How to Use AI to Prepare for Meetings

TL;DR: The expensive part of a bad meeting happens before it starts: nobody read the thread, nobody remembers what was decided last time, and the first twenty minutes go to reconstruction. AI fixes the input problem cheaply, paste what exists (prior notes, the email chain, the doc) and ask for a one-page briefing, an agenda with time boxes, and the three questions that actually need answering. What it can’t do is know anything you didn’t give it, so the review pass, checking names, numbers, and what’s changed since the sources were written, stays yours.

What AI can and can’t do here

Meeting prep splits into two kinds of work. Synthesis, pulling the relevant facts out of five scattered sources and arranging them by what matters, is mechanical and slow for humans, fast and reliable for a large language model given the sources. Judgment, knowing that the real agenda item is the one nobody put in writing, or that a stakeholder’s position shifted in a hallway conversation yesterday, never entered any document, so no model can produce it.

The practical division: AI drafts the briefing and the agenda from what’s written down; you spend your saved time on the part that isn’t.

Setup

1. Gather the sources before you prompt. For a recurring internal meeting: last meeting’s notes (this is where an AI meeting notes habit pays compound interest), the relevant thread or channel excerpt, and any doc on the table. For an external meeting: the email history, your CRM notes, and, if your assistant has web access, the company’s recent public announcements. The briefing is only as good as this pile.
2. Check what’s allowed in the prompt. Internal strategy docs are usually fine on a business-tier plan; board material, personnel matters, and anything under NDA need a look at your company’s AI policy first. When in doubt, summarize the sensitive doc yourself and paste the summary.
3. Save a briefing prompt. One reusable prompt (template below) that takes your source pile and returns a fixed-format briefing. The fixed format matters: when every briefing has the same five sections, you read it in two minutes instead of re-orienting each time.
4. Ask for the agenda as a decision list, not a topic list. “Discuss Q3 pipeline” produces discussion. “Decide whether to extend the trial by 30 days, owner: Sam, 10 min” produces a decision. Instruct the model to phrase every agenda item as a question to answer or a decision to make, with an owner and a time box.
5. Do the delta check. The model’s briefing reflects the sources’ timestamps, not today. Before the meeting, scan for what’s changed since the newest source, one Slack search or a 30-second question to a colleague. This is the step that catches the deal that closed yesterday.

Example prompt

The briefing prompt worth saving:

“Prepare a meeting briefing from the sources below. Meeting: [purpose, attendees and their roles, duration]. My role in it: [decider / presenter / contributor]. Sources: [paste last meeting’s notes, the thread, the doc, label each one]. Produce, in this order:

1. Situation, 3 sentences max on where this stands now, per the sources.
2. Decisions made previously, with who made them, quoted or paraphrased from the notes.
3. Open items, anything unresolved or assigned but not confirmed done, with owners.
4. Proposed agenda, every item phrased as a decision to make or question to answer, with a suggested owner and time box, fitting the meeting length.
5. Three questions I should be ready to answer, given my role. Use only the sources. If something important is missing or the sources contradict each other, say so explicitly in a ‘Gaps’ line instead of smoothing over it.”

That last instruction does the most work. Default model behavior is to produce a confident, seamless briefing whether or not the sources support one, hallucination in a prep doc means walking into the room sure of a number nobody agreed to. Forcing a “Gaps” line converts silent invention into a visible to-do.

For external meetings, add: “From public sources only: recent announcements, leadership changes, or product launches at [company] in the last 6 months, with dates. Flag anything you’re not certain of.”

Where this pays off most

Recurring meetings with poor institutional memory are the sweet spot: pipeline reviews in sales, sprint and roadmap reviews in product, vendor QBRs in operations, interview debriefs in HR. Anywhere the same group meets weekly and spends the first quarter of each session reconstructing the last one, a two-minute briefing read replaces that entire segment. One-off, high-stakes meetings benefit too, but there the AI briefing is a starting draft for your own thinking, not a substitute for it.

Pitfalls

• Briefing from nothing. “Help me prep for my 2pm” with no sources gets you generic advice shaped like a briefing. The model synthesizes; it doesn’t know your meeting. No sources, no value.
• Trusting the model’s memory of your company. If you didn’t paste it, the model is guessing from patterns, plausible-sounding project status is the most dangerous kind of wrong. Everything factual must trace to a source you provided.
• Skipping the delta check. Sources age. The briefing is a snapshot of the pile’s newest timestamp; the gap between that and the meeting is your job.
• Agenda as topics. Topic lists produce meetings that “covered a lot.” Decision lists produce outcomes. Enforce the phrasing in the prompt and again when you edit.
• Over-prepping the wrong meetings. A 15-minute standup doesn’t need a briefing. Spend this workflow on meetings where being the most prepared person in the room changes the result.

Checklist: before you walk in

• Briefing read; every number and name spot-checked against a source
• Delta check done, nothing major changed since the newest source
• Agenda items phrased as decisions/questions, each with an owner and time box
• You can state the meeting’s one required outcome in a sentence
• The three questions you’re likely to be asked have answers

FAQ

Can AI prepare a meeting agenda automatically? Yes, if you supply the purpose and the source material, notes, thread, doc. The output is a strong draft, and forcing decision-phrased items makes it a genuinely better agenda than most humans write. Without inputs, it’s a template with your meeting name on it.

Is it safe to paste internal documents into an AI assistant for meeting prep? Business/enterprise plans generally exclude inputs from training; check yours, and check your company’s AI policy for the sensitive categories, board material, legal, personnel, customer personal data. Summarize-then-paste is the fallback for borderline docs.

How is AI meeting prep different from AI meeting notes? Notes capture what was said after the fact; prep synthesizes what’s known before. They compound: this meeting’s notes are the next meeting’s best prep input, which is why setting up both closes the loop.

Can AI research the people I’m meeting? Web-connected assistants handle public company research well, announcements, funding, product moves. Verify before you repeat: models confuse similarly named people and report outdated titles with full confidence.

Get one practical AI implementation brief per week, join the free newsletter.