meets and decides. The record of it meeting and deciding is the only evidence that either happened.

Why minutes carry more weight than the charter

Governance is the first of the five areas the Joint Commission’s Responsible Use of AI in Healthcare certification standards are organised around. The certification asks whether an organisation has a formal governance structure to support and manage the use of health AI across the organisation.

What follows is inference rather than published fact, because the element-level standards are not public. It rests on how the Joint Commission has assessed hospitals for decades: a standard states an expectation, and an assessor asks for dated evidence that the expectation operated. Not that a policy exists. That it ran. On that pattern, the charter answers the structure question and the minutes answer the operation question, and the second is the one that is hard to produce late.

The practical version is simpler. Anyone can adopt a charter in an afternoon. Nobody can produce eighteen months of minutes in one.

What each entry has to contain

Per decision, not per meeting. This is the structural change most committees need to make, because minutes organised by agenda item read as a discussion and minutes organised by decision read as a record.

Seven fields carry a decision entry. The tool, named as it appears in your inventory rather than by its internal nickname. The decision taken, in the committee’s own words, with pilot approval and production approval minuted separately because they are separate decisions. The evidence that was in front of the committee, listed by document rather than summarised, so a later reader can tell what was known at the time. Who was present, which matters because quorum is meaningless if attendance is not recorded against it. Any dissent, named. Conditions attached to the approval, including the monitoring metric and its threshold. And the review date with a named owner.

The test to apply: can somebody reading this entry in twelve months tell why the committee said yes, and what would have changed the answer? If not, it will not answer a question from an outsider either.

Minutes and the decision log are different documents

Minutes are chronological and record deliberation. The decision log is one row per tool and records current state: name, owner, risk class, date approved for pilot, date approved for production, monitoring metric and threshold, last review, status.

Both, or neither works. A reader who wants the current position on one tool should not have to read a year of minutes to assemble it, and a reader who wants to know how a decision was reached should not be looking at a status row.

The decision log ships in the readiness starter pack as part of the evidence register. The seats, the quorum rule and the escalation route the minutes have to reflect are set out in the committee charter guide.

The four ways minutes fail

They record attendance and topics. Present, discussed, noted. This is the most common failure and it produces a document that proves a meeting happened and nothing else. Any entry whose verb is “discussed” is either incomplete or belongs somewhere other than the minutes.

Nothing was ever declined. A committee that has approved every tool put in front of it has been processing rather than governing, and that pattern is visible from outside without anyone having to say it. If the committee genuinely has never declined anything, the finding is about the committee rather than about the minutes, and it is worth acting on before somebody else notices.

Dissent is smoothed out. Minutes that record unanimous agreement on every clinical AI decision describe a committee that is not doing the work. Named dissent is not a liability. It is the clearest available evidence that the assessment was real, and a decision taken over a dissenting safety voice with the reason recorded is a stronger artifact than a clean one.

They were written afterwards. Reconstruction is detectable, and it converts a documentation gap into a credibility problem. A set of minutes with a single creation date, no conditions, no reversals and no dissent reads as what it is.

Retention, versioning and what to redact

Keep minutes at least as long as the tool is live, plus your standard retention for quality committee records, and longer where a safety event or monitoring finding references the decision. The early minutes are the valuable ones because they establish that governance predates the scrutiny, so they are the ones to protect when systems get migrated.

Version deliberately. Draft minutes, approved minutes, and the date of approval at the following meeting. An unapproved draft is not yet a record, and a set of minutes that has never been formally approved is a weaker artifact than one that was approved with corrections noted.

On redaction, the useful distinction is between commercially sensitive detail and the decision itself. Contract pricing can sit in an appendix with restricted circulation. The decision, its basis and its conditions should not, because a minute with the reasoning removed is no longer evidence of reasoning.

Where to start if the record is thin

Do not backfill. Start the record properly at the next meeting, and separately write a single dated memo that states plainly what governance existed before this point and what was decided outside a minuted process. An honest gap statement is a stronger position than a reconstructed history, and it is the approach the readiness guide argues for at length.

Then work the rest of the evidence. The five areas and what each plausibly asks for are in the certification requirements guide, and the free readiness score will name the artifact you are missing in about twenty minutes.

Disclosure

The Healthcare AI Institute is not affiliated with, endorsed by, or accredited by the Joint Commission or CHAI, and nothing here is official guidance from either body. It is an independent working document written by a physician executive. Where the Joint Commission has published something it is quoted; where it has not, this page labels the reasoning as inference rather than filling the gap. This is not legal advice, and your own counsel should read anything you adopt.

Questions people actually ask

What should AI governance committee minutes contain?

Per decision: the tool, the decision taken, the evidence in front of the committee, who was present and who dissented, any conditions attached, the review date and the named owner. Attendance and a list of topics discussed is not enough to show that governance operated.

How long should you keep AI governance committee minutes?

At least as long as the tool is live plus your organisation’s standard retention for quality committee records, and longer if any monitoring or safety event references the decision. The value of the record is that it reaches back before the tool went in, so early minutes are the ones worth protecting.

Are minutes and a decision log the same thing?

No, and keeping both is the point. Minutes are chronological and record deliberation. The decision log is one row per tool and records current state: risk class, approval dates, monitoring metric and threshold, last review, status. A reader looking for the current position should not have to read a year of minutes.

Can you write up AI governance minutes retrospectively before a survey?

You can write a document, but you cannot manufacture a record. Minutes that all share a creation date, carry no dissent and record no declined tool read exactly as what they are. Start the record now rather than reconstructing it later, because its value is a function of how long it has been running.

Should minutes record a decision the committee got wrong?

Yes. A committee that approved a tool, saw the monitoring drift and withdrew it has demonstrated governance working. Editing that sequence out leaves a record of a committee that has never been wrong, which is less credible than the version with the reversal in it.