Questions to ask an AI notetaker about your clients' data

An AI notetaker is safe for client calls when its vendor can answer a short list of questions in writing, specifically. This guide is a vendor-neutral checklist for freelancers and consultants who take client calls on Google Meet and Microsoft Teams: what to ask about model training, subprocessors, isolation between clients, access to recordings, deletion, backups and certifications, and what a good answer sounds like. TrackHeros' own answers come last, so you can hold them to the same standard.

Are AI notetakers safe for client meetings?

Some are, and the category label tells you nothing. A notetaker holds recordings of other people's meetings, sends them to at least one AI provider, and keeps them somewhere you will never look. Whether that is safe depends on answers you can collect in an afternoon.

Ask in writing and keep the replies. Sooner or later a client will ask you the same questions, and forwarding the vendor's answer is better than recalling it.

What does a good answer sound like?

It names a mechanism, a document and a time limit. "Our model provider's enterprise terms exclude training on customer content; here is the clause" is an answer. "We take security very seriously" is not.

Signs of a weak answer:

• Adjectives where a mechanism should be.
• A link to the homepage instead of the privacy policy, the terms or a data processing agreement.
• "We may" anywhere near your content.
• A certification logo offered in place of a reply.
• "It depends on your plan" for a privacy promise.

Does the vendor, or its AI provider, train models on your meetings?

This is the first question and the one most often answered loosely. Ask:

1. Do you train any model on our recordings, transcripts or notes?
2. Do the AI providers you send them to? Under which terms?
3. Is that the default for every account, or a setting someone has to switch off?
4. Is "de-identified" or "aggregated" data an exception?

A good answer: "No, and neither do our providers. The AI calls run under enterprise terms that exclude training on customer content, and that applies to every account." The same company can offer a consumer AI app and an enterprise cloud service under very different terms, so ask which one the vendor uses.

Who else processes the recordings?

Ask for the subprocessor list: every company that handles the data on the vendor's behalf, and what each does. Then ask:

• Who transcribes the audio, and who writes the summaries?
• Does any of them keep the audio after processing?
• Who hosts the database and the files?
• How will you tell us when the list changes?

A good answer is a short published list with one purpose per line. "Carefully selected partners" is not a list.

How are clients kept apart inside one account?

General security questionnaires skip this one, and it matters most if you work for several clients. Ask:

• Is the separation enforced on every query, or is it a filter in the interface?
• Can one calendar account feed two clients?
• When the AI assistant answers, what does it read?
• Do cross-client views, such as search or a dashboard, label every row with its client?

A good answer: "Every read is scoped to one client before it runs, and a query without a client fails." Then test it. Open one client and ask a question only another client's meetings could answer; one notetaker account for all your clients has the full ten-minute test.

Who can open a recording?

Ask:

• Are there share links or public URLs? How long does a playback link last?
• Are recordings or summaries emailed to attendees automatically?
• Which of your staff can open customer recordings, and when?

A good answer: no public links, short-lived playback links tied to one meeting, nothing sent anywhere unless you send it, and staff access limited to the people who run the systems.

What does deletion actually reach?

Ask:

• Can I delete one recording, one client and the whole account?
• Which stores does deletion reach: the database, file storage, the meeting bot's own copy, search indexes?
• Is it immediate in the live systems?

A good answer is a list of stores and the words "at once". "We will delete your data on request", with no list, is a promise you cannot check.

How long do backups keep deleted data?

Ask how long backups are retained, and whether deleted data is ever restored into a live account. A good answer is a number of days and "never restored". A vendor that keeps backups indefinitely is telling you deletion stops at the live systems.

Where is the data held, and what proof is there?

Ask which countries the data is stored and processed in, and whether the vendor holds a security certification you can read the report for. A good answer gives regions on request and either a report or a plain "we have none". The plain answer is more useful than a badge you cannot check.

What should you do with the answers?

Keep them, date them and reuse them. A short file per vendor, with each question, the reply and a link to the document it came from, does three jobs:

• It is what you send when a client's security team sends you their questionnaire.
• It shows what changed when the vendor updates its terms. Read the new version against your file rather than from memory.
• It tells you which client calls the tool is fit for. A vendor that is fine for weekly status calls may not be fine for calls covered by a strict NDA; AI notetakers and client NDAs covers that check.

Ask again once a year, and whenever the vendor announces a new AI feature. New features are where new providers and new terms tend to arrive.

How does TrackHeros answer these questions?

TrackHeros is an AI meeting notetaker that connects every Google and Microsoft calendar you hold, recognises close to 100 spoken languages on its own, writes the notes in nine of them, and keeps every client's notes in a sealed workspace of its own. Its answers come from its privacy policy, and the four promises behind them are at /#privacy.

Question	TrackHeros' answer
Training	No, for its own models or anyone else's. Every model call except building the search index goes through OpenRouter, and every request requires a provider that keeps no copy of the content and refuses any use of it for training. Google Cloud's Vertex AI runs the Gemini and Claude models, and Amazon Bedrock runs Claude when Vertex AI cannot. Meetings in a language other than English, or in more than one, are transcribed by DeepInfra, Together AI or Groq, under terms that exclude training and keep no copy of the audio once it is transcribed. The search index is computed on Vertex AI directly
Subprocessors	Listed in the privacy policy with what each does: OpenRouter, Vertex AI, Amazon Bedrock, DeepInfra, Together AI and Groq for the AI work, plus hosting, billing, email delivery and website analytics
Isolation	Every query on client data is scoped to one client before it runs; an unscoped query is an error. Postgres row-level security checks the same rule underneath. Each Google or Microsoft account belongs to exactly one client, the assistant reads one client, and cross-client views are read-only with every row labelled
Recording access	Playback inside the app through short-lived, meeting-bound links. No public URLs, no share links, no automatic email of summaries. Access to production systems is limited to the people who operate them
Deletion	One recording (its transcript and notes stay), a whole client, or the whole account; the last two ask you to type the client's name or your email address. Deleting a client or the account removes the database rows, the stored recordings and transcripts, and the copy held by the meeting-bot service, at once
Backups	Expire within 30 days of a deletion, and are never restored into a live account
Location	Servers run by the hosting provider, and the AI providers above; any of them may process data outside your country. The current regions for a client of yours are given on request
Certifications	None held, none claimed. Ask, and TrackHeros will say exactly what it does and does not have

The detail is in client workspaces that never mix, does TrackHeros train AI on my meetings? and what happens to recordings when you delete a client.