AI for meetings: before, during, and after

AI for Meetings: Before, During, and After

The Meeting That Almost Derailed a Product Launch

In early 2023, the product team at Shopify was running three-hour planning sessions every week trying to align engineering, marketing, and customer success on a major feature rollout. Each meeting ended with a shared Google Doc full of action items that nobody owned and a vague sense that important decisions had been made — but nobody could agree on exactly what they were. The meeting organizer spent 45 minutes after each session reconstructing a summary from memory and scattered notes. By the time it reached inboxes, half the attendees had already moved on to conflicting interpretations of what was decided.

This isn't a Shopify-specific failure. It's the default state of knowledge work. Microsoft's 2023 Work Trend Index surveyed 31,000 workers across 31 countries and found that the average professional spends 57% of their time in meetings and on communication — leaving less than half the workday for the focused work those meetings are supposed to enable. More damaging: 64% of workers say they don't have enough time and energy to do their actual jobs after meetings end. The meetings aren't just unproductive. They actively consume the capacity needed to act on their outcomes.

What changed the dynamic for teams like the one at Shopify wasn't a new meeting methodology or a stricter agenda format. It was AI entering the workflow at three specific points: before the meeting to sharpen preparation, during the meeting to capture what actually happened, and after the meeting to transform discussion into durable action. Each intervention looks simple in isolation. Together, they compress what used to take hours of pre- and post-meeting overhead into minutes — and they produce outputs that are more accurate, more actionable, and more consistently followed through on.

The Principle Hidden in the Chaos

The real problem in that Shopify scenario wasn't the meetings themselves — it was the cognitive overhead surrounding them. Preparing well takes research and synthesis. Capturing decisions accurately takes total attention while also participating. Writing a useful follow-up takes time and judgment that most people don't have at 5pm on a Tuesday. AI doesn't make meetings shorter. It removes the three tax brackets that drain the most energy: preparation overhead, in-meeting documentation burden, and post-meeting synthesis lag. When those three costs drop, the meeting becomes what it was always supposed to be — a focused human decision-making event, not an administrative marathon.

How a 12-Person Consulting Firm Rewired Its Client Calls

McKinsey & Company published internal guidance in 2023 encouraging consultants to use AI tools for meeting preparation — specifically to synthesize client background, recent news, and prior engagement notes into a pre-meeting brief. Smaller firms quickly adopted the same approach. Consider a 12-person strategy consultancy running discovery calls with new enterprise clients. Before AI, a consultant would spend 90 minutes before each call reading annual reports, scanning LinkedIn for attendee backgrounds, reviewing past email threads, and trying to construct a mental model of the client's current priorities. The resulting brief was usually incomplete, sometimes outdated, and always took too long.

With ChatGPT or Claude, that same consultant pastes in the client's recent press releases, a summary of the email thread, and the names of who'll be in the room. They ask the model to identify the three most likely strategic concerns this client is facing, suggest five questions that would surface unstated assumptions, and flag any tensions between what the client says publicly and what the business data suggests. The output isn't perfect — it still needs human judgment to filter and prioritize. But it compresses 90 minutes of background work into 20 minutes of review and refinement. That's not a marginal improvement. It's a structural change in how preparation scales.

The key insight from consulting practice is that AI-assisted preparation shifts the consultant's role from researcher to editor. You're no longer generating the brief from scratch — you're stress-testing and sharpening a draft. This is faster, but it also produces better outcomes because you spend your cognitive energy on judgment rather than retrieval. The same principle applies whether you're a consultant preparing for a client call, a manager preparing for a performance review, or a marketer preparing for a campaign debrief.

A Different Industry, the Same Structural Problem

Shift industries entirely — from consulting to healthcare administration — and the pre-meeting problem looks almost identical. Hospital department heads at systems like Kaiser Permanente run weekly operational reviews covering staffing, patient throughput, budget variances, and incident reports. Before AI tools entered the workflow, a department head would spend Sunday evening manually pulling data from three separate systems, building a summary table in Excel, and drafting talking points for Monday's meeting. The prep work wasn't intellectually demanding — it was just slow, manual, and error-prone. Mistakes in that summary had real consequences: misallocated staff, delayed decisions, and meetings that ran long because the baseline data was disputed.

Teams at several hospital systems began using Microsoft Copilot (embedded in Teams and Excel) to automate the data-pull and first-draft summary for these operational reviews. The model pulls from connected data sources, surfaces anomalies — a ward with unusually high agency staff hours, a budget line running 15% over — and generates a structured briefing document. The department head reviews and edits rather than builds from zero. Monday morning prep dropped from two hours to 30 minutes. More importantly, the summaries became more consistent, which meant meetings started from shared facts rather than disputed ones. Decision quality improved not because the AI was smarter than the department head, but because it removed the variance that came from rushed, manual preparation.

