Lawyers Are Shutting Off AI Notetakers Mid-Meeting. The Problem Isn't AI — It's the Vendors
By Lexi Banks · · AI & Technology
An AP report found lawyers evicting AI notetakers from meetings over data privacy. Here are the 4 questions to ask any vendor — and how Mira answers each one.
Key takeaways
- A July 9, 2026 AP report by Cathy Bussewitz found professionals — including attorneys and HRCI CEO Amy Dufrane — refusing AI notetakers outright, because vendors may train models on transcripts, resell meeting data, build voiceprints without consent, and retain metadata indefinitely.
- The legal exposure is real and current: in February 2026 a New York federal judge ordered a defendant to turn over materials shared with Anthropic's Claude, treating the AI as a third party that can waive attorney-client privilege. Illinois' BIPA already treats voiceprints as biometric identifiers requiring notice, consent, and a documented destruction schedule.
- Banning the category is the wrong fix. The AP's reporting effectively hands buyers a four-question vendor checklist: do you train on my calls, do you sell my data, what's your retention and destruction schedule, and what speaker ID are you doing?
- Mira answers all four in its published privacy policy — no training on your call data, no selling or renting personal information, 7/30-day recording retention with deletion within 30 days of an account-deletion request, named subprocessors, AES-256 at rest and TLS 1.2+ in transit, plus SOC 2 Type II and GDPR compliance (HIPAA, SSO, SAML and a DPA on Enterprise).
The people telling you to shut off the AI notetaker are the ones who'd get sued first
On July 9, 2026, the Associated Press published Cathy Bussewitz's report on the quiet revolt against AI notetakers — the bots that join your Zoom, transcribe everything, and email a recap before you've closed the tab.
The pushback isn't coming from technophobes. It's coming from lawyers and compliance chiefs.
Amy Dufrane, CEO of HR certification body HRCI, didn't hedge: "There are huge risks to the organization on AI notetakers. I don't think companies should use it at all."
Justin Daniels, a corporate attorney at Baker Donelson, has a rule for any call with a bot in it: "I won't start talking about anything substantive until it's shut off, because I just don't want to take the risk."
Read that back. Professionals are now stopping meetings to evict the software that was sold to them as a time-saver.
The problem was never the transcription. It's that nobody can tell you where it goes.
Daniels put his finger on the actual issue: "People who use AI notetakers, they don't always know where the data goes."
That's not paranoia. The AP spells out exactly what "somewhere else" can mean:
- Training data. Some vendors use confidential meeting transcripts and recordings to train their AI models. Your pricing floor, your severance conversation, your term sheet — now weights in somebody's model.
- Resale. Some companies resell the data collected by the notetaking tools they built.
- Voiceprints. Speaker ID works by building an acoustic signature of your voice — a biometric, closer to a fingerprint than a username. Danielle Kays, a privacy and employment partner at Fisher Phillips, says: "If there is some sort of speaker ID or voice recognition, really understand what that is and how it works."
- Forever. Metadata can sit in a vendor's systems indefinitely, even after you've deleted the content.
Thorin Klosowski, senior security and privacy analyst at the Electronic Frontier Foundation, explains why the transcript is the real asset: "Storing a bunch of video isn't easy, it's costly and hard to look through, but text is much easier to search and cheaper to store."
Text is cheap, permanent, and searchable. That's the entire pitch — for you, and for anyone else who ends up holding it.
The privilege problem already happened
In February 2026, a New York federal judge ordered a criminal defendant to hand prosecutors documents he had created for his own attorneys — because he'd shared them with Anthropic's Claude. Share privileged material with a third party and you can waive the privilege. The court treated the AI as exactly that third party.
Daniels' version: "if the data goes anywhere else and they're not aware of it, that attorney-client-privileged conversation may not be attorney-client-privileged anymore."
