Bot-Free vs On-Device AI Notetakers on Mac: Why 'No Bot' Doesn't Mean Local (2026)

If your reason for switching to a bot-free AI notetaker was privacy, you may have solved the wrong problem. "Bot-free" — the marketing category pioneered by Granola and now claimed by Jamie, Fathom, Otter Desktop, Laxis, and half a dozen others — describes what your meeting participants see: no third-party bot in the participant list, no "Notetaker has joined" announcement. It says nothing about where your audio actually goes. In most cases, it still goes to Deepgram, AssemblyAI, OpenAI, and Anthropic. This guide explains the difference between bot-free and on-device on Mac in 2026, why it matters more than any feature comparison, and which apps genuinely process meeting audio locally on Apple Silicon.

The category confusion: three overlapping definitions

The AI notetaker market has three distinct architectures that get muddled in a single word — "private." Getting them straight is the whole game.

Bot-based cloud tools (Otter, Fireflies, Fathom's default mode, Zoom AI Companion) send a visible participant into your Zoom, Google Meet, or Teams call. Audio streams to vendor servers in real time and is typically retained. This is the design that produced the consolidated federal class action In re Otter.AI Privacy Litigation, currently pending in the Northern District of California and awaiting a ruling on Otter's motion to dismiss.

Bot-free cloud tools (Granola, Jamie, Otter Desktop, Laxis, Supernormal) capture system audio directly on your Mac — no bot — then stream that audio to third-party transcription and LLM providers. There is no visible bot, but the data-flow problem is identical.

On-device tools (Basil AI, and to varying degrees Talat, Meetily, Hyprnote, Mumble) run transcription and, in the best cases, summarization entirely on your Mac. Nothing leaves the device unless you export it. As Mumble's May 2026 roundup of local notetakers put it bluntly: "bot-free does not mean local."

Why this distinction became urgent in 2026

The Brewer v. Otter.ai lawsuit reframed the conversation. Filed August 15, 2025 by California resident Justin Brewer and consolidated with three later suits on October 22, 2025, the consolidated case In re Otter.AI Privacy Litigation (5:25-cv-06911-EKL, N.D. Cal.) alleges violations of the federal Wiretap Act and the California Invasion of Privacy Act (CIPA). The motion-to-dismiss hearing was held on May 20, 2026 and, as of late June 2026, Judge Eumi K. Lee had not yet ruled.

The legal theory does not depend on the bot being visible. A November 2025 legal analysis in Best Law Firms notes that Otter's terms of service push consent obligations onto users, and that this "strategy, familiar in the tech sector, may not survive judicial scrutiny when the vendor builds, controls and profits from the recording infrastructure itself." What matters legally is whether audio was captured and transmitted to a third-party server without all-party consent — bot or no bot.

That's the compliance problem hiding underneath the "bot-free" label. As Krisp's 2026 Granola review observes, in two-party or all-party consent jurisdictions like California and parts of the EU, "the lack of a bot makes the process less obtrusive but does not change the legal baseline."

Where each "bot-free" tool actually sends your audio

Here is the specific data flow for the most popular bot-free notetakers on Mac, sourced from each vendor's own documentation.

Granola

Granola captures device audio locally without joining as a visible participant. Granola's own security page confirms the app "uses best-in-class transcription providers (like Deepgram and Assembly) and AI providers (like OpenAI and Anthropic) to summarize your meeting." Audio is deleted after transcription, but during transcription it is streamed to those third-party providers. Granola contractually prohibits those third parties from training on your data, and users can opt out of Granola's own anonymized training in settings — though as one detailed 2026 review noted, org-wide training opt-out is gated to the $35/user/month Enterprise plan. Transcripts and notes persist in a US-hosted AWS Virtual Private Cloud.

Jamie

Jamie is bot-free and GDPR-oriented, with audio capture on-device followed by processing in EU/German infrastructure. Audio is deleted after transcription. This is better than most US cloud vendors for European teams, but the audio still leaves the Mac to reach Jamie's servers.

Otter Desktop

Otter's desktop app removes the visible meeting bot but remains part of Otter's cloud transcription platform. All the data-retention and training concerns raised in Brewer v. Otter.ai apply to the desktop version as much as OtterPilot.

Fathom, Supernormal, Laxis

All added bot-free capture modes in 2025 or 2026, but the underlying pipeline still routes audio to their cloud infrastructure for transcription and summary generation.

