1 in 3 US workers met an AI notetaker.

The notetaker joined before anyone asked

AI notetakers spread through meetings faster than the rules about them. One in three US workers, 33.4%, say an AI notetaker or transcription bot has been in their work meetings, according to a survey published by Stacker in July 2026. Another 22.4% cannot say whether they have been recorded. The survey was commissioned by a law firm, Kolmogorov Law, and run through Pollfish among 500 employed US adults; it is small and has an interest in the topic. Its direction is still hard to ignore.

Among the workers who had met a notetaker, only 34.7% say they were always asked first. 36.5% were asked only sometimes, 25.1% were never asked and 3.6% found out afterwards. That adds up to 65.3% for whom consent was inconsistent or missing. 18.8% of all workers have discovered a recording they did not know about.

65% of US workers who met an AI notetaker say consent was inconsistent or missing (Kolmogorov Law, Jul 2026)

Source: Kolmogorov Law survey via Pollfish, 167 exposed respondents, fielded 8 Jul 2026, published by Stacker 16 Jul 2026.

Most people do not know the law applies

Awareness lags use. 41.8% of US workers had no idea that recording laws apply to work meetings, and only 35% knew that some states require the consent of every participant. 64.4% did not know that pasting confidential data into a personal chatbot can be illegal. Shadow AI in meetings is not a fringe problem; it is the default.

41.8% of US workers did not know that recording laws apply to work meetings (Kolmogorov Law via Stacker, Jul 2026)

The first test case is in court. In August 2025 a plaintiff filed a class action against Otter.ai in the federal court for the Northern District of California, NPR reported. The suit alleges that Otter's notetaker can join meetings without the consent of participants and that recordings are used to train its AI without clear notice. Otter denies the allegations and the case is open. The company says about 25 million people use its tools and that it has processed more than one billion meetings since 2016. Scale like that turns a consent default into a legal question.

In Europe the bar is higher

European law starts from a stricter place. In Germany, recording the privately spoken word without consent is a criminal offence under section 201 of the Criminal Code. A meeting bot that records by default is therefore not a convenience setting but a legal risk. Where a tool can monitor employees' behaviour or performance, the works council has a right of co-determination under section 87 of the Works Constitution Act, so introducing a notetaker across a company usually needs a works agreement.

The EU AI Act adds two layers. Since 2 February 2025, emotion recognition in the workplace is prohibited under Article 5, with narrow medical and safety exceptions. Meeting analytics that score mood, engagement or sentiment from faces or voices need careful review against that ban. Since 2 August 2026, under Article 50 of the AI Act, people must be told when they interact with an AI system, and providers of systems that generate synthetic audio must mark it in a machine-readable way. Systems already on the market have until 2 December 2026, under the Digital Omnibus on AI.

What a consent-first setup looks like

Consent is mostly a configuration question, and most tools can be set up properly. A consent-first setup has five parts. The bot announces itself by name when it joins, in the meeting and in the invitation. Every participant can decline, and the host can remove the bot without technical help. Recording, transcript and summary are stored in a region and system the organisation controls, with a fixed retention period. The provider's right to use recordings or transcripts for training is switched off or excluded by contract. And external participants, such as customers, candidates or patients, get the same notice as colleagues, because they are the people most likely to object and least likely to know the tool.

None of this slows a meeting down. It does change the buying process: the questions above belong in the tender, not in a later policy. A tool that cannot meet them should not join a meeting with customers or job candidates. Teams that settle these points once, at company level, avoid a hundred individual decisions by people who simply clicked "add to calendar".

Voice is now an attack surface

The same speech technology that writes the minutes can clone a voice. Cyber incidents remain the top business risk in the Allianz Risk Barometer 2026 at 42%, and AI has climbed to second place at 32%. Voice impersonation of executives is a fraud vector for payment approvals, password resets and supplier bank changes. The ETHORITY analysis of synthetic faces and voices sets out the rule: a voice or video call alone should never approve a payment.

Language is where the value is

None of this argues against the category. Transcription, summaries, translation and dubbing are among the most useful AI jobs in daily work. They make meetings searchable, include people who missed them and open content to other languages. The Good AI Prism maps them in Meetings, Voice & Language, next to Customer Service for voice agents and Media & Entertainment for dubbing and versioning.

Language coverage depends on data. Mozilla Common Voice on its card holds 33,815 hours of speech in 137 languages under CC0, including low-resource languages, in its 22.0 release of June 2025. European models such as Aleph Alpha's Pharia-1-LLM cover English, German, French, Spanish, Italian, Portuguese and Dutch, but only under a research licence. For a multinational, the practical question is which languages a tool handles well enough to put in the minutes.

The Good AI Prism Global Market Study names the two outcomes to measure in this area: meeting time saved and consent compliance. The second is the one most pilots forget.

In practice

Use case: a lawsuit as a buying checklist. The August 2025 class action against Otter.ai alleges two things: a bot that joins without every participant's consent, and recordings used for training without clear notice. Otter denies the claims. Both allegations map onto settings a buyer controls: announcement and opt-out in the meeting, and a contractual exclusion of training.

Scenario: an interview with an external candidate. A recruiter in Germany adds a notetaker to a video interview. In a consent-first setup the invitation names the bot, the candidate can decline without penalty, the transcript stays in the company's EU tenant for a fixed period, and no mood or engagement score is produced. Without those settings the interview touches section 201 of the Criminal Code and the EU ban on emotion recognition at work.

What to do Monday

1. Find every bot. List the notetakers, transcription and translation tools in use, including those individuals add to calendars, and where recordings and transcripts are stored.
2. Turn consent on by default. Configure tools to announce themselves and ask every participant, and give people a simple way to decline.
3. Agree it with the works council. In Germany, a company-wide notetaker normally needs a works agreement; settle storage, access and deletion there.
4. Switch off emotion scoring. Review meeting analytics against the EU ban on emotion recognition at work, in force since 2 February 2025.
5. Separate voice from approval. Add a second, non-voice check to payment approvals, password resets and bank-detail changes.