And Illinois isn't waiting. Its Biometric Information Privacy Act already classifies voiceprints as biometric identifiers, requiring written notice and informed consent before collection — plus a documented retention schedule and destruction policy. BIPA is the statute behind some of the largest privacy settlements in the country.
The bill for "we let a bot sit in on everything" is already being written.
"Don't use it at all" is the safe answer and the wrong one
Dufrane's advice is defensible. It's also a tax on your own team.
The value of reviewing sales calls is not imaginary. It's how a new rep ramps in weeks instead of quarters, how a manager coaches on what was actually said instead of a rep's memory of it, and how a forecast stops being vibes. Ban the category outright and you don't get safety — you get managers guessing, and reps rehearsing the same mistake for another six months.
The AP didn't just publish a warning, anyway. Read it again and it published a procurement checklist.
Four questions. Ask them in writing.
Strip the anxiety out of the article and you're left with the questions any AI that joins your calls should answer before it's allowed in:
- Do you train your models on my recordings or transcripts?
- Do you sell, rent, or share my data with third parties?
- How long do you keep recordings, transcripts, and metadata — and what's the destruction schedule?
- What speaker identification are you doing, and how does it work?
Chris Pluymers, an attorney at The Dillon Law Group, argues you can't punt on this: "In the world of AI, the world of data and privacy, the world of biometric identification, I don't think you can have such a lax approach to it. I think getting out ahead of it is crucial."
If a vendor won't answer those four in writing, you already have your answer.
How Mira answers them
Mira is Kalyxi's AI sales coach. It records, transcribes, and scores sales calls — so every one of the AP's questions lands squarely on us. Here's where we stand, and all of it is in our privacy policy, not a sales deck:
We don't train on your calls. In our own words: "We do not use your call data to train AI models." Your discovery calls don't quietly become model weights.
We don't sell your data. Also in writing: "We do not sell, rent, or trade your personal information or Google user data to third parties." No resale, no brokers.
We delete on a schedule. Recordings are retained for your plan's window — 7-day or 30-day access — then automatically deleted from our recording provider. Ask us to delete your account and your personal data is gone within 30 days. That's a documented retention schedule and destruction policy, which is precisely what BIPA asks for and what the AP found most vendors won't commit to.
We name our subprocessors. Supabase, Recall.ai, OpenAI, Resend, Vercel. You don't have to wonder where the data goes — it's listed.
We encrypt it. AES-256 at rest, TLS 1.2+ in transit.
We're audited, not self-certified. SOC 2 Type II and GDPR compliant. Enterprise adds HIPAA, SSO, SAML, audit support, and a DPA.
On question four — the one Kays says everyone forgets — Mira produces speaker-separated transcripts in 27 languages so you can see who said what. If you want to know how that works before you turn it on, ask us and we'll walk you through it. That's the standard the AP is really arguing for: not "trust us," but "here's the answer."
Scope is a security feature
Most of the AP's anxiety comes from a bot that lurks in every meeting — HR one-on-ones, board sessions, privileged legal calls — vacuuming up everything and reporting to no one in particular.
Mira isn't that, by design. It does one job: make your sales team better. It records sales calls from Zoom, Meet, or Teams, scores them line-by-line against MEDDIC, SPIN, BANT, GAP, NEAT, or a rubric you upload yourself, and cites the evidence behind every score with a confidence rating you can audit. Reps get a "what worked, what didn't" report in minutes. Managers get skill heatmaps instead of hunches.
A tool with a defined job has a defined data footprint. That's the difference between "we record everything, trust us" and a policy you'd be comfortable forwarding to your general counsel.
The takeaway
The notetaker backlash isn't a rejection of AI. It's a rejection of vendors who won't say where the data goes.
Dufrane says don't use them at all. We think that's the right instinct aimed at the wrong target: don't use the ones that can't answer four questions in writing. Ask them. Ask us. Then pick the tool whose data policy you'd be fine reading aloud in a deposition.