AI meeting notes and summaries with AI in 2026: how far it works, which tools, what it costs
AI notetakers can record a call, transcribe it and draft a summary with action items. The strongest open speech models average between 5.6% and 7.4% word error on English benchmarks, but summaries still need a human check for names, numbers and decisions. Recording people without clear consent is now a legal risk: wiretap and privacy claims against Otter.ai survived a motion to dismiss in August 2026. Pick one tool, tell attendees, and read the notes before they go out.
Where AI stands: AI meeting notes and summaries · October 2026
As of October 2026, AI is ready to transcribe meetings and draft recaps you can edit, still early for bot-free note capture, and not ready to be error-free, to record without consent, or to hear every accent equally well.
- Transcribing meeting speech
Backed by trials or large deployments
- Drafting recaps and action items
- Bot-free note capture
Promising, limited outcome data
- Transcripts with no errors
- Recording without consent
- Equal accuracy across accents
Evidence says avoid or add safeguards
What works today
- 1
Proven Transcribe English speech at an average word error rate of 5.6% to 7.4% in the strongest open models.
Evidence
Open ASR Leaderboard paper (October 2025 snapshot of 64 models). The average covers the leaderboard's English datasets, including 9 hours of AMI meeting speech. Top listed model 5.63%, OpenAI Whisper Large v3 7.44%. Noisy calls with crosstalk score worse. Source ↗
- 2
Early Draft meeting recaps and highlight lists that attendees can edit and share.
Evidence
Microsoft Research prototype study, CSCW 2024. Seven users tested highlight and structured-minutes recaps in their own work meetings. Small sample, and the paper reports no accuracy figures. Source ↗
- 3
Early Take notes from your own device's audio, with no bot joining the call.
Evidence
Granola pricing page (checked October 2026) describes bot-free, device-audio capture. Vendor statement with no independent test, and attendees should still be told. Source ↗
What doesn't work yet
- 1
Can't Guarantee that every transcribed phrase was actually spoken.
Evidence
Koenecke et al., FAccT 2024. In OpenAI Whisper transcripts of speakers with aphasia and a control group, about 1% of transcriptions contained fabricated phrases, and 38% of those carried explicit harms. Tested in 2023; speech with long pauses was hit hardest. Source ↗
- 2
Can't Record or transcribe people without telling them.
Evidence
In re Otter.AI Privacy Litigation, N.D. Cal. No. 5:25-cv-06911. On 13 August 2026 Judge Eumi K. Lee let the federal Wiretap Act, California CIPA and Illinois BIPA claims proceed, finding it plausible that Otter retained and used recordings for its own commercial purposes (as reported by UC Today and RecordingLaw.com). These are allegations, and no liability has been found. Source ↗
- 3
Can't Transcribe every accent and dialect equally well.
Evidence
Koenecke et al., PNAS 2020. Five commercial systems (Amazon, Apple, Google, IBM, Microsoft) averaged a 0.35 word error rate for 73 Black speakers against 0.19 for 42 white speakers, over 19.8 hours of interviews. Systems were 2019 versions, so test your own tool. Source ↗
Where to start
Team lead
Pilot one notetaker on internal calls for two weeks, and have the meeting owner check each summary against the transcript before it is shared.
Freelancer or consultant
Tell clients at the start that you are recording, get a yes on the call, and choose a tool that joins without a bot if a client objects to one.
IT or legal lead
Review retention, model-training and deletion terms in the vendor's data processing agreement, and set a rule for calls that must never be recorded.
