TrueMinutes

Checks your AI notetaker's summary against the actual transcript and holds the recap before it goes out if it invented something.

Problem

Otter, Fireflies, Granola, Fathom, and Zoom AI Companion turn a call into action items in seconds, and most people forward that recap without re-listening to check it. A 2024 study (Kirstein et al.) found hallucinated content — invented dates, names, or commitments nobody made — in up to 37% of AI-generated meeting summaries. Once a fabricated deadline or task owner lands in a recap email, it becomes the record everyone works from, and nobody notices until the deadline is missed or the client asks why nobody did the thing they never actually agreed to.

Target user

An account manager, project lead, or founder running client or cross-team calls who relies on the AI notetaker's output to tell people what happens next, and doesn't have time to replay the recording after every meeting. What they want: know, within a minute of the call ending, whether the summary is safe to send as written or needs a fix first.

MVP scope

• Ingest a transcript plus its AI-generated summary — pasted, uploaded, or pulled via webhook from Otter, Fireflies, Granola, Fathom, or Zoom AI Companion exports
• Break the summary into atomic claims (who committed to what, by when) with a claim-extraction pass
• For each claim, search the transcript for supporting text and return a confidence score plus the exact quote and timestamp it's grounded in
• Flag three failure types: unsupported claims with no matching transcript text, misattributed owners (right task, wrong person), and omitted commitments — something promised on the call that never made it into the summary
• "Hold for review" mode that pauses the automatic Slack or email send until flagged items are cleared or a human overrides them, configurable per meeting type
• A redlined view of the summary showing flagged text next to the transcript excerpt it was checked against

Monetization

Freemium. Free tier: paste a transcript and summary, get the flagged report, capped at a handful of checks a month, no webhook automation. Paid tier ($15/user/month, or $99/month flat for small teams): webhook integration with the major notetakers, automatic hold-until-verified, and a Slack app for one-click approval of flagged recaps.

Why now

A July 17, 2026 survey found an AI notetaker has sat in on meetings for 33.4% of employed Americans, and adoption is still climbing across Otter, Fireflies, Zoom AI Companion, and Microsoft Copilot. At the same time, the accuracy problem hasn't gone away — the 37% hallucination-rate figure has been in circulation since 2024 and none of the major notetaker vendors ship a transcript-grounding check on their own output. The gap between how fast these summaries get trusted and how often they're wrong is the whole opportunity.

Risks & open questions

• If the notetaker's own transcript is wrong — real-world word error rates run from under 9% in clean audio to over 50% in noisy multi-speaker calls, per a PMC systematic review — TrueMinutes is grounding claims against a flawed source and inherits the error
• Otter, Fireflies, or Granola could ship a grounding check as a native feature, which would close the gap this product depends on
• People who don't currently double-check their AI summaries may not see the value in a tool whose whole job is to make them wait for a second opinion before hitting send
• Claim extraction and grounding need to run fast enough to sit between "call ends" and "recap goes out" — that rules out a slow multi-turn agent and pushes toward a narrow, fast model pass
• Webhook access depends on notetaker vendors' APIs, which could get restricted or paywalled for third-party integrations

Next step

Build a single web form that takes a pasted transcript and summary and returns the flagged claims; test it against 20 real meeting recordings and see how often it catches something a human reviewer would also flag.

Sources

• https://blog.saner.ai/ai-note-taking-statistics/ — cites hallucinated content in up to 37% of AI-generated meeting summaries (Kirstein et al., 2024) and word error rates from under 9% to over 50% depending on audio conditions
• https://abc17news.com/stacker-business-economy/2026/07/17/ai-notetakers-have-sat-in-on-1-in-3-us-workers-meetings-but-only-a-third-say-they-were-asked-first/ — July 17, 2026 survey: 33.4% of employed Americans report an AI notetaker present in their meetings