4 of 10 AI Support Vendors Won't Say If Your Data Trains Their AI

Your support inbox is the most sensitive text your company holds. Order histories, addresses, complaints, refund arguments, occasionally a photo of somebody's damaged medication. It is other people's data, held on their behalf, and you agreed to look after it.

So: does any of it end up training your vendor's AI? On four of the ten platforms we checked, you cannot find out from anything they publish.

Two answer it thoroughly. Two say a flat no. One says its suppliers do not train on your data while never quite saying whether it does. One has written the answer down and put it behind a login. The remaining four do not address it in their AI terms, their DPA or their privacy policy.

There are two questions here and vendors answer the easier one

The first is whether the vendor trains its own models on your conversations. The second is whether the model provider underneath — OpenAI, Anthropic, Azure, Google — trains on them, because most support platforms do not run their own models and the AI layer is an API call to somebody else.

A vendor can truthfully say it does not train on your data while the layer it does not own remains an open question. Only three of the ten name a provider at all.

For fairness: the major API providers have defaulted to not training on business inputs since 2023. OpenAI's API terms say data sent through it is not used for training unless you opt in. So the exposure here is usually smaller than the silence suggests. The complaint is not that these vendors are secretly feeding your tickets into a model. It is that half of them make you take it on trust.

How this was read

For each platform, in this order: AI-specific terms, then the trust or security page, then the data processing agreement, then the privacy policy. We stopped at the first document that addressed AI training. Where none did, the row says not stated — an absence we went looking for, not one we gave up on.

No vendor was contacted. This is a study of what a buyer can establish alone, at eleven at night, before a call is booked. Several of these companies would no doubt answer immediately if asked, and that is the point: it should not require asking.

Freshworks sets the standard, and it costs them nothing

Freshworks is the only one of the ten that offers a choice. It does not train third-party hosted models on customer data. It does train its own, on data from consenting customers, and splits that into two kinds: custom models trained only on your organisation's data and used only in your account, and collaborative models trained on de-identified data pooled across customers.

Either can be opted out of by contacting support, and the page states the trade-off rather than burying it: opting out may reduce performance in some features or turn them off.

That is what a complete answer looks like. It admits to training, explains what kind, offers an exit, and names the cost of taking it. It is also, notably, not a competitive disadvantage — nobody is switching away from Freshworks for being clear about this.

Zendesk is nearly as good, and the sanitisation detail is worth reading

Zendesk confirms it trains its own models on customer data and then does something unusual: it describes how the data is stripped first. Fields that are identifiers by definition, username and email, are excluded from the training set. A natural language pass removes further identifying information from free text. What survives is tokenised into vectors that are not human readable.

On the provider layer it is one sentence and it is the right sentence.

“No third-party will use your inputs to train their models or otherwise improve their services.”— Zendesk AI Data Use Information, read 6 August 2026

What is missing is a choice. Nothing describes an opt-out, so on this account training on your data is a condition of using the product.

HubSpot answers the question it was not asked

HubSpot is precise about its suppliers: it does not permit the AI service providers it engages to use your data for model training. It names OpenAI, AWS and Google, and commits to minimising retention, including zero-day retention where possible. That is better than most of this table manages.

It is also not an answer to whether HubSpot trains on your data. We could not find that addressed anywhere. The sentence is about providers, and it is easy to read as covering both if you are skimming — which is the problem with it.

Ada has written the answer down and made you ask for it

Ada's trust centre contains a document called AI Training Data. It requires authorisation to open.

This is a different failure from silence and in some ways a worse one. The work has been done, the position exists, somebody wrote it out — and reading it costs you a form and a wait. A buyer comparing five platforms on a Tuesday evening does not do that for one of them.

The flat noes, and the one that names its providers

eesel AI gives the cleanest answer of the ten: your data is never used for model training, and never included in any training data, period. It also names OpenAI, Anthropic and Google as the providers underneath, which almost nobody else does, and describes per-workspace isolation.

Chatbase says it does not train on your data and explains why it does not need to — it uses retrieval rather than fine-tuning, so your content is fetched at answer time instead of absorbed. It names no provider, so the layer underneath is unaddressed.

The four that say nothing

Fin, Gorgias, Tidio and Front do not address AI training in the documents where a buyer would look.

Fin's and Gorgias's DPAs are thorough about other things. Both prohibit selling personal data and restrict processing to delivering the service, which a lawyer might argue covers training by implication. Neither says so. Tidio's privacy policy contains exactly one sentence about model training and it is about Google Workspace APIs rather than Lyro. Front acknowledges that some of its service providers supply generative AI, then neither names them nor says what they may do with your data.

None of this is evidence of wrongdoing and this post does not suggest otherwise. It is evidence that the question was not anticipated, on products sold specifically for handling other people's personal information.

What to ask, and what a good answer sounds like

Four questions, one email, before the trial ends:

• Do you train your own models on our conversation data? If yes, is it isolated to our account or pooled across customers?
• Which model provider handles inference, and is our data excluded from their training by contract rather than by their current default?
• Can we opt out, and what breaks if we do?
• How long is conversation data retained at every layer, including the provider's logs?

A good answer is Freshworks': yes to the first, here is the split, here is the opt-out, here is what it costs you. A bad answer is a link to a SOC 2 report, which describes how carefully data is handled and says nothing about what it is used for.

Check it yourself

Open the vendor's site and search its legal index for the word training. If the AI terms or trust page address it, the answer is usually one paragraph and easy to find. If you end up in a DPA reading about cross-behavioural advertising, the answer is not published, and you have learned something anyway.

Every row above is dated. If a vendor has published a clearer position since we read it, send it over and the record gets corrected and re-dated — that has already happened twice this month on other pages.

WRITTEN BY AR · UPDATED 2026-08-06

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

How we work · About the author