AI Meeting Tools Compared

The Marketing Director Who Stopped Dreading Monday Recaps

Take a different role entirely: a marketing director at a mid-size e-commerce brand running a weekly campaign review with her team of six. Every Monday, she recapped the previous week's performance across paid search, email, social, and SEO — pulling numbers from Google Ads, Klaviyo, Meta Business Suite, and Google Analytics. The recap itself was valuable. The 45 minutes she spent building it Sunday night was not. She started using a two-step AI workflow: first, she exported raw data from each platform as a CSV or copied key metrics into a structured prompt for ChatGPT; second, she asked the model to identify the three biggest performance shifts, explain likely causes based on the data patterns, and draft three agenda items for Monday's discussion. The model doesn't have access to her accounts — she's feeding it the data manually — but the synthesis it produces is faster and often more pattern-aware than what she'd catch in a tired Sunday-evening scan.

What she discovered after four weeks was something she hadn't anticipated: the AI-generated agenda items were better discussion starters than the ones she'd been writing manually. Because the model wasn't attached to any particular hypothesis, it surfaced patterns she'd unconsciously been explaining away — a consistent Thursday dip in email open rates, a paid search campaign that was technically hitting its CPC target but generating zero repeat purchasers. Those observations became the most valuable 15 minutes of Monday's meeting. The AI didn't run the meeting. It gave her something sharper to bring into it.

What This Looks Like in Real Workflow Terms

The three examples above — the consultant, the hospital department head, and the marketing director — each used AI differently, but they all moved the same cost. In each case, the expensive pre-meeting work was synthesis: taking distributed, messy information and turning it into a focused, actionable starting point for a human conversation. That synthesis task is exactly where large language models operate at their best. They don't get tired at 10pm. They don't have a vested interest in the conclusion. They process the inputs you give them and return structured outputs faster than any human can. The professional's job becomes curation and judgment — which is where human expertise actually lives.

There's an important constraint to name clearly: AI-assisted preparation is only as good as the inputs you provide. The consultant who pastes in a thin email thread and no client context gets a thin brief. The marketing director who feeds the model last week's numbers without flagging that a major promotion ran on Thursday gets a confused analysis. The model doesn't know what it doesn't know. Your job in the preparation phase is to be a rigorous curator of inputs — giving the model enough context to reason well, then editing its output with the domain knowledge it lacks. This is a collaboration, not a delegation.

This also means the prompts you use for pre-meeting preparation are worth building and saving. A consultant who writes a strong discovery-call prep prompt once can reuse and adapt it for every new client engagement. A marketing director who nails her weekly performance-review prompt runs that synthesis in three minutes every Sunday instead of 45. The investment in learning to prompt well compounds over time — each meeting type you crack becomes a reusable template that scales across your whole calendar. By the time you're working through the full before-during-after framework, you'll have a small library of prompts that fundamentally change how you operate.

What the Stories Teach Us

1. The real cost of meetings isn't meeting time — it's the preparation and follow-up overhead that most professionals absorb invisibly, averaging 4.8 hours per week.
2. AI tools work best in the pre-meeting phase when you shift your role from researcher to editor: let the model synthesize, then apply your judgment to the output.
3. Different tools solve different parts of the problem — Otter.ai and Fireflies.ai focus on in-meeting capture, while ChatGPT and Claude excel at flexible preparation and synthesis tasks.
4. The 'feed it the data' approach lets you get AI-quality analysis from any platform without requiring direct integrations — manual input of exported data is fast and effective.
5. AI preparation is only as strong as the inputs you provide. Thin context produces thin output. Rigorous curation of what you give the model is a core skill, not a workaround.
6. Prompts built for recurring meeting types compound in value — a strong weekly-review prompt used 50 times a year returns far more than the 30 minutes it took to build.
7. The consultant, the hospital administrator, and the marketing director all used different tools and different workflows — but the structural principle was identical: move synthesis overhead from human to AI, keep judgment with the human.

Key Takeaways from Part 1

• 57% of the average professional's day goes to meetings and communication — the overhead surrounding meetings is as costly as the meetings themselves.
• AI enters the meeting workflow at three points: before (preparation), during (capture), and after (synthesis). Part 1 covers the 'before' phase in depth.
• Pre-meeting AI preparation works by compressing research and synthesis tasks — turning 90-minute manual prep into 20-minute review and editing.
• ChatGPT and Claude are flexible tools for pre-meeting briefs; Microsoft Copilot, Fireflies.ai, and Otter.ai are purpose-built for meeting workflows.
• Your role in AI-assisted preparation is curation and judgment — you provide context and edit output, not generate everything from scratch.
• Reusable prompt templates for recurring meeting types are one of the highest-ROI investments in your AI workflow.

