Is AI Call Transcription Legal for Business? What the 2026 Notetaker and Voice AI Lawsuits Mean

Short answer: Yes. AI call transcription is legal for a business when everyone on the call knows a transcript is being made, the people in all-party consent states agree to it, and the transcription vendor isn't keeping its own copy of the conversation for its own purposes. That last condition is the one most businesses miss. Every major lawsuit against AI transcription in 2026 (Otter, Granola, Fireflies, SoundHound) turns on the same question: did a third-party software company hear the call, and what did it do with the words afterward? A plain "this call may be recorded" may not cover that.

We build AI phone systems at Wrk Less, mostly Retell voice receptionists with n8n moving transcripts into each client's CRM. Every one of those builds creates a transcript, so we've had to settle what gets transcribed, who hears it and where the text ends up. Our AI call recording laws guide covers recording consent for AI receptionists. This post is about the transcript: meeting notetakers, AI notes in your phone system, and the voice AI that answers your line.

This is not legal advice. Have your attorney approve your notices and vendor contracts.

Is transcribing a call the same as recording it?

For wiretap purposes, mostly yes. Laws like Florida's apply to anyone who "intercepts" a communication, meaning they acquire its contents. They don't only cover saving an audio file. Florida makes that lawful only when "all of the parties to the communication have given prior consent" (Fla. Stat. § 934.03). A live transcript acquires the contents as the words are spoken, so "we only keep the text, not the audio" doesn't take you outside the statute.

Federal law is looser. The Wiretap Act lets one party consent, but not when the interception is for a criminal or tortious purpose (18 U.S.C. § 2511(2)(d)). Plaintiffs now use that exception against transcription vendors themselves, as the Otter case shows below.

Why are AI transcription companies getting sued in 2026?

Because a transcription tool can look like a third party listening in, not a tape recorder. Four cases show the pattern. All four are allegations or pleading-stage rulings, not findings of liability.

• In re Otter.AI Privacy Litigation (N.D. Cal.). On August 13, 2026, Judge Eumi K. Lee largely denied Otter's motion to dismiss. Federal wiretap, CIPA § 631, Illinois BIPA, unjust enrichment and unfair competition claims went forward. Otter argued the one-party exception, and the court answered that "committing a tort and seeking a profit are not mutually exclusive" (National Law Review summary). Otter filed an answer on September 17 denying the core allegations. Class certification isn't due until May 13, 2027 (October status roundup).
• Chamberlain v. Granola (N.D. Cal.). Granola runs without a visible bot and captures audio from the user's own computer. The complaint quotes Granola's marketing line that "[o]ther people in the room won't know," and alleges that Granola trains on meeting content by default (complaint). The theory is simple. Participants who don't know a transcriber is there can't consent to it.
• Fricker v. Fireflies.AI (N.D. Ill.). This is a biometric case, not a wiretap case. It alleges that Fireflies' speaker recognition created voiceprints of meeting participants who never signed up (consolidated complaint). The motion to dismiss was fully briefed on August 26 and is still undecided.
• Thompson v. SoundHound AI (Alameda County Superior Court, filed July 30, 2026). This is the closest case to an AI receptionist. A caller phoned a restaurant, a voice AI answered, and the complaint says the call was transcribed in real time and kept to train SoundHound's models (CIPAWorld breakdown). The restaurant isn't a defendant. The complaint argues that the "will be recorded" notice the caller heard said nothing about a third-party AI company receiving and using the call (ZwillGen, Oct 22, 2026).

In California, CIPA § 637.2 allows $5,000 per violation without proof of harm (analysis of the SoundHound complaint). That's why a single phone line with heavy call volume draws plaintiffs' firms.

Does California's new CIPA law fix this?

No. SB 690, signed September 30, 2026, cuts off private lawsuits only for pen-register and trap-and-trace claims about website and app tracking. The wiretap section (§ 631) and the confidential-recording section (§ 632), the ones used in the Otter and Granola cases, were dropped from the final bill. Transcribing a California caller still requires everyone's consent.

Which kinds of AI transcription carry the most risk?

