Most executive assistants who try AI for meeting follow-ups stop at the obvious win: paste the transcript into ChatGPT, ask for a summary, paste it back into the team channel. It saves time. It also misses the actual job.
The job after a meeting isn’t to summarise what happened. It’s to surface what your principal will need before the next meeting, what got softly committed without anyone noticing, and what the room agreed not to talk about. None of that comes out of a generic “summarise this transcript” prompt. The EAs getting genuine value from AI are running a three-prompt chain that pulls each layer out separately.
Capture the room before the model touches it
Your transcript quality decides everything that follows. The free AI notetakers that ride along on calls have improved enough that the difference between Otter.ai and Fireflies and a built-in option like Microsoft Copilot in Teams is mostly about which calendar you live in. Pick one and use it for every internal meeting your principal attends, even ones they think don’t need notes. The notes you take for yourself are usually more valuable than the ones you take for them.
One specific failure mode to watch for. Generic notetakers will mis-attribute speakers when two people on the same domain talk back to back, especially on a call with two engineers named Mark. Always glance at the attribution column before you feed a transcript into anything else. A wrong attribution becomes a wrong action item, and a wrong action item assigned to the wrong VP is the kind of mistake people remember.
Run three prompts, not one
The summary prompt that gets shared on LinkedIn produces a tidy block of text that nobody reads twice. Replace it with three short prompts that each do one job.
The first prompt gets the explicit decisions. Paste the transcript into ChatGPT or Claude with this:
“You are reading a transcript of an internal meeting. Extract only the explicit decisions made during this meeting, in the form ‘X was decided’ with the named decision-maker and a one-line rationale if it was given. Do not include action items, do not include topics that were discussed but not decided, and do not infer decisions that weren’t actually made. If a decision was deferred, list it under a separate ‘Deferred’ heading. Transcript: [PASTE TRANSCRIPT].”
This is the one you can actually circulate. People will challenge a decisions list if it’s wrong, which is what you want. They will not challenge a wall-of-text summary, which is how mis-recorded decisions calcify into reality.
The second prompt is the one most EAs never write. It pulls out the implicit commitments, the half-promises that weren’t framed as action items but will be remembered as ones:
“Read the same transcript again. List every commitment that was made informally, including phrases like ‘I’ll take a look at that’, ‘we should probably’, ‘leave that with me’, ‘send me something on that’. For each one, identify who made it, what they implicitly agreed to, and the most likely person who will follow up if they don’t. Format as a table with columns: who committed, what they said, what it likely means, who will chase. Transcript: [PASTE TRANSCRIPT].”
You’re not going to circulate this one. It’s for you. It’s the list that tells you what your principal will be asked about in the next meeting that they don’t yet remember promising. A counterintuitive thing happens once you start tracking these: your principal stops being surprised by their own commitments, and your reputation for catching things shifts upward in a way that’s hard to articulate but very obvious in performance reviews.
The third prompt builds the action tracker, but only after you’ve reviewed the first two. Take the cleaned decisions list and the implicit commitments list, paste them in together, and ask:
“Combine the explicit decisions and implicit commitments below into a single action tracker. For each row, give me: task description (rewritten so it’s specific and measurable), owner, suggested deadline based on the urgency cues in the discussion, the meeting date, and a one-sentence note on context that the owner will need to actually do this. Flag any task where the owner was not clearly assigned, do not guess. Output as a markdown table. Decisions: [PASTE]. Commitments: [PASTE].”
The flag for unassigned owners is the part that makes this useful. The flag is what you bring to your principal at the end of the day and ask, “do you want me to chase Sarah on this or does it actually belong to David?” That single question is the difference between an EA who runs a tight ship and one who pushes paper.
Brief your principal before the next meeting, not after this one
Here’s where the chain pays off. Most EAs stop at the action tracker. The ones whose principals say “I don’t know how I’d function without them” take one more step: they generate a forward-looking briefing for the next meeting on the calendar that touches the same project.
Open the action tracker, pull the relevant rows, and run:
“My principal has a meeting tomorrow with [ATTENDEES] on the topic of [TOPIC]. Below are the open items from previous meetings on this topic that involve any of these attendees. Generate a one-page briefing for my principal that includes: outstanding decisions still owed by them, outstanding decisions owed to them by others, the two questions they should expect to be asked given the open items, and one question they should ask given what’s still unresolved. Items: [PASTE RELEVANT ROWS].”
This is the work nobody outside the executive support world realises is the job. The principal walks into the meeting already three steps into it. The other attendees will assume they read the briefing on the flight. They didn’t. You wrote it in four minutes.
A note on what AI can’t do here
The transcript-to-tracker workflow is solid because the underlying task (extracting structured information from messy text) is what large language models are reliably good at. The forward-looking briefing is shakier, because it asks the model to infer political context it doesn’t have. Don’t trust the “questions they should expect” output without sanity-checking it against your knowledge of the room. The model will sometimes invent tensions that aren’t there or miss the ones that are. The corrective is your judgement, which is the part of the job AI can’t replicate, and the part that makes a senior EA expensive to replace.
The post-meeting hour used to be where good EAs lost their evenings. Run the chain above and that hour becomes fifteen minutes plus a five-minute conversation with your principal about what to chase. The remaining forty minutes are the ones you spend on the work that actually compounds: relationships, anticipation, the quiet operating system of the office that nobody documents but everyone notices when it breaks.