The comparison table: bot-free vs on-device across 6 dimensions

This is the matrix your security team should actually be looking at.

Sources: vendor security pages and independent 2026 reviews cited throughout this article.

The technical primitives: what "on-device" actually means on Apple Silicon

The phrase "on-device" is often used loosely. On Apple platforms it has a specific meaning tied to concrete APIs.

Apple's requiresOnDeviceRecognition property on SFSpeechRecognitionRequest is a Boolean that, per Apple's developer documentation, "determines whether a request must keep its audio data on the device." Setting this to true forces the entire recognition pipeline through the local model on the Neural Engine and prevents any audio from being sent to Apple's servers — or anyone else's.

supportsOnDeviceRecognition is the companion API that reports whether the recognizer can operate without network access for a given locale — critical for developers who want to gate their app's availability on true offline capability.

With iOS 26 and macOS Tahoe, Apple shipped a new SpeechAnalyzer class alongside the old SFSpeechRecognizer. A November 2025 technical guide to SpeechAnalyzer describes it as "designed for performance, flexibility, and full offline operation," with modular components including DictationTranscriber, SpeechTranscriber, and SpeechDetector. The new framework is what lets developers manage on-device speech models for specific locales, download them explicitly, and use them without a network round-trip.

These are the technical primitives that separate genuinely on-device Mac apps from "desktop app with a cloud backend" pretenders. When a vendor claims "on-device," the honest question to ask is: which API? If they can't name it, they're probably running Whisper on a server they own.

The regulatory decision tree: which scenarios require on-device?

Bot-free is often good enough. On-device is required in specific, high-stakes contexts. Here's the decision rule per scenario.

HIPAA (protected health information)

Any call between a clinician and a patient, or between covered entities discussing patient care, is subject to HHS HIPAA Privacy Rule requirements. If audio leaves the device, the vendor must sign a Business Associate Agreement (BAA). Granola is not HIPAA compliant as of early 2026. Fully on-device processing eliminates the BAA question entirely because no PHI ever reaches a covered vendor.

Attorney-client privilege

Any time a lawyer discusses matter strategy with a client, transmitting that audio to a third-party server risks a waiver argument. On-device transcription keeps the privilege intact by keeping the communication within the client-lawyer relationship.

All-party consent states (California, Florida, Illinois, Massachusetts, Pennsylvania, Washington, Maryland, Montana, Connecticut, New Hampshire, Oregon, Vermont)

Bot-free doesn't cure the consent problem. If you're recording or transcribing a confidential communication in these states, you need affirmative consent from every participant — whether the capture happens via bot, via device audio, or via an on-device model. On-device does, however, keep the recording out of a third-party's evidence-retention obligations, which matters if the call is later subject to subpoena.

GDPR (EU participants)

Under Article 5 of the GDPR, personal data must be processed lawfully, transparently, and only for specified purposes (data minimization). On-device processing is the cleanest path to compliance because it eliminates the cross-border data-transfer question and reduces the vendor's role from processor to (in effect) nothing at all. See our analysis of AI meeting notes and legal discovery for how on-device architecture affects work-product claims.

NDA-heavy client work (consulting, M&A, competitive strategy)

Most NDAs prohibit disclosing confidential information to unnamed third parties. A bot-free-but-cloud tool that pipes your client's data through Deepgram, OpenAI, and Anthropic is disclosing that data to three unnamed third parties. On-device isn't just cleaner — it may be the only architecture consistent with the NDA you signed.

The Basil AI test methodology: our "Nothing Leaves" audit

To claim "on-device," a tool should survive a network audit. Here's the methodology we use for every Basil AI release. We call it the Nothing Leaves audit, and it's the same test we recommend before trusting any vendor's on-device claim.

1. Baseline: Install the app on a fresh Mac running macOS Tahoe. Note the app version and macOS build.
2. Network monitoring: Enable Little Snitch or the built-in macOS Network Extension framework to log all outbound connections from the app's bundle identifier.
3. Airplane test: Disconnect Wi-Fi. Start a recording. Verify that transcription still occurs in real time.
4. Summary test: Still offline, generate an AI summary. Verify it completes without a network prompt.
5. Reconnect and diff: Reconnect Wi-Fi. Confirm no queued network traffic flushes to a vendor endpoint. Any post-hoc upload is a fail.
6. Report: Document the exact date, macOS version, and app version tested. Publish the results.

