The Product
In early 2024, we launched a feature we were convinced would be a game-changer: AI-powered meeting summaries. The pitch was compelling—connect your calendar, we’ll join your meetings, and you’ll get a beautiful summary with action items afterward.
We spent four months building it. We launched with a splashy announcement. And within six weeks, we quietly sunset it.
This is what went wrong.
Mistake #1: We Built for a Persona, Not a Problem
Our target users were “busy executives who attend too many meetings.” We had data showing they attended 25+ meetings per week. We assumed they’d value anything that reduced meeting overhead.
What we missed: busy executives aren’t the ones taking action items—they’re delegating them. The summary wasn’t solving their problem; it was solving a problem for their direct reports, who weren’t our users.
Lesson: Validate the problem exists for the person who will use the solution. A real problem for the wrong persona is still the wrong bet.
Mistake #2: We Underestimated the Privacy Concerns
Having a bot join every meeting sounds useful in theory. In practice, it created awkward moments:
- “Who is ‘Notetaker AI’ and why is it in our call?”
- “Can you remove that bot? We’re discussing sensitive HR matters.”
- “Our client refuses to have meetings recorded.”
We’d anticipated some friction, but not the degree to which people simply couldn’t use the feature in their most important meetings.
Lesson: Features that require behavioral change from other people (not just your user) face compounding adoption friction.
Mistake #3: The Core Value Arrived Too Late
Even for users who did adopt it, the value came after the meeting—sometimes hours after, once the AI had processed the transcript. By then, they’d already written their own notes or moved on.
The async summary was neat, but it wasn’t solving an urgent need. “Nice to have” features don’t drive adoption.
Lesson: Time-to-value matters enormously. If your value arrives after the moment of need has passed, you’re not solving a problem—you’re creating a nice artifact.
Mistake #4: We Didn’t Kill It Fast Enough
The warning signs were there by week two:
- Activation rate: 12% (terrible)
- 7-day retention: 8% (worse)
- Support tickets: “How do I turn this off?”
But we kept iterating. “Maybe if we improve the summary quality…” “Maybe if we add Slack notifications…” We spent another month polishing a product that fundamentally wasn’t working.
Lesson: Set kill criteria before launch. “If we don’t hit X activation by week 4, we’re shutting down.” Sunk cost bias is real.
What I’d Do Differently
Prototype the Hard Parts First
The bot-joining-meetings flow was the riskiest part. We should have tested that interaction with real users before building the AI summarization layer.
Talk to Stakeholders, Not Just Users
The person affected by a meeting summary isn’t just the host—it’s everyone in the meeting. We should have interviewed meeting participants, IT admins, and compliance teams.
Define Success Metrics Up Front
“We’ll know it’s working if…” should be answered before writing code. We launched without clear success criteria and spent weeks debating whether 12% activation was “good enough.”
Have the Courage to Kill Faster
When the data says it’s not working, trust the data. Iterating on a fundamentally flawed concept is not the same as iterating toward product-market fit.
The Silver Lining
The failure wasn’t a total loss. We learned:
- Our users do want meeting-related features—just not this one.
- There’s a workflow opportunity around meeting prep, not just meeting follow-up.
- Integrations that require trust (like recording) need slow, thoughtful rollouts.
Six months later, we launched a simpler feature: meeting agenda templates synced from your project context. No recording, no AI joining calls—just helpful prep delivered before the meeting.
It’s now one of our most-used features.
The Meta-Lesson
Failure is expensive, but not as expensive as not learning from it. This launch cost us four months of engineering time and probably some customer trust.
But it also sharpened our product intuition. We’re more rigorous about validating assumptions. We’re faster to kill bad bets. We’re more humble about what we don’t know.
Would I trade the failure for a success? Of course. But since that’s not an option, I’ll take the lessons instead.