The AI Notetaker I Rolled Out to Nine People Earned Its Keep in Month Three, Not Week One

I put an AI meeting notetaker in front of nine people in March and for the first week it looked like a magic trick. The bot joined every calendar invite, sat through every standup, and dropped a clean transcript into our Slack channel before most people had found their coffee. Everyone was impressed, and I was briefly the person who fixed meetings. The real story is what happened in month three, when the pricing page’s omissions stopped being theoretical and started being line items.

The honest before-and-after: before, one person in every meeting spent it half-listening while typing notes, and the notes were whatever survived that split attention — usually action items with no owner. After, nobody took notes. The transcript landed within a minute of the call ending, and the AI summary pulled out decisions and named the person next to each action item. Across nine people averaging maybe six meetings a week, that recovered roughly twenty to twenty-five minutes per person, per week. Not dramatic. Real.

The pricing page said “unlimited AI minutes.” What it did not say, and what support confirmed only after I asked twice, is that meeting recordings and transcripts older than six months get deleted on the free-to-mid tiers unless you pay for the top one. That is not a limit on minutes. It is a limit on memory, and it matters precisely for the meetings you actually want to look back at — the ones from six months ago when you are trying to reconstruct why a decision was made.

The second omission is what happens when you leave. Your notes are not exactly yours. The export I tested produced a folder of per-meeting text files with timestamps but no speaker attribution preserved, and the shared “spaces” I had built — the thing that made the tool useful for a team — did not export at all. Leaving would have meant choosing between paying to keep access to two years of context, or losing the connective tissue and keeping raw transcripts. Neither was on the sales page.

Nobody budgets the migration, because migration is invisible until you do it. We had two years of notes scattered across a wiki, a shared drive and a few people’s personal folders. Importing them was the entire first week I had not planned for, and about half of it was cleaning up things we should have cleaned up years ago. The switch cost was not the subscription. It was the week, and the week is what the pricing page never quotes.

The seat math was the subtlest surprise. I assumed nine seats meant the nine people who joined calls. In practice a “seat” also got consumed by anyone who opened a transcript link, which at our size was closer to fourteen people once a couple of stakeholders wanted read access. So the real per-seat cost was about fifty percent higher than the sticker, and the sticker had been quoted to me per seat specifically so the annual number would look smaller than it was.

Was it worth it? Three months in, seven of the nine people use it without being reminded, which is a higher hit rate than anything else I have rolled out to this team. The time recovered is real but modest; the thing I did not predict is that the transcripts became the de facto institutional memory, and people now check them instead of asking each other. On balance I would switch again. I would just budget for the export problem on day one, not discover it on the way out.

If I had to compress it: before the demo, ask three questions and do not accept a vague answer. What happens to recordings older than a year? What exactly does the export contain, and can I see a sample? Who counts as a seat — everyone who joins, or everyone who reads? If the salesperson answers all three cleanly, the tool might be fine. If they reach for the top-tier plan, that is your answer.