Meeting AI needs evidence before it needs autonomy

A fluent meeting summary is easy to trust and difficult to audit. Useful meeting intelligence must preserve the source, respect the audience and separate a proposed action from authority to perform it.

Keep the transcript as evidence

Normalise speakers and timestamps, but do not replace the authorised source. Decisions, risks and actions should link back to the transcript segments that support them. When the model is uncertain or speakers disagree, the output should show that rather than inventing consensus.

Do not flatten permissions

A global meeting index can disclose content through summaries even when the underlying recording is protected. Retrieval and publishing need to respect the Microsoft 365 identities and permissions of both the source and destination. Derived content requires the same care as raw content.

Separate proposal from authority

The system may propose an action, owner and due date. A named person should approve or correct it before another system changes. The executing identity should have only the required scope, and the result should be verified without duplicating side effects.

Govern the derivatives

Raw transcripts, prompts, summaries and action records may have different retention requirements. Deletion and legal-hold processes need to locate derivatives. The model provider should receive no broader tenant access than the processing task requires.

Measure correction, not applause

Accuracy is better measured by unsupported claims, missed decisions, permission failures and human corrections than by whether a summary sounds polished. The aim is a trustworthy working record, not an impressive paragraph.