The meeting-intelligence policy - governing for candour (2026)
The situation
We record meetings and process the transcripts with AI to extract decisions, actions and context. For an organisation where one person holds context across commercial, legal, technical and product domains at once, the operational gain is substantial. It is also the single feature most likely to change how people behave at work, and the part most AI policies treat as an afterthought.
The read
The failure mode nobody designs for is behavioural. Once people suspect a recording might be replayed against them - not will be, only might be - they stop disagreeing openly, stop thinking aloud, stop raising the half-formed concern that turns out to matter, and start performing for an audience they cannot see. The meeting still happens. The recording still captures it. What it captures is worth less than the candour it cost, and nothing ever appears to go wrong, so the cost never surfaces in a governance review. I treated that as the primary design constraint rather than a cultural side-effect.
What I built
Five structural decisions. The raw record belongs to the room: only attendees may access a recording or transcript, and everyone else - including management and division heads - receives the AI extraction, enforced as a storage permission rather than a courtesy. Humans stop and the AI continues: after month one, participants work from extracted outputs rather than replaying recordings, making the AI the only reader of the raw material, which is defensible precisely because it has no social memory and no career interest in what anyone said. Correction is an active step with a deadline - the extraction is circulated within three working days and after day 31 it becomes the record. Purpose is limited to three things: context for the work, actions within scope, and the decisions reached. Retention is a scheduled, evidenced deletion - one year, or project close - with legal hold as the only extension. Supporting all of it, an AI processing log recording what was extracted, the expunge schedule and every access granted, which doubles as the evidence base for PDPA data-subject requests and client audits.
The judgement call
My first position on HR use was absolute: this material is never used for evaluation, discipline or investigation, no exceptions. On review that was the wrong answer to the right instinct. It promised something the company cannot deliver, since no internal policy survives a court order; it would read from outside as an organisation holding evidence of harassment and declining to look; and it stripped access from the person who most deserves it, because a complainant was in the room and under my own participant rule that is their record of what happened to them. I redrew the line between proactive and responsive rather than between permitted and prohibited. The company never initiates - no monitoring, searching, mining or sampling, no opening a matter from meeting material and no going looking once a matter is open, no access by seniority. What remains is narrow and responsive only: a participant's own access to their own meeting, requested through the DPO rather than a line manager, and lawful compulsion handled by group legal. That preserves the whole behavioural guarantee without the company making a promise it would have to break, and without accidentally shielding the wrong party.
What this shows
Identifying a failure mode that compliance frameworks do not name, and engineering against it structurally rather than with a values statement. It also required holding four things at once - the legal position and its limits, client agreements that already referenced recording in both directions, whether the tooling could actually enforce participant-scoped access, and how people behave when they believe they are being watched. Conduct policy was not mine to write, so I bounded my own instrument and deferred explicitly to the process that owns the question. And I changed my own design when a better argument arrived rather than defending the first position.