Three years ago, an AI note taker felt like magic. A bot joined your call, and twenty minutes later you had a transcript, a summary, and a list of action items. Startups raised on it. Fireflies crossed a $1B valuation in June 2025. The category looked like a winner.

It's over. Not because the tools got worse - because three forces converged that no amount of better summarization can fix.

The platforms ate the product

Transcription accuracy is broadly commoditized - most tools land in the same 90–95% range. And the thing that used to be a product is now a checkbox: Zoom ships AI note-taking natively, Microsoft bundles Intelligent Recap into Copilot, and Google's Gemini notetaker now works even in in-person meetings — and on Zoom and Teams calls.

When the platform you're meeting on does your core feature for free, "good enough transcription plus summary" is over as a standalone product. You can't out-summarize a bundled default.

Nobody reads the notes

The dirty secret of the category: the output mostly goes unread. Roughly 70% of AI-generated action items are never completed. One analysis of 30,000+ AI summaries found nearly four in five miss critical context that changes the meaning of what was discussed. Teams report the 90 seconds saved on generation gets consumed by 12 minutes of verification and correction.

Atlassian found a third of action items never got documented before AI. Now they're documented — in a format teams ignore. A perfectly captured meeting that changes nothing afterward is a very expensive filing cabinet.

The bots are getting kicked out of the room

The backlash is no longer anecdotal. A federal class action, In re Otter.AI Privacy Litigation, alleges wiretap violations for recording non-users without consent. The NYC Bar Association warned in December 2025 that notetakers without strict client consent risk breaching confidentiality duties. And starting Q3 2026, Microsoft is rolling out admin block lists and tenant-level enforcement against third-party bots in Teams - naming Otter, Fireflies, Fathom, and Read.ai.

"Bot fatigue" is now a purchasing criterion. Companies are banning meeting bots outright via admin policy, and client-facing teams increasingly refuse meetings where a bot shows up.

What replaces them: agents that do the work

The category isn't dying because meetings stopped needing help. It's dying because the bar moved. The distinction that matters in 2026: notetakers transcribe and summarize; agentic platforms turn the conversation into actual work - filed tasks, drafted documents, sent follow-ups, updated systems of record.

Microsoft's Copilot can now take an intention from a meeting and execute across Word, Excel, Outlook, and Teams without a human between steps. The evaluation criterion across the market is no longer text quality but the ability to act in real workflows. Even the survivors prove the point: the tools still growing are the ones that specialized into verticals - advisor platforms like Jump and Zocks winning on compliance automation and CRM orchestration - not the ones that write prettier summaries.

Why we built Earmark this way

This is the thesis Earmark was built on - and each of the three failure modes above is a design decision in the product.

The bot problem: Earmark never joins as a bot. It captures audio on-device - no avatar in the participant list, no consent theater, no attendee ever sees it. That's why it works on Zoom, Teams, Meet, or across a coffee table, and why Microsoft's tenant-level bot enforcement doesn't touch it.

The nobody-reads-notes problem: Earmark doesn't produce notes to read - it produces work to ship. Live conversations become PRDs, Jira and Linear tickets with acceptance criteria, decision logs, and stakeholder updates. The 70% of action items that die in summaries die because they never reach a task system; Earmark pushes them into Linear before the meeting ends.

The passive-capture problem: Earmark works during the meeting, not after it. Task agents draft the deliverables while you're still talking, and real-time guidance surfaces the questions you haven't formulated yet - so you leave the call with the follow-up already started, not a to-do list for the second shift.

The obituary, in one line

AI note takers solved capture. But capture was never the problem - follow-through was. The meeting tools that survive won't be the ones that remember what you said. They'll be the ones that make sure it happens.

That's the product we're building.