What Happens When You Run Meetings at Scale

In 2022, Shopify made headlines when it deleted 76,000 recurring calendar events in a single day. The company calculated that a 50-person meeting costs roughly $700 per hour in salary alone — and most of those meetings were producing nothing close to that value. Their diagnosis wasn't that people were lazy or disorganized. It was structural: without a system for capturing and acting on meeting output, every discussion evaporated. Teams would reconvene a week later, rehash the same ground, and leave with the same vague commitments. The problem wasn't the meetings themselves. It was the gap between what was said and what actually happened next.

Shopify's radical calendar purge forced teams to justify every meeting from scratch. But most organizations can't — or won't — do that. What they can do is close the gap between conversation and action using AI. The Shopify story reveals a principle that holds across industries and team sizes: meeting value isn't created in the room. It's created in the documentation, the decisions, and the follow-through that come after. Every tool, technique, and prompt in this section is designed to capture that value before it disappears.

The During-Meeting Problem Nobody Talks About

Most professionals assume the hardest part of meeting productivity is the aftermath — the notes, the follow-ups, the action items. But there's a subtler problem that happens in real time. When you're actively listening, contributing, managing dynamics, and tracking the agenda simultaneously, your cognitive load is maxed out. Something always gets dropped. Usually it's the meta-level tracking: noticing when the group drifts off-topic, flagging a decision that was made implicitly without consensus, or catching a commitment someone made casually that nobody wrote down. These are exactly the failure points that turn good meetings into wasted ones.

Otter.ai and Fireflies.ai both address this by running as silent participants. They transcribe in real time, identify speakers, and tag action items as they're spoken. During a 45-minute product review at a mid-sized SaaS company, Fireflies might capture 14 distinct action items — while the human facilitator's handwritten notes show only six. The gap isn't negligence. It's the natural limit of divided attention. When you're defending a roadmap decision or navigating a stakeholder's concerns, you're not simultaneously logging commitments. The AI is.

This changes the facilitator's job in a meaningful way. Instead of being the meeting's stenographer, you become its conductor. You can focus entirely on the quality of the conversation — pushing for clearer decisions, surfacing disagreements early, managing the energy in the room — because the documentation layer is handled. Teams that adopt this approach consistently report that their meetings become shorter and more decisive. When participants know everything is being captured, they stop repeating themselves for emphasis and start moving faster toward resolution.

A Hospital System That Reclaimed 11 Hours Per Week

Healthcare is one of the most meeting-heavy industries in existence — and one of the most documentation-burdened. A regional hospital network in the American Midwest piloted AI meeting tools across their administrative and clinical leadership teams in late 2023. The specific problem they were solving wasn't efficiency in the abstract. It was compliance: clinical decisions made in meetings had to be traceable, documented, and accessible for audits and handoff protocols. Their previous system relied on a rotating notetaker role, which introduced inconsistency, created bottlenecks, and occasionally resulted in critical decisions being poorly documented or lost entirely.

After deploying Otter.ai with custom vocabulary trained on medical terminology, the network reported a 40% reduction in time spent on post-meeting documentation — roughly 11 hours per week across their leadership team. More significantly, the quality of documentation improved in ways that mattered for compliance: action items were consistently captured with owners and timestamps, and the AI's speaker identification meant accountability was traceable rather than assumed. The hospital's Chief Operating Officer noted that disputes about what was decided in a meeting — previously a frequent source of friction — dropped to near zero within the first quarter.

Comparing the Major AI Meeting Tools

The Consultant Who Stopped Writing Proposals from Scratch

A management consultant at a boutique firm in London describes her pre-AI workflow as a familiar kind of tax: every client discovery meeting was followed by two to three hours of writing. First the meeting notes, then the synthesis document, then the first draft of the proposal or recommendation. The content of those documents was almost entirely derived from what was discussed in the meeting — but translating spoken conversation into polished written output took hours every time. Multiply that across four to five client engagements per week, and she was spending close to 15 hours weekly just converting meetings into documents.

