Does Your Meeting Data Train the Model? How to Find Out
The reason this question is hard to answer is that it is actually two questions, and most sources answer only one of them. Does the notetaker vendor train its own models on your meetings, and do the third-party AI providers that vendor sends your audio to train on it? A tool can honestly answer no to the first while your transcripts still pass through a model provider under separate terms. This guide is a method rather than a scoreboard: where the answer lives, what words to search for, and what to ask when the documentation will not say. Do it once and you can answer it for any tool, including the one you switch to next year.
It is two questions, not one
Write them down separately, because a vendor's marketing page usually answers the first and stays quiet on the second.
Question one: does the vendor train on your data? This is about the company whose product you installed using your meetings to improve its own models.
Question two: what do their AI providers do with it? Most notetakers do not build their own speech and language models from scratch. They send your audio or transcript to a third-party provider. That provider has its own terms, and the question of whether it trains on data it receives is separate from anything your notetaker promises about itself.
A vendor addressing both layers explicitly is a good sign in itself. Granola, for example, states in its own materials that it contractually prohibits third-party AI providers from training on customer data, and separately describes organization-wide opt-out from its own anonymized training on enterprise plans (checked 2026-08-04). Whatever you think of that specific arrangement, it is the right shape of answer: it addresses the vendor and the subprocessor as two different commitments.
If a tool's documentation only ever says "we do not train on your data," you have half an answer.
Why you cannot answer this from a comparison post
A great deal of what ranks for this question is written by notetaker vendors about their competitors. Those pages are sales material, and a company describing a rival's privacy defaults is not a source you should rely on for a decision about your client conversations.
The other problem is decay. Defaults in this category change quietly and often, and a roundup accurate in one quarter can be wrong in the next without anyone updating it. Anything you read about a specific tool's training policy, including anything on this page, needs re-checking against that vendor's current documentation before you act on it.
That is precisely why the useful skill is knowing where to look rather than memorizing an answer.
Where the answer actually lives
Four places, in descending order of how likely they are to be honest and specific:
- The Data Processing Agreement or DPA. This is the contractual document, and it is where subprocessors are usually named. If you want to know which third-party model providers see your data, this is the document that lists them.
- The security or trust page. Usually more specific than the privacy policy and written for buyers who ask hard questions. Look for a subprocessor list and any certification claims.
- The privacy policy. Comprehensive but written to cover the company, not to inform you. Use search rather than reading it end to end.
- The settings screen in the product. Sometimes the real answer, because a toggle that exists tells you the default it departs from.
The order matters. A marketing page saying "your data is private" is not evidence of anything. A DPA naming subprocessors is.
The words to search for in a policy
Open the policy and use your browser's find function. These terms surface the relevant clauses fast:
- train, training, model improvement: the direct question
- improve our services: the common euphemism, and the phrase that most often covers training without saying so
- aggregate, anonymized, de-identified: these qualify the promise, and they usually mean training happens on data the vendor considers stripped of identifiers
- subprocessor, third party, service provider: the second question
- opt out, opt in: tells you whether a control exists
- retention, delete: the companion question covered in where your meeting recordings live
Pay particular attention to anything sitting next to "anonymized." A commitment not to train on your data that quietly permits training on anonymized derivatives of your data is a meaningfully different commitment, and it is extremely common. Whether that is acceptable is your call, but you should make it knowingly.
Opt-in, opt-out, and the default that decides it for you
Three arrangements exist, and the difference between them is mostly about what happens when nobody does anything.
Opt-in: nothing is used for training unless you actively agree. The strongest position for you.
Opt-out with a control: training happens by default and stops when you find the setting. Workable, but only for the people who know to look, which in practice is very few.
No control: training is a condition of using the product.
The practical point is that opt-out and opt-in can be described in almost identical language on a marketing page, and they produce opposite outcomes for the majority of users who never open settings. When you find the answer, note which of the three it is, not just whether a control exists.
Consumer plan and business plan are often different products
This catches people constantly. The same tool frequently applies different data handling to a free or personal account than to a business or enterprise one, and the stronger commitments are usually gated to the paid tiers.
So "does this tool train on my data" has no single answer for a given product. It has an answer per plan. If you evaluated a tool on a company account and then installed it on your personal one, or trialed it free and read the enterprise security page, you may be holding a commitment that does not apply to you.
Organization-wide enforcement is often gated the same way. An individual toggle protects one person; an admin-enforced setting protects everyone, and that capability tends to sit on the higher tiers.
What to ask a vendor directly, in one email
If the documentation will not resolve it, ask. Vendors answer this question, and how they answer is itself informative. Keep it to four questions so it is easy to reply to:
- Do you use customer meeting audio, transcripts, or summaries to train or improve your own models? If yes, is that opt-in or opt-out, and on which plans?
- Which third-party AI providers process our data, and are they contractually prohibited from training on it?
- Does that answer differ between our plan and your enterprise tier?
- Can training be disabled organization-wide by an administrator, rather than per user?
A vendor who answers all four plainly is telling you something real. A vendor who responds with a link to a marketing page, or who answers question one and ignores question two, is also telling you something.
What to do if you cannot get a straight answer
Treat an unclear answer as a no-confidence signal and act accordingly rather than waiting for certainty.
Stop pointing the tool at your most sensitive meetings, which is worth doing regardless of the answer. Restrict it to internal, low-stakes conversations while you decide. Check whether a capture mode exists that keeps less in the first place, since a tool that deletes audio after transcription has less to train on by construction. And if your organization has a procurement or security function, hand them the four questions above; this is exactly what they are for.
The broader question of what every tool your company uses does with company data is wider than any one notetaker and is not covered on this page. Keep this one scoped to the meeting tool, because that is the one recording your conversations.
Data policies in this category carry a short shelf life. Anything cited here should be re-checked against current vendor documentation before you rely on it. If you are still choosing a tool, the hub guide on how to choose an AI meeting note taker covers the wider criteria, and bot in the call versus bot-free capture explains why the capture method does not change any of this.
FAQ
Does my AI notetaker train on my meetings?
It depends on the tool and on your plan, and the answer changes over time. Check the vendor's DPA and security page rather than a comparison post, and confirm both whether the vendor trains on your data and whether its third-party AI providers do.
What does "we do not train on your data" usually leave out?
Two things: what the third-party model providers do with data the vendor sends them, and whether training on anonymized or aggregated derivatives is still permitted. Both are common and both are frequently unmentioned.
Where do I find the real answer?
The Data Processing Agreement is the most reliable, because it usually names subprocessors. The security or trust page is next, then the privacy policy, then the product's own settings screen.
Is opt-out as good as opt-in?
Not in practice. Opt-out means training happens by default for everyone who never opens settings, which is most people. Note which arrangement you are under, not just whether a control exists.
Does my free account get the same data protections as a business plan?
Often not. Stronger data commitments and admin-enforced controls are commonly limited to paid or enterprise tiers, so verify the answer for the plan you are actually on.