AI Notetaker That Does Not Train on Customer Data: How to Verify

An AI notetaker that does not train on customer data is one where that promise is written into a contract or enforced by how the software works, not just stated on a landing page. The strongest guarantees come from tools that run entirely on your device, or from vendors that sign a data processing addendum (DPA) binding them and their sub-processors. Everything else is a policy you are trusting.

This guide separates those cases, shows what vendors actually publish, and gives you five questions to put to any supplier. It also covers an awkward gap: privacy-first tools are often built for English, and your team may not speak English all day.

Policy promises versus architecture: what "does not train" really means

A policy promise says "we will not train on your data". An architectural guarantee means the data never reaches anyone who could. The difference matters because policies change, staff make mistakes, and sub-processors have their own terms.

There are four common setups:

Many roundups blur these together. They also confuse encryption in transit (TLS) with a training opt-out. TLS protects data on the way to a server. It says nothing about what the server then does with it. A vendor can use TLS on every connection and still train on what arrives.

So when a page says "secure" or "encrypted", ask a separate question: is training prohibited, and where is that written down? For a wider set of checks, see our guide on whether AI note takers are safe.

Local and offline notetakers: audio never leaves the device

Local tools remove the question rather than answer it. If transcription and summarisation run on your hardware, no vendor receives your meeting.

Several tools take this approach, according to their own pages:

• Private Notetaker says it runs Whisper-class transcription and local language model summarisation entirely on your computer, with no bot joining the call and no central server collecting or storing meeting data.
• Notely says it runs all AI transcription and summarisation locally, using GPU acceleration across NVIDIA, AMD and Apple Silicon, and works completely offline so meeting data never leaves the device.
• Private Whisper offers a fully on-device pipeline using local Parakeet speech-to-text and a local Gemma model run through llama.cpp.
• OpenWhispr uses local speech-to-text by default with open-source models, so no audio leaves the computer. Its terms also commit that customer content (audio, transcriptions, notes or AI prompts) will never be used to train AI or machine learning models.

What you give up

Local does not mean free of trade-offs. You carry the hardware cost: Notely's own description mentions GPU acceleration, so check what your team's laptops can actually run. Summaries from small local models may be less polished than those from large cloud models. And someone has to keep the software updated.

Local tools also tend to be single-user. If you want shared meeting history, search across a team, or a central admin, you are likely to need a cloud component, which brings back the questions in the next sections.

OpenWhispr shows the combined model

OpenWhispr is worth noting because it pairs the architecture with a contract. Local processing is the default, and its Data Processing Addendum states that it will not use customer personal data to train, retrain or improve any AI or machine learning model, its own or a third party's. It also requires sub-processors to disable or opt out of training. That is the pattern to look for: a technical default backed by a document you can point to.

Cloud notetakers with contractual zero-training terms

A cloud notetaker can be a reasonable choice if the zero-training terms are in a signed agreement and extend to every sub-processor. The agreement is what you can enforce; a web page is not.

Two agreements matter most:

• DPA (data processing addendum): governs how personal data is processed. Look for explicit language that the vendor and its sub-processors will not use your data to train or improve models.
• BAA (business associate agreement): relevant for health information. CareNotes says it protects patient health information by executing BAAs with providers such as OpenAI with Zero Data Retention, where data is not logged, stored or saved to disk. Speak AI says it operates under BAAs with infrastructure partners that contractually prohibit model training on customer data and limit use to operational processing.

Hybrid designs: local audio, cloud text

Some tools split the work. Cleft says voice audio stays on-device using Whisper on iOS and macOS. Encrypted notes are hosted under AES-256-GCM in Ireland (EU), and only text, never audio, goes to OpenAI or Groq, under DPAs that prohibit model training.

Private Whisper's cloud route works similarly. It says its cloud speech-to-text defaults to Deepgram, where data retention and training are turned off by default for API customers.

Ephemeral processing

DiscMeet describes ephemeral AI inference: no third-party AI provider stores or retains audio or transcripts, and audio is kept only in an encrypted Cloudflare R2 bucket during processing. The idea is that data exists only as long as the job needs it.

Whatever the design, read the sub-processor list. A vendor's own promise is only as good as the weakest company it passes your data to.

Bring your own key, enclaves and ephemeral processing

These are the middle options between fully local and standard cloud. Each shifts trust somewhere narrower than "the vendor's goodwill".