Her current workflow looks like this: Otter.ai runs during every client call and exports a cleaned transcript. She pastes that transcript into Claude with a prompt specifying the document type she needs — discovery summary, proposal outline, or follow-up email. The first draft is ready in under two minutes. She spends 20 to 30 minutes editing for tone, adding proprietary insights, and aligning with her firm's voice. Total post-meeting writing time dropped from 2.5 hours to 45 minutes per engagement. Over a year, that's roughly 300 hours recovered — the equivalent of more than seven full working weeks. The work didn't get worse. Her clients consistently rate her follow-up documentation as clearer and more structured than before.

What Good AI-Assisted Meeting Output Actually Looks Like

There's a meaningful difference between an AI summary that technically captures the meeting and one that's actually useful to the people who weren't there. The first kind reads like a transcript compressed into bullet points — technically accurate but narratively flat, with no signal about what matters most. The second kind reflects the meeting's logic: what the group was trying to decide, what evidence or arguments shaped the outcome, what the decision was, and what happens next. Getting from the first kind to the second requires prompting with intent, not just pasting a transcript and hitting enter.

The most effective meeting summaries treat the output as a communication artifact — something a busy executive or absent team member can read in 90 seconds and understand everything they need to. This means structuring output hierarchically: the most critical information (decision made, who owns what) at the top, supporting context below, and open questions at the end. When you prompt ChatGPT or Claude with this structure explicitly, the output reflects it. When you prompt vaguely, you get a vague summary. The model's quality ceiling is high — your prompting determines how close you get to it.

One underused technique is asking the AI to write the same meeting summary in two different formats: a short version for stakeholders who weren't in the room, and a detailed version for participants who need a reference document. These serve different audiences with different needs, and generating both takes less than a minute of additional prompting. Teams that distribute the short version to senior leadership and the detailed version to the working team find that alignment improves significantly — because both groups are working from the same source of truth, just presented at the right level of detail for each.

Principles Extracted from the Examples

1. Meeting value is created after the meeting. The conversation is the raw material — documentation, decisions, and follow-through are the actual output. AI tools make that conversion faster and more complete.
2. Divided attention is the enemy of good documentation. When you're running a meeting, you can't simultaneously track every commitment and decision. AI transcription removes that tradeoff entirely.
3. Prompt structure determines output quality. Pasting a transcript and asking for 'a summary' produces mediocre results. Specifying format, hierarchy, and purpose produces outputs people actually use.
4. Different audiences need different formats. A 3-sentence stakeholder summary and a detailed action item log serve different people with different needs — and generating both costs you nothing extra.
5. Recurring meetings deserve saved prompts. Building a prompt library for your standard meeting types is a one-time investment that pays off every week for as long as you run those meetings.
6. AI catches what humans miss. In real-time meetings, action items made casually or in passing are the most likely to be forgotten. Transcription tools don't filter by perceived importance — they capture everything.
7. Compliance and accountability are byproducts of good AI documentation. Industries with documentation requirements — healthcare, legal, financial services — gain an additional layer of value from AI meeting tools beyond pure efficiency.

Key Takeaways from This Section

• Shopify's $700/hour meeting cost calculation is a useful mental model: if your meeting isn't generating that level of value in decisions and output, the problem is usually the documentation gap, not the meeting itself.
• Otter.ai and Fireflies.ai are purpose-built for live transcription; ChatGPT and Claude are more powerful for post-meeting processing — your stack should include both types.
• The consultant example demonstrates a concrete ROI: 15 hours of weekly writing reduced to under 4 hours, with no loss in output quality.
• Microsoft Copilot and Google Gemini are strong if you're already in those ecosystems — but at $30/user/month, Copilot requires clear ROI justification before rolling out broadly.
• Two-format output (stakeholder summary + participant reference doc) is one of the highest-leverage habits you can build into your post-meeting workflow.

After the Meeting: Where AI Turns Chaos Into Clarity

In 2023, Shopify made a controversial internal decision: it banned unnecessary recurring meetings and required written summaries for those that remained. The mandate forced teams to produce clear, searchable records of every decision made. What Shopify discovered — and what productivity researchers have confirmed — is that most meeting value evaporates within 24 hours if it isn't captured. People leave a room with different mental models of what was agreed. Action items get confused with suggestions. Follow-through drops. The problem wasn't the meetings themselves. It was the absence of a reliable, structured record that everyone could trust and act on.

Shopify's teams that adopted AI-assisted note-taking saw a measurable shift. Instead of one person frantically typing while trying to participate, the whole team could focus on the conversation. Afterward, tools like Notion AI or Claude processed the transcript and produced a structured summary: decisions made, owners assigned, deadlines stated. The summary became a shared source of truth. Disagreements about what was agreed dropped significantly. That's not a soft cultural win — it's a direct reduction in the rework and miscommunication that costs companies billions annually. The post-meeting document became as important as the meeting itself.