Tools for this task, by pricing model
Free
| Tool | Pricing model | Free tier | Starting price | Compliance claims | Official site | Last verified |
|---|---|---|---|---|---|---|
 Granola | Free | Not verified | Not verified | Not verified | Online (redirects) (2026-10-01) | Not yet |
Freemium
| Tool | Pricing model | Free tier | Starting price | Compliance claims | Official site | Last verified |
|---|---|---|---|---|---|---|
 tl;dv Review: Meeting Notes, Consent, Privacy, and Security | Freemium | Not verified | Not verified | Not verified | Online (2026-09-26) | Not yet |
 Transcri | Freemium | Yes: Free plan, no credit card required | Not verified | Not verified | Online (2026-09-29) | 2026-10-01 |
 Vatis Tech | Freemium | Not verified | Not verified | Not verified | Online (2026-09-30) | Not yet |
Typical cost
Otter.ai Pro is $16.99 per user per month ($8.33 billed annually) and Business is $30 ($19.99 annually); the free plan gives 300 transcription minutes a month and 30-minute meetings. Fireflies Pro is $18 per seat monthly or $10 billed annually, and Business is $29 monthly or $19 annually. Granola Business is $14 per user per month, and its free plan keeps 30 days of meeting history.
Seat prices add up across a team, and meeting-length caps on free plans matter more than the headline fee. Prices were checked in October 2026.
Data & compliance requirements
General rules that apply to most AI tools, whatever the category. Check which ones apply to your use.
| Name | Applies to | Official source | When it applies |
|---|---|---|---|
| EU AI Act (transparency obligations, Art. 50) | EU | digital-strategy.ec.europa.eu | Users must be told when they are interacting with a chatbot or viewing AI-generated content |
| GDPR | EU/EEA (and non-EU services targeting EU users) | commission.europa.eu | Processing personal data, including personal data in training data |
| US Copyright Office — AI & copyright guidance | US | copyright.gov | Copyright ownership and registration of AI-generated content |
| FTC Act §5 & Endorsement Guides | US | ftc.gov | Claims about AI capabilities; AI-generated reviews or endorsements |
| COPPA | US | ftc.gov | The product may be used by children under 13 |
| NIST AI Risk Management Framework | US (voluntary) | nist.gov | Reference for risk assessment when buying or deploying AI |
| ISO/IEC 42001 (AI management system) | International (voluntary) | iso.org | A vendor claims an AI-governance certification |
This list is a starting point for your own compliance review, not legal advice. Last reviewed 2026-09-24.
Failure modes
- Wrong names, numbers or owners in action itemsSummaries compress speech, and speech models can invent text: about 1% of Whisper transcriptions in one study contained fabricated phrases. Check names, figures and deadlines against the transcript or recording before the notes are shared.Source
- A notetaker joins without everyone's consentPlaintiffs allege Otter's bot recorded and transcribed people, including non-users, without consent; wiretap and privacy claims survived dismissal in August 2026. Say at the start that an AI notetaker is running, get a yes in the invite or on the call, and turn it off for sensitive meetings.Source
- Recordings kept or used to train modelsThe court found it plausible that Otter retains and uses recordings for its own commercial purposes, including model development. Read the retention, training and deletion terms and the data processing agreement before use, and set a retention period.Source
- Some speakers are transcribed worse than othersFive commercial systems had a 0.35 average word error rate for Black speakers against 0.19 for white speakers (2019 versions). Test the tool on your own team's calls and spot-check transcripts for the speakers it handles worst.Source
- Free-plan limits cut off long meetingsOtter's free plan allows 300 transcription minutes a month and 30 minutes per conversation. Compare plan minute and duration limits with your meeting lengths before relying on a free tier.Source
Recent changes: tools for this task
Last verified: 2026-10-03 · auto-checked 2026-10-09 · Reviewed by Toolsfine editorial
- Notta · AI alternative note · — → A general chatbot can summarize a tra… ·
- Notta · editor verdict · — → Meeting and file transcription tool w… ·
- Notta · integrations · — → ["Zoom","Google Meet","Microsoft Team… ·
- Notta · compliance · — → [{"name":"SOC 2 Type II","status":"cl… ·
- Notta · pricing page URL · — → https://www.notta.ai/en/pricing ·
- Notta · free tier · — → {"exists":true,"limits":"120 transcri… ·
- Notta · pricing model · — → Freemium ·
- n8n · AI alternative note · — → A general LLM cannot run scheduled wo… ·