Bring your own key (BYOK)

In a BYOK setup, you supply your own API key for the speech or language provider. Your account with that provider, and its terms, then apply to your data. For an organisation that already has an enterprise agreement upstream, this can mean inheriting those terms, including no-training and no-retention clauses.

BYOK is not a guarantee by itself. Protection depends on the terms attached to your key and account type, which vary by provider and plan. Get those terms in writing and check that the notetaker does not also copy transcripts into its own storage or logs.

Confidential enclaves

A trusted execution environment (TEE) is a hardware-isolated area where data is processed so that the host cannot read it. Private Whisper says it supports enclave processing through Tinfoil, encrypting text on-device so neither the vendor nor intermediaries can read it. It also offers zero-retention cloud cleanup through Groq.

Enclaves are attractive because they cut trust in the operator. They are also newer and harder for a non-specialist to audit. Ask for documentation on how the enclave is verified.

Where this fits

If you cannot run models locally but cannot accept a standard cloud tool either, enclaves and BYOK are the middle ground. They add technical or contractual control without asking each person to run models on their own laptop.

Multilingual reality check for Southeast Asia and Nigeria

Accuracy in your own languages is the part buyers test too late. Smaller local models in particular tend to be weaker on lower-resource languages such as Burmese, Hausa or Yoruba, and on mixed-language speech.

Treat any claim of broad language support as unproven until you test it. A practical way to do that:

1. Pick three real recordings: one clean, one noisy, and one with code-switching (for example Burmese and English in the same sentence).
2. Run each through the tool on the hardware your team actually uses.
3. Have a native speaker mark errors in names, numbers and decisions, since those are the errors that cause damage.
4. Note the processing time on your slowest machine.

This matters more for local tools because you cannot rely on a larger cloud model to rescue weak audio. If a local tool fails your test, a cloud tool with a signed DPA may be the safer trade: strong accuracy with contractual limits, rather than private but unusable notes.

For language-specific detail, see our posts on Burmese speech-to-text accuracy, Burmese-English code-switching and transcription software for Southeast Asian languages.

Five questions to ask before you buy an AI notetaker that does not train on customer data

Ask these in writing. A vendor with a real answer will give one quickly, and vague replies are information too.

1. Is the no-training commitment in a contract or only on the website? Ask for the DPA, and a BAA if you handle health data. Check that the wording covers training, retraining and improvement of models.
2. Does it cover sub-processors? Ask for the list, and for confirmation that each is bound by the same no-training and, where relevant, no-retention terms.
3. Where does audio go, and for how long? Distinguish between audio, transcripts and derived data such as summaries or search embeddings. Ask what is retained, where, and how deletion works.
4. Who holds the keys? Ask how stored data is encrypted and who can decrypt it. Remember that encryption at rest protects against stolen storage, not against the vendor's own use of the data.
5. Does it work on your languages and hardware? Run the test above. A private tool nobody can use accurately will be bypassed.

If you need a policy for staff rather than a vendor checklist, our confidential meeting recording policy template sets out tiers for what can and cannot be recorded.

A simple decision rule

• Board, legal or clinical content: prefer fully local, or a signed DPA or BAA with documented sub-processors.
• Routine team meetings: a cloud tool with a clear no-training contract is usually enough.
• Mixed-language meetings: test accuracy first, then decide how much privacy architecture you can afford.

Where Loka Note fits

Loka Note is a cloud tool, so it sits on the contractual side of this spectrum, not the local one. It records in the browser with no bot joining the call, or takes uploaded audio and video, and produces a transcript, summary, decisions and action items. It is built Burmese-first, with English, Thai, Vietnamese, Chinese, Yoruba and Hausa, including mixed Burmese-English meetings.

On data use, our security page states that customer audio and transcripts are processed for transcription and summarisation only and are never used to train AI models. Transcripts, summaries, action items, notes and search embeddings are encrypted with AES-256-GCM under a unique per-user key, wrapped by a key held in Google Cloud KMS. This is defence-in-depth for stored data, not end-to-end encryption. Users can delete any meeting or their whole account at any time, and account deletion destroys the per-user key. Our sub-processors are Vercel, Supabase, Google Cloud, Polar, a transactional email provider, PostHog and Sentry. Recordings are kept for 6 months and can be deleted at any time.

The enterprise tier adds a signed DPA, configurable retention, SSO and centralised administration. If you need audio that never leaves a laptop, a local tool from the list above is the better fit. If you need Burmese or other regional languages handled well, with clear data-use terms, that is what we built.

Try Loka Note with pay-as-you-go minutes