Agendas and notes: separating fact from interpretation

By the end, you can

• Explain the difference between a fact, an interpretation, a decision and an unknown in a meeting record, and why blurring them creates risk later (PM-3).
• Use AI to draft first-pass meeting notes from an agenda or rough notes, and explain why a named attendee must confirm which statements are facts before the notes are treated as accurate (PM-3).
• Classify a sample set of meeting statements as fact, interpretation, decision or unknown (PM-3).

Before you start

This is Module 4, Lesson 1 of the AI-Powered Project Management course, opening the module's work on meetings and status. It assumes you have completed Modules 1 to 3 — you should be comfortable with a bounded project brief, a reviewed task breakdown and a decision/risk packet — and have used at least one AI chat assistant for drafting text. It does not require project-management certification, software or a specific meeting tool.

An agenda fixes what gets discussed; notes decide what gets remembered

A meeting without an agenda tends to drift: whoever speaks loudest or last shapes what actually gets covered, and something that needed a decision quietly doesn't get one. An agenda is a short list, shared before the meeting, stating what will be discussed and, ideally, what each item needs by the end — a decision, an update, or just awareness. It does not need to be formal. Three lines circulated the day before ("1. Confirm the delayed delivery date. 2. Review the draft budget. 3. Agree who owns the client email.") does the job.

APM's project management glossary defines communication as "the process of exchanging information and confirming there is shared understanding." That second half is easy to skip past: exchanging information is not the same as confirming everyone understood it the same way. An agenda is the setup for that confirmation — it tells everyone what shared understanding the meeting is meant to produce. The notes taken during and after the meeting are what carry that understanding forward to anyone who wasn't in the room.

Four things a meeting record blurs together if you let it

A rough set of meeting notes often reads as one continuous stream: "The vendor's delivery was late. They're clearly struggling with capacity. We're going to add a penalty clause next time. Not sure yet why this one slipped." Written like that, all four sentences look equally solid. They are not the same kind of statement, and a meeting record that doesn't separate them passes on a hidden problem to whoever reads it later.

The risk isn't writing an interpretation down — reading a situation is often exactly what a meeting is for. The risk is writing it down indistinguishably from a fact, so that three weeks later someone repeats "the vendor is struggling with capacity" as a settled fact in a client conversation, when it was one person's guess that nobody ever checked. A meeting record that keeps these four apart lets a reader trust the facts fully, weigh the interpretations for what they are, see clearly what was actually decided, and know what still needs finding out.

• **A fact** is something that happened and can be checked: "the delivery arrived four days after the agreed date."
• **An interpretation** is a judgement someone drew from a fact: "they're clearly struggling with capacity" is one attendee's reading of the late delivery, not something anyone confirmed with the vendor.
• **A decision** is a choice the group actually made and who is accountable for it: "we will add a penalty clause to the next contract, owned by the procurement lead."
• **An unknown** is something nobody in the room actually knew: "we don't yet know why this delivery specifically slipped."

AI can draft the notes; a named attendee confirms what's actually a fact

Given an agenda and a rough set of notes typed during the meeting — or a transcript, where one exists — an AI assistant can draft a clean, readable set of meeting notes quickly: tidy sentences, grouped by agenda item, in place of a scrawled page. That is genuinely useful; turning fragmentary notes into something a colleague can actually read is real work.

It is not the same as knowing what was actually said, meant or agreed. The Open Worldwide Application Security Project's (OWASP) generative-AI security guidance on misinformation warns that "overreliance occurs when users place excessive trust in LLM-generated content, failing to verify its accuracy." A tidied-up draft can smooth an attendee's tentative "I think maybe the vendor is stretched thin" into a flat, confident line — "the vendor is struggling with capacity" — simply because confident prose reads better than hedged prose, not because anything became more certain. It can also blend a decision the group actually made with an option someone merely floated, if the two weren't clearly distinguished out loud. The draft was not in the room; it only has the words it was given, and the US National Institute of Standards and Technology's (NIST) AI Risk Management Framework's description of an AI system as something that "can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions" is exactly why that draft is a starting text, not a verified account of what happened.

Confirming which lines in a drafted note set are actually facts, which are interpretations, which were genuine decisions and which remain unknown — then circulating that record to people who will act on it — is itself the kind of call the OWASP guidance on excessive agency has in mind when it recommends that a human "approve high-impact actions before they are taken." Once meeting notes go out to people who weren't there, they become the record those people work from; getting that record wrong is a real cost, and checking it belongs with a named attendee, not with the draft that first tidied it up.

A separate, practical note worth flagging here: many teams now use tools that record or transcribe a meeting automatically, and hand that transcript to an AI assistant to draft notes from. Before relying on one, there are two separate checks to make, not one. The first is your organisation's general confidentiality policy — the same check that governs what meeting content can be shared with any AI tool at all. The second is specific to recording and transcription: whether you're allowed to record this particular meeting, and whether the people in it need to actively agree first rather than simply be told it's happening. Passing the first check does not settle the second — consent to recording is its own question, and the answer depends on the law that applies to you and the people in the room (e.g., in the UK the ICO publishes guidance on recording conversations and workplace monitoring). This course does not give legal advice, and this lesson does not attempt to settle that question for you: check your organisation's policy and get clear agreement from the people in the meeting before you rely on a recording or transcription tool.

A worked example: Elm & Stone's weekly site meeting

Elm & Stone, a two-person architecture practice, holds a weekly site meeting with a client partway through a kitchen extension. The practice's lead architect types rough notes during the call, then asks an AI assistant to draft a clean set of meeting notes from them and the week's agenda.

The draft reads well: "The steel beam delivery was delayed by five days. This is likely to push the whole build back a week. The client has agreed to a revised completion date. The cause of the delay is being investigated." The architect checks it against what actually happened. The delay itself is confirmed with the supplier — a genuine fact. "Likely to push the whole build back a week" was her own rough estimate said aloud in the meeting, not anything the client or the builder confirmed — she relabels it as her interpretation, not a fact, and flags that the builder still needs to confirm the actual schedule impact. The client didn't in fact agree to a new date; he said he'd need to check his own availability before agreeing to anything — she corrects the draft to record this as still open, not decided. The cause of the delay is a genuine unknown: the supplier hasn't yet said why. She sends the corrected notes to the client and the builder the same day, with the interpretation and the open item both clearly marked as unresolved.

Accessibility notes

This lesson is text-first, with no images, audio, video or downloadable artifacts. The practice exercise's model answer sits behind a native disclosure control that is reachable and operable by keyboard and correctly announced by screen readers. The knowledge check uses native radio-button inputs with a visible question and options, and posts its result to a live status region so assistive technology announces the outcome without a page reload.

Sources and limits

This lesson synthesises the sources below into a practical learning model. It is not a security standard, legal advice or a guarantee that any particular agent design is safe.

1. Project management glossary — APM (Association for Project Management). Defines communication as the process of exchanging information and confirming there is shared understanding.
2. AI Risk Management Framework 1.0 — NIST AI Resource Center. Frames an AI system as an engineered system that generates outputs such as predictions, recommendations or decisions — not a self-directing decision-maker.
3. LLM09:2025 Misinformation — OWASP Gen AI Security Project. Defines overreliance as placing excessive trust in AI-generated content without verifying its accuracy.
4. LLM06:2025 Excessive Agency — OWASP Gen AI Security Project. Recommends human-in-the-loop control requiring a person to approve high-impact actions before an LLM-connected system takes them.