The principle Shopify stumbled onto applies across every industry: the post-meeting output is where value either compounds or disappears. A brilliant strategy session means nothing if three people walk away with three different action plans. AI changes the economics of documentation. Previously, writing a thorough meeting summary took 20-40 minutes and usually fell to the most junior person in the room — who also had the least context. Now, a 60-minute recorded meeting can be summarized by Claude or ChatGPT in under two minutes, with a level of structure and completeness that a tired human rarely matches. The bottleneck has been removed.

What Good AI Summaries Actually Look Like

The quality of your post-meeting AI output depends almost entirely on how you prompt it. Paste in a raw transcript and ask for 'a summary' and you'll get a mediocre paragraph that buries the decisions. Give the model a structured prompt with explicit output requirements and you get something you can paste directly into Slack, forward to a client, or drop into your project management tool. The difference between a generic summary and an actionable one is about 30 extra words in your prompt — a trivial investment for the return it produces.

Two Industries, Same Problem Solved Differently

At a mid-sized law firm in Chicago, associates were spending three to four hours per week writing up internal case review meetings — time that billed at $300+ per hour. After piloting Otter.ai combined with GPT-4 summaries, they cut that documentation time by 70%. The structured summaries also surfaced a pattern no one had noticed: certain case types consistently produced unresolved questions about jurisdiction, signaling a training gap. The AI didn't just save time — it made the firm's collective thinking visible in a way manual notes never had.

In a completely different context, a product team at a European fintech used Claude to process their weekly sprint retrospectives. The AI generated a 'recurring themes' section that tracked which problems kept resurfacing across weeks. Within two months, the team had a data-backed case to present to engineering leadership: three infrastructure issues were responsible for 60% of sprint delays. That argument — built from AI-synthesized meeting records — directly led to a resource reallocation. The meeting summaries became institutional memory, not just administrative paperwork.

A senior consultant at McKinsey described her workflow to a professional forum in early 2024: she records every client working session, runs the transcript through Claude with a tightly structured prompt, and has a draft client update ready before she's left the building. Her clients receive follow-up documentation within an hour of the meeting ending. The response from clients, she noted, was striking — they perceived her as unusually organized and thorough, which strengthened trust and accelerated project timelines. The AI wasn't replacing her thinking. It was removing the lag between her thinking and its delivery.

What this consultant's example reveals is something worth sitting with: speed of follow-up is itself a form of professional quality. When a client receives a clear, accurate summary 45 minutes after a meeting, it signals competence, respect for their time, and organizational maturity. Most professionals send follow-ups the next day, or not at all. AI makes same-hour documentation not just possible but easy — and it shifts the baseline expectation for what 'professional' looks like. This is a compounding advantage. The first few times you deliver it, clients notice. Over months, it defines your reputation.

The practical implication is a simple new ritual: within 30 minutes of any significant meeting, paste the transcript or your rough notes into ChatGPT, Claude, or Gemini with a structured prompt. You don't need perfect notes — even a rough bullet list of what was discussed gives the model enough to work with. The output becomes your first draft. You edit, not create from scratch. Over a full working week, this shift reclaims one to three hours that previously disappeared into the documentation black hole. Across a year, that's weeks of recovered time redirected toward actual work.

1. Post-meeting documentation is where meeting value either compounds or evaporates — AI makes thorough documentation the default, not the exception.
2. Prompt structure determines output quality: generic prompts produce generic summaries; explicit section requirements produce actionable documents.
3. Speed of follow-up is a professional signal — same-hour summaries build trust and differentiate you from peers who send notes the next day or not at all.
4. AI meeting summaries create institutional memory — over time, they reveal patterns (recurring problems, unresolved questions) that manual notes never surface.
5. Always review AI output before distributing — two minutes of checking prevents the misattribution or missed nuance that undermines credibility.
6. The biggest gains come from consistency — one great summary is useful; a system that produces great summaries every week is transformative.

• Paste transcripts into Claude or ChatGPT with a structured four-section prompt to get summaries you can send immediately.
• Assign owners and deadlines explicitly in your prompt — the model will flag gaps you might have missed.
• Send AI-generated summaries within 30-60 minutes of the meeting; this alone separates you from 90% of peers.
• Save your best-performing prompts as reusable templates — your system improves each time you refine them.
• Use accumulated summaries to spot recurring themes across meetings — this turns documentation into strategic intelligence.
• Review every AI summary before sending — accuracy is non-negotiable when the document goes to clients or leadership.