The AI Notetaker That Stayed On After the Firing: What Marathon Engineering Teaches Every Employer About HR Calls

Published September 15, 2026

September 15, 2026 · 10 min read

If your company uses a cloud AI notetaker on video calls, there is a specific fact pattern from 2026 you should have taped to your monitor. It has nothing to do with the technology being wrong. It has to do with the technology doing exactly what it was designed to do — record the meeting, transcribe it, and email the transcript to participants — during a moment nobody in the room thought was still “the meeting.”

The Marathon Engineering matter is that fact pattern. And read alongside the August 13, 2026 federal ruling in In re Otter.AI Privacy Litigation, it changes what a reasonable employer should assume before it lets a bot into any HR, termination, investigation, or privileged conversation.

What actually happened at Marathon Engineering

According to a complaint filed in New Jersey Superior Court and reported by Law360 on August 14, 2026, environmental scientist Lindsay Waninger was terminated on a February 3, 2025 video conference with three Marathon Engineering & Environmental Services representatives after fewer than three months on the job. An administrator had activated Fireflies.ai to record and transcribe the meeting.

The termination itself was, by the complaint’s account, brief and vague — “not working out.” Waninger logged off. But The Employer Handbook’s coverage of the complaint notes what happened next: the three remaining representatives stayed on the call and continued talking, and Fireflies kept transcribing. As Weinstein + Klein summarized, the conversation turned to what the ideal replacement looked like — with one supervisor allegedly volunteering that the ideal candidate would “[h]opefully [be] a relatively strapping young man.”

Then Fireflies did what Fireflies is supposed to do. It emailed the full transcript, post-termination commentary included, to the recently-terminated employee. She used it to file a gender-discrimination suit under the New Jersey Law Against Discrimination.

The failure mode in one sentence

The bot leaves when the host ends the meeting, not when the agenda does.

That single sentence, drawn from independent analysis of the Marathon and Otter events read together, is the entire lesson. Nobody at Marathon needed to want to record what came after. The default did the work.

Why the timing of the Otter ruling matters

One day after the Waninger complaint was filed, on August 13, 2026, Judge Eumi K. Lee of the U.S. District Court for the Northern District of California ruled on Otter.ai’s motion to dismiss in the consolidated class action In re Otter.AI Privacy Litigation, No. 5:25-cv-06911-EKL. According to UC Today’s reporting, the court granted the motion in part but allowed the federal Wiretap Act, California Invasion of Privacy Act, and Illinois biometric-privacy claims to proceed toward discovery.

The reasoning matters more than the result. As Lawsuit Intelligencer’s analysis of the order explained, the CIPA Section 631 holding did not turn on whether the notetaker was invited. It turned on the fact that Otter retains recordings and uses them to train its own models — which took the vendor outside the line of cases treating a service provider as a mere extension of its client. The court in effect treated the AI notetaker as a third-party eavesdropper for pleading purposes.

Read together, the two events do something that has not been true before: they make the employer’s downside real on both sides of the equation. On one side, the transcript can become the plaintiff’s evidence (Marathon). On the other side, the fact that the vendor retained and reused the recording is itself an actionable theory that runs against both vendor and customer (Otter). We covered the specific procedural mechanics of that ruling in our explainer on the In re Otter.AI Privacy Litigation hearing.

The two liabilities you are actually stacking

When an employer runs a cloud AI notetaker on an HR call, it takes on two distinct exposures at the same time. Most policy discussions only address one.

The Workplace Privacy Report’s analysis of Cruz v. Fireflies pointed out that HR interviews, witness interviews, and applicant interviews are exactly the venues where voiceprint capture accumulates fastest — because the same conference room hosts many different non-employees in short succession.

Why the transcript is discoverable even if it is “wrong”

A common in-house reaction to the Marathon story is: “We’ll challenge the AI transcript on reliability grounds.” That is a real argument, but it is not a shield. Writing at Mondaq, employment counsel noted that a Frye or analogous reliability hearing on the transcription tool itself is likely — but that even if the transcript is ultimately inadmissible, its existence will need to be disclosed during discovery, will identify additional witnesses, and can create damaging factual disputes.

This dovetails with what NJBIZ’s March 2026 e-discovery analysis put plainly: AI-generated meeting transcripts are electronically stored information (ESI), and once litigation is reasonably anticipated they are subject to the same preservation rules as emails, contracts, and other business records. We walked through the full spoliation and litigation-hold implications in our deep dive on AI meeting transcripts as discoverable evidence.

The consent problem cloud notetakers cannot solve for you

Vendors have responded to the litigation wave by adding disclaimers, banners, and account-level toggles. None of them cure the structural issue. As the National Law Review noted about Otter’s configuration, non-user attendee notifications are gated behind the most expensive plan tier, and default configurations do not notify non-users that they are being recorded. The vendor’s architecture assumes the account holder handles consent; the account holder assumes the vendor does.