We ran this test against Basil AI on July 8, 2026 on macOS Tahoe 26.1 with Basil AI 3.2. Zero outbound audio traffic. Transcription and summarization both completed with Wi-Fi disabled. See our deep dive on Apple Neural Engine transcription for the technical breakdown of how this works.

Where truly local Mac apps sit today

Beyond Basil AI, a handful of Mac apps are genuinely working toward full local processing. Mumble's roundup of local AI meeting notetakers for Mac in 2026 tested eight tools across three groups and found that Talat "publicly positions itself as using 100% on-device AI, with recordings, transcripts, and notes stored in a local database that does not leave the machine." Meetily and Hyprnote are open-source and locally configurable but require developer-level setup for most users. Basil AI is the only tool that combines fully on-device processing with a consumer-grade UX and native iPhone, iPad, and Mac apps.

The rest of the "privacy-first" category — including Otter's privacy policy and Fireflies' privacy policy — is still fundamentally cloud. Their marketing has become more careful in the wake of Brewer, but the architecture hasn't changed.

How Basil AI solves this: 100% on-device across iPhone, iPad, and Mac

Basil AI is the fully on-device answer to the bot-free trade-off. Here's what "on-device" means for us, in specifics:

• Transcription: Uses Apple's SFSpeechRecognizer with requiresOnDeviceRecognition = true, plus the new iOS 26 SpeechAnalyzer and SpeechTranscriber pipeline where available.
• Summarization: Runs on Apple's on-device foundation models, executed on the Apple Neural Engine — no OpenAI, no Anthropic, no Deepgram, no AssemblyAI in the loop.
• Storage: Recordings, transcripts, and summaries persist in a local database and (optionally) sync via your own iCloud account — never through a Basil-controlled server.
• Bot behavior: None. There is no bot to join a meeting, because Basil doesn't join meetings — it captures device audio and your microphone locally.
• Offline: Full functionality in airplane mode, including on 8-hour continuous recordings. See our offline airplane-mode test of the best iPhone notetakers.

The consequence is architectural: there is no vendor cloud that could be subpoenaed, breached, or repurposed for training. You own the data because it never left your hardware. For a head-to-head on the Granola comparison specifically, see our breakdown of notetaker vs assistant vs transcription categories.

What to ask any vendor claiming "on-device"

If you're evaluating an AI notetaker and the vendor uses the phrase "on-device," "local," or "private by design," ask these six questions before you sign anything:

1. Which Apple API do you use for transcription — SFSpeechRecognizer with requiresOnDeviceRecognition, the iOS 26 SpeechAnalyzer, or a server-side Whisper deployment?
2. Does summarization happen on-device (Apple foundation models, local Llama, MLX) or via a third-party LLM API?
3. Can you provide the list of sub-processors that receive audio, transcript, or summary content? (Granola's list includes Deepgram, AssemblyAI, OpenAI, and Anthropic.)
4. What is deleted after transcription, and what persists — and where geographically?
5. Does the app function fully in airplane mode, end-to-end?
6. Will you sign a BAA (for HIPAA) or a DPA with Standard Contractual Clauses (for GDPR)?

Any vendor unable to answer these clearly is not on-device. They may still be a good tool. They are not the tool for privileged, protected, or regulated conversations.

The bottom line

Bot-free was a real improvement over the visible-bot model that produced the Otter lawsuit. It solves a genuine social problem on client calls and eliminates one obvious form of the consent failure. But if your reason for going bot-free was privacy, compliance, or professional obligation, the marketing category doesn't reach far enough. Your audio still leaves the Mac. It still passes through Deepgram or AssemblyAI. It's still summarized by OpenAI or Anthropic. It still lives — as text, at minimum — in a US-hosted AWS Virtual Private Cloud.

On-device is different. It's the architecture where the audit surface is your own hardware and no one else's. On Apple's privacy platform, that architecture is now genuinely production-grade. Basil AI is what it looks like when a consumer app is built entirely on top of it.

Try the only fully on-device AI notetaker on Mac and iPhone

Basil AI: bot-free and on-device. 8-hour recording, real-time transcription, AI summaries — all running on your device. Nothing ever leaves.