**I Built an AI System That Handles 70% of My Meeting Processing. Here's What Actually Happened.**

Last quarter, I was writing a Slack update at 7 PM for a meeting that ended at 11 AM.

That eight-hour lag — between "decision was made" and "my stakeholders know what was decided" — was the version of my workday I was done with.

I manage multiple regulatory compliance programs simultaneously across international markets. At any given week, I'm running or attending 8–12 meetings, coordinating across 15+ engineering, legal, product, and compliance teams, and trying to keep a biweekly executive update series timely, accurate, and actually worth reading.

The manual processing alone — reviewing Zoom transcripts, writing structured meeting notes, extracting action items, drafting Slack summaries, updating the program health tracker — was consuming 3–5 hours a week. Hours that were supposed to go toward the actual program work.

So I built a system to handle it. Not a vendor tool. Not a purchased workflow. I built it myself, with Claude Code, Glean, and Airtable — and it now handles roughly 70% of my meeting processing with no manual intervention after setup.

Here's what I built, what I learned, and what I would do differently.

The first thing I had to do was stop treating meeting processing as "just part of the job."

It's easy to normalize the hours you spend converting a Zoom transcript into usable notes — because it feels like productive work. It isn't. It's transformation overhead: the cost of converting raw meeting content into a form that other people can actually act on.

When I mapped out the full processing chain for a single meeting, it looked like this:

1. Access Zoom transcript
2. Review for decisions, action items, open questions
3. Write structured notes with speaker attribution
4. Send notes to attendees
5. Draft Slack summary for the senior stakeholder channel
6. Extract action items and push to the tracker with owners and due dates
7. Update program health status
8. Refresh the program health dashboard

Eight steps. One meeting. Multiply that by 8–12 meetings a week across multiple programs.

That math is not sustainable for one person.

The system automates steps 1 through 7 of that chain.

Transcript ingestion

I access Zoom meeting transcripts through Glean, our enterprise search platform, which indexes them automatically. Using an MCP (Model Context Protocol) integration, the transcript goes directly into Claude Code without any copy-paste.

Structured notes generation

Claude reads the full transcript and produces a structured meeting notes document: date, attendees, decisions with rationale, action items with owners and due dates, and a parking lot for items raised but not resolved. The output is ready to send with minimal editing — usually a quick pass to verify name attributions and catch anything the AI missed.

Slack executive summary

From the same transcript, Claude drafts a leadership-ready program update for the Slack channel where my senior stakeholders track health. It follows a template I defined: health status, key decisions this week, what's coming, what needs executive attention. I review and edit before posting — but the draft is 80–90% there before I touch it.

Action item push to Airtable

Using an Airtable MCP integration, Claude extracts action items from the structured notes and writes them directly to the program tracker — assignee, description, and due date pre-populated. I review and confirm before anything is committed.

Program health dashboard update

After action items are confirmed, the program health dashboard — a GitHub Pages site reading from Airtable — refreshes automatically. Any executive or stakeholder can see current health across all programs without asking me for a slide deck.

Start with one workflow, not all of them.

I made the mistake of trying to design the full system before building any of it. The better approach: pick your single most painful step, build that piece alone, run it reliably for two weeks, then add the next layer. For me, the right starting point would have been just transcript → structured notes. Immediate visible value, and it builds your confidence in the output before you start automating downstream decisions.

Template the output before you automate it.

Output quality improves dramatically when you give the AI a structured template to fill rather than asking it to generate structure from scratch. My notes template has fixed sections with consistent headings. That means output from meeting A looks like meeting B — easier to maintain, easier for attendees to read, and easier for the AI to get right every time.

Human review is not optional — it is the feature.

I didn't build this to remove myself from the loop. I review every document before it goes to attendees or stakeholders. The system removes creation time; I still own accuracy and judgment. That's the distinction between automation that amplifies a professional's work and automation that replaces their accountability. I want the first kind.

The numbers: 70% reduction in manual processing time. Roughly 3–4 hours recaptured per week and redirected to actual program work — stakeholder conversations, dependency resolution, early risk identification.

The less obvious shift: the quality of my executive communication improved because I stopped composing Slack updates at 7 PM after finally finishing notes. The draft exists within two hours of any meeting ending. Leadership gets timely, accurate status instead of status that was accurate two days ago.

And the thing I didn't anticipate: building this made me a better program manager.

Designing the workflow forced me to think clearly about what a meeting's output actually needs to be. Decisions. Actions. Owners. Health signals. Everything else is noise.

The AI system is only as good as the structure you impose on it. And imposing that structure taught me to run sharper meetings.

You don't need to be an engineer. I'm not one.

What you need:

1. An AI tool that connects to your existing systems (Claude Code with MCP integrations worked for me; other options exist)
2. 2–3 hours to define your output templates before you start prompting
3. A willingness to run version 1 imperfectly and iterate from there

The version I have today is the fourth major iteration. Version 1 produced passable notes and nothing else. Version 2 added the Slack draft. Version 3 added the Airtable push. Version 4 added the dashboard sync.

The most important thing I did: ship version 1 before I had a plan for versions 2 through 4.

I'm a Senior Technical Program Manager at eBay, where I lead regulatory and compliance programs for eBay Live. I write about program management at scale, AI tooling for practitioners, and the TPM career path.

What's one part of your workflow you wish you could automate? Drop it in the comments — I'm curious what the pain points look like across different orgs.