The Wealth Solutions Report distilled the practical implication: only the account holder consented. Otter allegedly captured the other participants’ voices and used them to train its AI models even though those participants consented to nothing. In an HR call — where the “other participant” is often the person about to be fired, or a witness in an investigation — that gap becomes the entire lawsuit.

Why HR calls are the highest-risk surface

Employment counsel writing at SocialTalent’s coverage of the notetaker suits made a point worth internalizing: interview recordings carry higher legal risk than internal meetings because candidates are external, protected characteristics surface naturally in conversation, and recordings can later become evidence in hiring disputes. The same logic applies with even more force to terminations, disciplinary meetings, accommodation discussions, and internal investigations.

The Skoler Abbott analysis of the Waninger complaint captured the executive summary in one phrase: your AI note-taker could be your former employee’s star witness. Even if you consent to the notetaker attending a meeting with your own attorney, you may not be able to guarantee that the conversation stays confidential once the transcript exists on a third-party server.

Meetings that generally should not have a cloud AI notetaker on them

• Terminations, layoffs, RIF conversations
• Formal disciplinary meetings and performance improvement plan discussions
• ADA accommodation and FMLA leave discussions
• Harassment, discrimination, or retaliation investigation interviews
• Whistleblower intake calls
• Attorney-client privileged strategy sessions
• Compensation, offer, and separation-agreement negotiations

These are the meetings where your general counsel and CHRO — not the vendor’s marketing page — decide the recording rules.

Cloud AI notetaker vs on-device transcription for HR calls

How Basil AI solves this — the honest version

Basil AI is a fully on-device meeting recorder and transcription app for iPhone and Mac. It uses Apple’s Speech framework and the Apple Neural Engine to transcribe audio locally. There is no vendor server that holds the recording, no default email to participants, no training pipeline that ingests the audio, and no bot that “stays in the meeting” after you close the app or end the call.

That eliminates the specific failure modes in Marathon and Otter:

• No auto-emailed transcript. Sharing is a deliberate export action by the device owner, not a default vendor behavior.
• No “kept listening” problem. Recording stops when the device owner stops it, not when a host in a different app ends a meeting.
• No vendor retention, no training use. There is no vendor server holding the audio in the first place, so there is no retention theory to plead against the vendor.
• No third-party voiceprint extraction. Speaker labeling runs locally with Apple’s frameworks; nothing leaves the device.

What Basil AI is not is a substitute for your employment counsel’s call on whether an HR meeting should be recorded at all. On-device processing is an architecture fact, not a compliance guarantee. Your CHRO and general counsel still decide which meetings warrant recording, which never should be recorded, and what your written policy says. Basil AI removes the vendor-server risk vector; it does not remove your obligation to have a policy. For the broader picture on which meeting-notes architecture fits regulated work, see our compliance-officer buyer’s guide for financial services.

A minimum-viable AI-notetaker policy for HR calls

Every employer that has deployed any AI notetaker — cloud or local — should be able to answer these five questions in writing, reviewed by employment counsel:

1. Which meetings are off-limits? Enumerate the categories above. Do not rely on judgment in the moment.
2. Who is authorized to enable recording, and how is that enforced? Administrative gating, not honor system.
3. What is the consent-announcement script? Written, memorized by managers, delivered before the substance starts.
4. What are the retention, access, and deletion rules? Including what happens the moment a duty to preserve attaches.
5. What is the shutdown procedure? Specifically: who confirms the bot has left before any post-meeting conversation begins, and where that confirmation is logged.

The last item is the Marathon rule. A written pre-call and post-call checklist that includes “confirm Fireflies / Otter / Zoom AI Companion has stopped” is the single cheapest control against the exact fact pattern that generated the Waninger complaint.

The bottom line

The Marathon Engineering complaint is not proof that AI notetakers are illegal. The In re Otter.AI ruling is not a finding that Otter broke the law. Both are early. But together they establish the working assumption every employer should now operate under: a cloud AI notetaker sitting quietly on an HR call is both a discovery witness for the plaintiff and a wiretap defendant with your name attached.

The cheapest defense is not deploying one on those calls at all. Where an employer does want a record of a sensitive conversation, moving the recording pipeline off vendor servers and onto a device the employer physically controls eliminates the specific failure modes both cases hinge on. Whether to record any given HR call remains your general counsel’s decision. Where the audio lives, if you do, is an architecture decision — and it should stop being your vendor’s.

Keep meeting audio off vendor servers

Basil AI runs 100% on-device. No cloud upload. No auto-emailed transcripts. No vendor retention.