合規團隊為何禁止機器人入會型 AI 會議助手：2026 年私隱風險指南

1. The EU AI Act Has Turned Meeting Privacy Into a Legal Deadline

€15 million or 3% of worldwide revenue. That's the exposure the EU AI Act created for organizations whose meeting assistants record calls without demonstrable disclosure — and it became enforceable on August 2, 2026.

The tools aren't the problem. One design choice is: a bot that joins the call as a participant and sends the audio to a third-party server. Compliance teams are banning that architecture, not AI itself.

2. Three Capture Models — and Why the Difference Is Legal, Not Cosmetic

A meeting bot does not just listen. It receives the same audio stream the platform sends to human attendees. Then it sends that stream to infrastructure the meeting host does not control.

Not all assistants work this way.

"Botless" does not mean private. A browser extension can capture locally and still upload everything to a cloud with unlimited retention and training rights. Privacy depends on three questions, not on whether a tile appears:

1. Where does the audio go after capture?
2. How long does it stay there?
3. What is the vendor allowed to do with it?

A tool with a visible bot and strict zero-retention can be safer than an invisible extension that trains models on your client calls.

3. The Legal Cascade: What Breaks When a Bot Is in the Room

If you're not a lawyer, here's the headline: a bot in a call creates at least three independent legal problems. Winning on one does not protect you from the others. Each can cost money on its own.

Consent and wiretap

US federal law needs one party's consent. Thirteen states need everyone's: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Michigan, Montana, Nevada, New Hampshire, Pennsylvania, and Washington. A bot joins because the host asked — but every other participant never agreed.

That's the theory in Brewer v. Otter.ai (four consolidated class actions, Aug–Sep 2025). On August 13, 2026, the court denied the motion to dismiss on core wiretap, California privacy, and Illinois biometric claims — finding it plausible Otter used conversation data for its own commercial purposes.

"The host consented" is no longer a defence that ends the case.

Biometric privacy (BIPA)

Speaker identification needs voiceprints — biometric data under Illinois' BIPA, which requires written notice, written consent, and a published retention policy first.

Cruz v. Fireflies.AI Corp. (filed Dec 18, 2025): an Illinois resident joined a normal meeting, never signed up for any AI tool, never consented. A bot joined at someone else's request, recorded the call, named speakers, produced attributed transcripts.

Why employers care: BIPA lets multiple parties be liable for the same collection. If any participant was in Illinois when recorded, BIPA applies regardless of where you're based.

Privilege and the third-party problem

Privilege depends on confidentiality. Put a vendor in the room, and opposing counsel has an argument. You do not need a human at the vendor to have listened. You need the vendor to have access under its terms.

4. Why "Just Configure Zoom or Teams" Falls Short

Platform controls have improved. But they all share the same limit: they govern admission to the room.

Microsoft Teams added "Unverified" bot labels (June 2026) and auto-blocking of detected external bots (August 2026). Zoom requires host/co-host to start AI Companion. Google Meet gates "Take notes for me" by org unit.

All three share a blind spot: a device-level tool never asks for admission. It captures system audio on the user's machine. No lobby entry, no bot signal, no tenant policy applies. Microsoft's own documentation concedes this — auto-blocking depends on detecting a participant as a bot and does not prevent local recording or screen capture.

The second gap: bans push usage into shadow channels. Block the approved tool and employees connect a personal Otter account to their work calendar. That bot arrives through an identity the admin doesn't manage. Prohibition without an approved alternative produces concealment, not compliance.

5. The Architectural Shift: Device-Level Capture

The question that matters is not "AI or no AI." It is: does the assistant become a participant in the conversation?

A bot-based assistant enters the room — triggering the consent, privilege, and biometric problems above. A device-level assistant does not enter the room. It captures audio on the user's device and adds no participant. The platform sees nothing join.

That difference is legal, not cosmetic. If there is no third party in the conversation, the wiretap argument and the privilege-waiver argument weaken by design.

CatchMeet is one example of this architecture: it runs on the user's computer, captures system audio from Zoom/Meet/Teams/Slack/Discord, adds no participant, stores data locally, and routes any processing through providers under zero-retention agreements.

The honest caveat: device-level is not a privacy guarantee. A botless tool can still upload everything to a cloud with unlimited retention and training rights — at which point the core risk returns through a different door. And device-level capture does not eliminate consent law: in an all-party-consent state, recording without disclosure is still a violation regardless of how the audio was captured. What device-level removes is the third-party-in-the-room problem — the specific trigger for the privilege, eavesdropping, and BIPA arguments above.

6. Compliance Checklist: What to Do This Week

Steps 1–5 are inventory. Steps 6–8 are the ones most organizations skip — and the ones that decide whether the policy survives contact with employees.

7. FAQ

Are AI meeting assistants illegal under the EU AI Act?
No. Article 50 requires transparency — disclosure and labelling — not a ban. Fines reach €15M or 3% of turnover for non-compliance. The burden falls on the deploying organization.

What makes bot-based assistants riskier?
They introduce a third party into the conversation. That single feature triggers consent, privilege, and biometric rules simultaneously. See Section 3.

Does "botless" mean private?
No. It describes capture method, not data flow. Three questions decide privacy: where audio goes, how long it stays, and what the vendor can do with it. See Section 2.

Which US states require all-party consent?
Thirteen: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Michigan, Montana, Nevada, New Hampshire, Pennsylvania, and Washington.

Can a bot waive attorney-client privilege?
It can create the argument. A vendor with contractual access to transcripts is a third party. The Brewer v. Otter ruling makes this argument credible.

What is the BIPA risk for employers?
BIPA allows multiple parties to be liable for the same biometric collection. An employer that selects or benefits from the tool can be sued alongside the vendor. See Section 3.

Can blocking external bots in Teams or Zoom solve this?
Partially. Those controls detect bots that try to join as participants. They don’t cover device-level capture, local recording, or personal accounts the admin doesn’t manage. See Section 4.