The average knowledge worker spends about 21 hours a week in meetings. Microsoft's 2023 Work Trend Index put the number at 23 hours for managers. If even a third of those meetings are avoidable, redundant, or could have been an email, that's 7 to 8 hours of your week going somewhere that doesn't produce work.
AI doesn't fix bad meeting culture. Your organization has to do that. But it can absorb most of the work that surrounds meetings, so the meetings you do attend cost you less, and it can give you the leverage to push back on the ones you shouldn't be in.
Here's what actually works, broken down by where in the meeting cycle the time leaks.
The Pre-Meeting Filter
Most meeting overhead starts before the meeting. You see an invite, you don't have time to figure out what it's for, you accept it anyway, and an hour disappears from your calendar. The cost compounds across a week.
Drop the meeting invite text, the agenda if there is one, and the attendee list into Claude with this prompt: "Based on this invite, what is this meeting actually for? What decisions are likely to be made? Do I need to be there, or can I get the outcome from notes? If I do need to be there, what's the one thing I should be prepared to contribute?" That four-part framing forces a specific answer. The model will tell you when the meeting reads like a status update you could skip, or when your specific role is unclear.
For recurring meetings, this is more useful than it sounds. Take the last three notes from a recurring meeting and ask: "Looking at these three sets of notes, what's the actual purpose of this meeting, and what percentage of the content required real-time discussion versus could have been async?" If the answer is over 70 percent async-suitable, you have data to bring to whoever owns the meeting.
Real Prep in 10 Minutes
Meeting prep usually means scanning the last thread, remembering what you committed to, skimming the doc nobody read, and figuring out what the other person actually wants. Most people skip half of this because they don't have time, then improvise in the meeting and produce worse outcomes.
For an internal meeting, paste the relevant Slack thread, the doc, and any previous notes into Claude. Ask: "Summarize the state of this conversation in five bullets. What does each person seem to want? What's the unresolved tension? What questions am I likely to be asked, and what's my best answer to each?" You're not outsourcing your thinking. You're outsourcing the catch-up so you can spend your prep time on the part that requires judgment.
For an external meeting, the same workflow with one addition. Run the company name through Perplexity with "recent news, funding, leadership changes, product launches in the last six months." Then ask Claude to synthesize that with anything you have on the contact. You walk in knowing what changed for them recently, which is the single biggest signal of being prepared.
Live Notes Without the Cognitive Tax
Taking notes while paying attention is a known impossibility. You do one or the other, usually badly. Granola, Otter, Fathom, and the Zoom and Teams built-in AI notetakers all solve this. The differences matter less than people think. Pick the one that integrates with how you already work and stick with it.
Granola is the one I'd recommend for Mac users who care about privacy. It runs locally, listens to your meetings, and produces structured notes you can edit. Otter is better if you need transcription as a primary output and want native Zoom and Google Meet integration. For Microsoft shops, Copilot's built-in recap in Teams is good enough that you usually don't need a third tool.
The non-obvious move: don't rely on the auto-generated summary. Right after the meeting, spend two minutes adding three things to the notes. The actual decision that got made, any commitment you personally made, and the next checkpoint. AI captures what was said. It misses what was decided, because decisions often come from tone, eye contact, or a ten-second exchange the model treats as filler.
Follow-Ups That Actually Land
The follow-up email is where most meetings die. People leave with a vague sense of next steps, no one writes them down, and two weeks later nothing has moved. A 15-minute meeting followed by a 90-second follow-up email is worth more than a 60-minute meeting with no recap.
Take your notes, the transcript if you have it, and run this prompt: "Draft a follow-up email to [attendees]. Three sections: what we decided, who owns what by when, and what we'll revisit at the next checkpoint. Keep it under 200 words. Match the tone of someone who respects the reader's time." Edit for accuracy, send it within an hour of the meeting ending. The faster you send the recap, the more likely the meeting outcomes stick.
For client-facing meetings, push it further. Ask the model to write a version of the email that flags the two things the client might push back on at the next meeting, with a draft response for each. That's the prep for the next meeting embedded in the follow-up of this one, and it costs you five extra minutes total.
When You Should Cancel Instead
AI also gives you the leverage to kill meetings you shouldn't be in. The hardest part of meeting reduction isn't deciding which meetings are useless. It's writing the message that gets you out of them without sounding dismissive.
Try: "Draft a Slack message to [person] explaining that I think this meeting could be handled async. Acknowledge what they're trying to accomplish, propose a specific written format that would serve the same goal, and offer to be available for a 10-minute clarification call if they need it. Keep it warm and direct." The output is rarely perfect, but it gives you a starting point that's 80 percent there. You edit for accuracy, send it, and reclaim an hour.
For meetings you can't cancel but shouldn't attend in full, the same prompt with a different ask: "Draft a message saying I'll join for the first 15 minutes to give input on [topic], then need to drop." Most meeting owners accept this if you frame it as partial commitment rather than a no.
The Math On Time Saved
Let's be specific. If you have 15 meetings a week, and AI saves you 10 minutes of prep on each one by absorbing the catch-up work, that's 2.5 hours. If it saves you 5 minutes per follow-up email, that's another 75 minutes. If it helps you cancel or shorten three meetings a week, that's 90 more minutes minimum. You're at 5 hours of recovered time per week without any change to your organization's meeting culture.
The 8 hours figure in the headline assumes you also push back on the worst recurring offenders, which is the part that requires actual organizational courage. AI gives you the data and the drafted message. Whether you send it is on you.
What Not to Do
Don't use AI to attend meetings for you. Avatar bots that join calls on your behalf are a bad look in most professional contexts and they erode trust faster than the time savings justify. The right move is to skip the meeting and read the AI-generated notes, not to send a synthetic version of yourself.
Don't transcribe sensitive meetings with consumer-grade tools without checking your organization's policy. Recording laws vary by jurisdiction and many companies have explicit rules about which tools can listen to client calls. The Granola local-only model is the safest default if you're not sure.
Don't trust the auto-generated action items without reading them. AI is good at extracting what was said and often wrong about what was decided. Read the action items before sending the follow-up. Five seconds of verification prevents a week of confused execution.