What Is an AI Notetaker? A Plain-English Guide to How They Work (and Where Your Audio Goes)

Published October 31, 2026

Published October 31, 2026 · 11 min read

If you have joined a Zoom, Teams, or Google Meet call in the last year, you have probably seen a small avatar labeled "Otter", "Fireflies", "Read.ai" or "Granola" sitting silently in the participant list. That is an AI notetaker. The category has exploded — and so has the confusion about what these tools actually do, where the audio ends up, and what you are agreeing to when one shows up on your calendar invite.

This is a plain-English guide for anyone who has typed "what is an AI notetaker" into Google and wants a real answer — not a vendor pitch. We will cover what an AI notetaker is, the three main architectures (cloud bot, desktop capture, on-device), what a notetaker actually produces, the data-flow questions you should ask before trusting one, the legal landscape as of late 2026, and a framework for picking the right tool for your workflow.

What an AI notetaker actually is

At the most basic level, an AI notetaker is a software tool that uses artificial intelligence to take notes during meetings or other interactions. The Wikipedia entry on AI notetakers describes them as tools created to eliminate office tasks like keeping detailed meeting minutes and setting agendas, so users can focus on the conversation rather than scribbling.

In practice, almost every AI notetaker is a pipeline of three capabilities glued together:

• Automatic speech recognition (ASR) — converts audio into text, often with timestamps.
• Speaker diarization — figures out who said what and labels each turn.
• Large language model (LLM) summarization — reads the transcript and produces a summary, action items, decisions, and follow-ups.

That three-stage pipeline is why "AI notetaker" is a meaningfully different product from a plain transcription app. A transcription app gives you a wall of text. A notetaker gives you a one-page recap you can send to the team — with the trade-off that the LLM may paraphrase, miss nuance, or hallucinate.

The three architectures: cloud bot, desktop capture, on-device

If you only remember one thing from this guide, make it this: the AI notetaker category hides three very different architectures under one marketing label. The user-facing experience looks similar, but what happens to your audio is wildly different.

1. Cloud notetaker with a meeting bot

This is the Otter, Fireflies, and Read.ai model. A bot joins the meeting as a visible participant, captures the audio stream, and uploads it to the vendor's cloud. ASR and summarization run on vendor servers (or on subprocessors like OpenAI or AWS). The transcript, summary, and often the full recording are stored in your vendor account. A UK Cumbria Fire & Rescue information security newsletter flagged exactly this scenario — AI bots like Read.ai turning up in Teams meetings and needing to be identified and dismissed before sensitive information is discussed.

2. Cloud notetaker without a bot (desktop capture)

Tools like Granola skip the visible bot and instead capture meeting audio directly from your Mac or PC. From a privacy standpoint this is often cosmetic — the audio still gets uploaded to vendor servers and vendor subprocessors for ASR and summarization. Fazit's lawyer-focused analysis describes bot-free cloud transcription as "ASR/LLM subprocessors, then deleted" — the bot is gone, the vendor exposure is not.

3. On-device notetaker

Here the entire pipeline — ASR, diarization, and summarization — runs on your phone, iPad, or Mac. The audio never leaves the device. Apple's on-device Speech framework and Apple Intelligence foundation models make this architecture possible on recent hardware. Apple's privacy page explicitly positions on-device processing as its core strategy for Siri and other intelligence features. Basil AI uses this same architecture for meeting notes.

What an AI notetaker actually produces

A modern AI notetaker typically generates several distinct artifacts for a single meeting. Fellow's zero-day retention guide breaks these out as separate "data layers" — each with its own retention schedule and compliance implications:

Those last two rows matter more than they look. Voiceprints are the subject of active litigation (see below), and action items plus decisions typically persist even if you turn on "zero-day retention" for the raw audio and transcript.

Cloud vs on-device: the comparison that actually matters

Buyer's guides love to compare notetakers on summary quality and CRM integrations. For almost every regulated or sensitive conversation, the comparison that actually matters is architectural:

The Laxis 2026 analysis of on-device vs cloud transcription frames it the same way: cloud tools have unclear data retention policies and third-party access provisions buried in service agreements, while on-device processing keeps data under complete local control.

Where does your audio actually go?

This is the question most "what is an AI notetaker" explainers skip. The honest answer is that it depends on the vendor, and reading the privacy policy — not the marketing page — is the only way to know.

For cloud notetakers, the audio typically travels along this path: your microphone → the meeting platform (Zoom/Teams/Meet) → the notetaker's ingestion endpoint → the vendor's ASR provider → the vendor's LLM provider → storage in your vendor account. Screenapp's workplace compliance guide notes that cloud tools typically send audio to third-party servers like AWS, Google Cloud, or Azure, and that data often crosses borders — potentially triggering GDPR obligations even for EU customers.

For on-device notetakers, the path is dramatically shorter: your microphone → your device's Neural Engine → local storage. Nothing leaves. That architectural difference is what shrinks every downstream risk — training exposure, subprocessor exposure, breach exposure, subpoena exposure, and privilege-waiver exposure.

The 2026 legal landscape every user should know about

The AI notetaker category went from "productivity gadget" to "active litigation target" in 2026. If you are picking one for professional use, you should know the headlines.

In re Otter.AI Privacy Litigation

On August 13, 2026, Judge Eumi K. Lee of the Northern District of California partly denied Otter's motion to dismiss in the consolidated privacy class action. The federal Wiretap Act (ECPA) claim, the California Invasion of Privacy Act §631 claim, and both Illinois BIPA voiceprint claims all survived and moved toward discovery. Plaintiffs allege Otter's AI notetaker recorded them without consent and trained models on their calls. The court let claims proceed on the theory that Otter is plausibly a third-party eavesdropper.