Transcription type	Who usually knows it's running	Main legal risk	What to set before you rely on it
Meeting notetaker bot (joins Zoom, Teams or Meet)	The host. Guests see a bot name, if they notice it	Vendor training on meetings, BIPA voiceprints from speaker ID	Turn off training for every seat, not just yours. Announce the bot out loud
Bot-free notetaker (captures your mic and speakers)	Only the user	No notice to anyone else, so no consent from them (the Granola theory)	Turn on any participant notice the tool offers and say it at the start of every call
Phone-system AI notes or summaries	Whoever turned the feature on	All-party consent on calls with FL, CA, WA and similar callers	Update the greeting so it says calls are recorded and transcribed
AI receptionist / voice AI on your main line	Nobody, unless the first line says so	Generic "may be recorded" notice that never mentions AI processing (the SoundHound theory)	First line names the business, says it's an AI assistant and that the call is recorded and transcribed

One platform change makes the meeting rows matter more this month. Microsoft's Teams update MC1459141 adds an admin policy to automatically block identified external meeting bots. It started reaching general availability in early October 2026 and is off by default. If you sell to larger companies, expect your notetaker to get bounced from their meetings, and don't switch to a silent bot-free tool to get around it. That's exactly the fact pattern in the Granola complaint.

How we handle transcripts on a Retell build

The transcript is the most useful thing a voice AI produces. It drives the CRM note, the callback task and the summary text to the owner. Here's the checklist we work through on every build before the line goes live:

• The greeting says "transcribed," not just "recorded." Our first line names the business, says the caller is talking to an AI assistant, and says the call is recorded and transcribed. That language covers the step the SoundHound complaint says was missing. The AI-identity side is covered in our disclosure laws post.
• Vendor training terms are read and handled in writing. Retell's Terms of Service (Section 11) say that if you don't opt out of recording, Retell may use call data for "the development, training, and improvement of artificial intelligence and machine learning models that are included in the Service," after de-identification and aggregation (Retell terms). Under the third-party theory in Otter and SoundHound, that clause matters. We take it up in the client's DPA and run the same check on every tool that touches the call: speech-to-text, the language model and the CRM.
• Storage is set on purpose. Retell lets each agent store Everything, Everything except PII, or Basic Attributes Only. With storage switched off, raw transcripts and recordings are deleted about 10 minutes after post-call processing (Retell data storage docs). So our post-call webhook has to grab what the client needs right away and write it to the CRM.
• Redaction matches the job. Retell's PII Redaction launched October 19, 2026 on all plans. It removes names, addresses, dates of birth, passwords and PINs from transcripts and recordings. A plumber needs the service address, so we redact card numbers and dates of birth and leave addresses alone.
• Diarization yes, voiceprints no. Labeling "caller" and "agent" inside one call is fine. Recognizing a person by voice across calls is the BIPA problem in the Otter and Fireflies cases. We identify repeat callers by phone number and CRM lookup. Our caller memory post covers how.
• A real path for "no." If a caller says "don't record me," the call goes to a person or a scheduled human callback. See transfer to a human.
• Proof of what the caller heard. The webhook logs the greeting version next to the transcript in the CRM.

Want us to run this checklist on your current phone or notetaker setup? Book an AI systems assessment.

What should a small business do this week?

1. List every tool that transcribes. Include meeting notetakers on staff laptops, AI notes in the phone system, call-tracking software and your AI receptionist. Shadow notetakers on personal accounts count too.
2. Find each vendor's training setting. Turn off training on customer conversations for every user, or get the no-training commitment in writing.
3. Fix the words callers hear. "Recorded and transcribed, including by AI," spoken before any personal details are collected.
4. Keep less. Store the summary and the fields you need, set a retention period, and delete raw transcripts you'll never read.
5. Keep transcription consent separate from texting consent. Agreeing to a transcript isn't agreeing to follow-up texts. See our missed-call text-back guide and FCC opt-out rules.

Bottom line

AI transcription is too useful to give up, and you don't need to. The 2026 lawsuits aren't about transcripts existing. They're about people not knowing a transcriber was there, and about vendors keeping the words for themselves. Say it up front, close the vendor-training gap in writing, and keep only what the business uses.

Related

• Recording consent for voice AI: AI call recording laws
• What a compliant setup costs: AI receptionist pricing 2026
• How we set up voice AI: Retell AI voice agent setup
• Product page: AI receptionist for small businesses

Get your transcription setup reviewed

We'll check your greeting, vendor training terms, storage and redaction settings, and CRM logging, then fix what's off. Book your AI systems assessment.