Does Your AI Note-taker Train on Your Calls?

There are two questions about model training, they have different answers, and a vendor can answer the easy one truthfully while leaving the hard one unaddressed. Reading only the first is the most common mistake made when evaluating these tools.

The two questions

The first is whether the third-party model providers a vendor uses will train on your content. Almost every vendor can answer no, because enterprise agreements with the major model providers routinely exclude it. This is genuinely worth having, and it is also close to table stakes.

The second is whether the vendor itself trains on your content, usually described as being de-identified or anonymised first. This is the question that actually varies, and the one where the answers in this directory diverge sharply.

Both statements can appear on the same page, and reading them quickly leaves you with an impression that only one of them supports.

What the range actually looks like

Among the tools covered here, the positions fall into three groups.

Some state flatly that customer content is never logged or used for training or testing, on every tier, with nothing to opt out of. That is the strongest position available and it requires no action from you.

Some train on de-identified customer data by default and offer a setting to stop it, with exclusion sometimes becoming the default only on an enterprise plan. That is a real difference from the first group even though both can honestly say a training opt-out exists — the difference is what happens if nobody goes looking for the setting.

And at least one uses de-identified customer recordings and transcripts to train its own models, while separately confirming that its AI service providers do not. Both sentences are accurate. Only one of them is the answer to the question most people think they are asking.

How to check in two minutes

Search the vendor's privacy or security page for the word "train" and read every sentence it appears in, not the first one. Then ask three follow-ups: is exclusion the default or a setting; does it apply on the plan you will actually buy or only on enterprise; and does the commitment cover the vendor as well as its model providers.

While you are there, check what happens to the raw audio. Several vendors delete it once the transcript exists, which quietly removes the most sensitive artefact from the question entirely. That is worth more than a carefully worded training clause, because deleted audio cannot be trained on, leaked, or subpoenaed later.