BIPA voiceprints and $1,000–$5,000 per-violation exposure

Illinois' Biometric Information Privacy Act creates liquidated damages of $1,000 per negligent violation and $5,000 per intentional one. The Sheppard Mullin analysis of the Otter ruling notes the court rejected Otter's extraterritoriality defense because the plaintiffs were Illinois residents physically in Illinois when their data was collected. For any speaker-identification feature that fingerprints a voice, the BIPA exposure is now real.

Attorney-client privilege waiver

Sending privileged audio to a third-party vendor is a disclosure to third parties — the classic ingredient for a waiver argument. We covered this in depth in our guide to AI notetakers for lawyers and privilege waiver.

Broader privacy and compliance risks

In April 2026, Goodwin Law's data privacy practice published an alert, "AI Transcription Tools Under Scrutiny," flagging BIPA violations, wiretapping claims, and privileged-information exposure as the main legal theories being tested against cloud transcription vendors. The alert notes that once audio and transcripts are generated, they often reside on systems controlled by the service provider rather than the organization that initiated the recording — and depending on license terms, vendors may retain data indefinitely, share it with third parties, or use it for model training.

GDPR, HIPAA, and the compliance surface a notetaker creates

When a cloud notetaker ingests your meeting, it becomes a data processor under Article 28 of the GDPR — which requires a written data processing agreement, documented security measures, and controls on subprocessors. Article 5's data minimization principle is in tension with the "record everything, store it indefinitely" default of most cloud notetakers.

If any of the meeting content qualifies as protected health information, HHS HIPAA Privacy Rule guidance treats the vendor as a business associate — which means a signed BAA, access controls, breach notification, and the whole Security Rule stack. Many consumer-tier notetakers simply do not sign BAAs.

For a deeper look at the competitor landscape, see our AI meeting notetaker comparison guide.

What vendors do with your audio: training, retention, and subprocessors

Three questions decide whether a notetaker is defensible for anything sensitive:

1. Does the vendor's default allow training on your meetings? Plaintiffs in the consolidated In re Otter.AI Privacy Litigation docket allege Otter retains transcripts indefinitely and uses them to train its speech-recognition and machine-learning models without disclosure. The Otter.ai privacy policy itself is the best primary source to check for current terms.
2. What is the default retention? "Indefinite unless you configure it" is a common and dangerous default. Zero-day retention — where raw audio and transcript are deleted immediately after AI processing — is a strong signal of a vendor that respects your containment perimeter.
3. Who are the subprocessors? Most cloud notetakers depend on third-party ASR (Deepgram, AssemblyAI, OpenAI Whisper API) and LLMs (OpenAI, Anthropic, Google). The Fireflies privacy policy and Zoom's privacy statement both disclose multiple subprocessors — each of which gets a copy of your audio or transcript.

How to pick an AI notetaker: a framework based on where your audio lives

Here is a decision tree that works for almost everyone:

• If your meetings never touch anything confidential — personal to-do lists, internal team syncs about low-stakes projects — a cloud notetaker is probably fine. Check the training default and retention anyway.
• If your meetings touch any kind of regulated or confidential content — client calls, deal rooms, HR, health, legal strategy, MNPI, source conversations — prefer on-device. The architecture removes an entire category of third-party risk rather than mitigating it through a DPA.
• If you work in a regulated industry — financial services, healthcare, law — on-device is the shortest path to a clean containment story. For deeper reading, see our guide to bot-free, on-device AI meeting notes.

What to check in writing before trusting any notetaker

• Processing location (vendor cloud vs on-device).
• Default retention for audio, transcript, and summary.
• Training-data clause in the DPA — opt-in, opt-out, or disclaimed by default.
• Full list of subprocessors.
• Whether the vendor will sign a GDPR Article 28 DPA and a HIPAA BAA if needed.
• Whether speaker diarization uses a persistent voiceprint, and how it is stored.
• Export and deletion controls — can you get everything out and prove it is gone?

How Basil AI solves this

Basil AI is an AI notetaker built on the on-device architecture described above. The whole pipeline — speech recognition using Apple's Speech framework, summarization, and action-item extraction — runs locally on iPhone, iPad, and Mac. The audio never leaves your device, and there is no Basil server holding a copy of your meeting.

Concretely, that means:

• No third-party copy of the audio. There is no vendor cloud to subpoena, breach, or train on.
• No bot joining the meeting. Basil captures audio locally on your device — nothing to kick out of the participant list, nothing to disclose as an external observer.
• User-controlled retention. Transcripts export to Apple Notes / iCloud or stay in the app. You decide the schedule.
• Offline-first. Works on a plane, in a SCIF-adjacent room, or anywhere else the vendor cloud cannot reach.
• 8-hour continuous recording. Full-day workshops, depositions prep, and board sessions in one capture.

That does not make Basil AI the "compliant" choice for every workflow — compliance is always the firm's determination. What it does is remove the vendor-cloud leg of the data flow entirely, which is the leg most of the 2026 litigation and regulatory scrutiny is attacking.

The bottom line

An AI notetaker is, in plain English, software that listens to your meeting and produces notes using AI. That label hides a lot. The thing that actually matters — for privacy, for compliance, for discovery, for peace of mind — is where your audio is processed and what the vendor does with it afterward.

Cloud notetakers trade a clean containment story for convenience and shared storage. On-device notetakers keep the containment story clean by never creating the vendor copy in the first place. Both can produce excellent summaries. Only one of them removes an entire category of risk by architecture.

If you are picking a notetaker in late 2026, that is the question to lead with — not "which one has the best Slack integration."

Try an AI notetaker that keeps your audio on your device.

Basil AI runs the whole pipeline — transcription, speaker labels, summaries, action items — locally on iPhone, iPad, and Mac. No cloud. No vendor copy. No training on your meetings.