Do not let AI turn discussion into agreement

“We discussed it” and “we agreed to it” are different records. When AI turns a messy meeting into a neat summary, a manager needs to check that the difference survived. A shorter document is not useful if it makes an unresolved decision look settled.

This matters when a small team uses AI to prepare project updates, management briefs or handover notes. The purpose is to help people act on what happened, not to make the meeting sound more productive. Begin with a draft for review, using only notes your organisation permits the tool to process.

Separate the record from the recommendation

Give the assistant two distinct jobs. First, extract what the source supports: decisions, proposals, objections, unanswered questions and assigned actions. Then, in a separate section, suggest what the team could clarify next. Do not let the suggestion become part of the historical record.

Those labels describe the status of a statement, not its importance. A proposal from the most senior person is still a proposal until the source records a decision. An objection from one person may contain the condition that makes a plan workable. Repeated discussion does not establish agreement.

For each claimed decision, ask for a pointer to the supporting passage: a note number, paragraph or timestamp already present in the source. A pointer helps a person check the draft. It is not proof by itself; the assistant can attach the wrong passage or misunderstand it.

Keep the disagreement that changes the plan

Consider this fictional discussion at a distributor serving Jamaica and Barbados. The team is considering a new customer-order form. The meeting notes contain three entries:

• Note 1: Sales proposes a Monday launch to reduce repeated calls about missing order details.
• Note 2: Operations says some customers order by phone and asks how those orders will be entered. No answer is recorded.
• Note 3: The chair asks operations to check the phone-order process and bring an option to Thursday's review. No launch approval is recorded.

A summary that says “The team agreed to launch the new form on Monday” has changed the decision. A summary that says “Some concerns were raised” is also weak: it removes the practical issue the next meeting needs to resolve.

The last line does not claim that no decision occurred. It describes the limit of the available record. That distinction protects the team from treating incomplete notes as a complete account.

Give the assistant a clear instruction

“Summarise this professionally” says little about what must survive. Try a narrower request:

This is an original example of applying the AI Fluency Framework's Description competency: communicate the required output and behaviour. The framework, by Rick Dakan, Joseph Feller and Anthropic, names Delegation, Description, Discernment and Diligence. In this task, Delegation leaves approval with people; Discernment checks meaning against the notes; Diligence covers appropriate information use, disclosure and responsibility for the record. A well-written prompt does not replace those checks.

Review the verbs before circulating the draft

Scan words such as agreed, approved, rejected, confirmed and committed. Each should have support. Then check the conditions: “if the supplier confirms” must not disappear when a paragraph becomes a bullet. Verify names, dates and who was actually assigned an action.

Ask the responsible participants to confirm the points that affect their work. Do not use an AI summary to declare consensus or to attribute motives. “Operations asked how phone orders would be handled” records an issue. “Operations resisted the change” adds an interpretation the notes do not establish.

The NIST AI RMF Playbook, GOVERN 4.1, encourages critical thinking and the questioning of important AI decisions. Preserving a relevant objection in an AI-assisted record is one practical management application of that principle. NIST does not prescribe this meeting-note template.

Know when the notes are not enough

If speakers are unclear, the transcript is incomplete or two accounts conflict, ask a person to resolve the gap. Do not ask the model to decide whose recollection is correct. In a multilingual meeting, have someone who understands the language and context check wording that changes an obligation, condition or decision.

Keep sensitive personnel, customer and commercial information within approved systems. Follow your organisation's recording, access and retention rules. Formal board minutes, legal matters and regulated decisions may require a different review process; an AI draft is not a substitute for it.

At the next project meeting, record decisions and open questions separately before asking AI to help. Review one proposed decision against its source. If the evidence is missing, leave the question open and name the person who can clarify it. The team needs an accurate starting point for the next action, not a cleaner version of events.