Search for whether AI support tools train on your customer conversations and look at who answers. Crescendo. Freshworks. Fin. CustomGPT. Help Scout. Every result on the first page sells a support tool, and every one of them concludes that the responsible approach happens to be theirs.
So we did the boring version. We opened the AI policy page for three platforms, read what each company had actually committed to in writing, and scored it. Same criterion, same weight, same date. Here is what they say.
| Help Scout | Zendesk | Intercom (Fin) | |
|---|---|---|---|
| Model provider trains on your data | No | No | No |
| Vendor trains its own models on your data | Limited samples only | Yes, with a control | Yes, anonymised |
| Default position | Off | Not stated | On |
| Opt out available | Yes | Yes | Yes |
| Score, out of 100 | 80 | 65 | 60 |
There are two questions here and most people only ask one
When a vendor tells you it does not train on your data, it is usually answering about OpenAI. That is the easy half. Every one of these three has a written commitment that its model provider will not use your inputs for training, and all three run generative features on zero retention endpoints. If your worry is that a customer’s angry message about a late delivery ends up baked into somebody’s foundation model, you can stop worrying. That question is settled across the category.
The harder question is what the support vendor does with your data for its own models. That is where these three go in completely different directions, and it is almost never what the marketing page is talking about.
Help Scout: off by default, and it says so
Help Scout’s AI transparency page contains the sentence we were looking for and did not find anywhere else: no global retraining on your account data. It goes on to say it does not push customer data back to OpenAI to aid training of the global foundation model.
AI Answers is off by default. Somebody at your company has to go into Beacon settings and switch it on before a single customer conversation touches it. You can ask support to remove your data entirely.
It scores 80 rather than higher for one line in that same page: limited samples of customer data may be used to evaluate and improve feature quality, such as tuning adapters for tone. That is a real carve out and it is honest of them to print it. The page also does not give a retention period, pointing you at the privacy policy and the data protection amendment instead.
Zendesk: more machinery, less clarity about the default
Zendesk documents its process in more detail than either of the others. Before any of its own models see your data, identifier fields like username and email are excluded, free text is de identified using natural language processing, and what is left gets tokenised. There is a customer control over whether it happens at all.
That is genuinely more engineering than Intercom describes, and it earns the extra five points. What Zendesk’s own AI data use page does not say is which way that control points if nobody touches it.
We do not fill in a default a vendor has not written down. It might well be off. If it is, Zendesk should say so on that page, because right now the most cautious reading and the most generous reading are both available to anyone who reads it, and only one of them is true.
Intercom: on by default, and honest about it
Fin may use anonymised customer data to fine tune models by default. You can opt out at any time, and deletion follows within thirty days.
Sixty is not a punishment for the opt out being difficult, because it is not. It is a straightforward setting. The score reflects that the default is on, and defaults are what most accounts run under forever. Nobody audits their AI settings in month fourteen.
Worth saying: Intercom writes this down plainly, in one place, without hedging. That is more than we can say for some of the category, and the 60 sits above the score a vendor would get for saying nothing at all.
What to ask, if you are the one filling in the security review
Four questions, in this order. The first one usually gets a yes and tells you nothing.
- Does your model provider train on our data? Expect no. Every vendor in this category can answer this cleanly now.
- Do you train your own models on our data? This is the question. Ask it separately and in writing, because the answer to the first one is often offered as though it answered this one too.
- If yes, is it on or off when we sign? Get the default in writing. An opt out you were never told about is not really an opt out.
- What is the deletion window after we turn it off? Intercom publishes thirty days. The other two do not publish a number.
What this does not tell you
These are scores for what each company has published, not for how carefully any of them behaves. We cannot audit a training pipeline from outside. Nobody writing a comparison of these products can, including the vendors writing about each other.
What we can check is whether a company was willing to put its position in public where a buyer can hold it to the words. Help Scout was clearest. Zendesk was most detailed and least clear about the part that matters most. Intercom was direct about a default that costs it points.
This is one criterion out of six. On the full index, weighing pricing, export, subprocessors, uptime history and attestations as well, the order shifts and Zendesk drops further back. The scorecards carry the evidence and a source link behind every number:
- Help Scout, 72.8 overall
- Intercom (Fin), 70.5 overall
- Zendesk, 61.2 overall
- The full AI customer support ranking
Scores here are append only. When any of these policies changes, the new reading lands as a new dated row and the one above stays on the page next to it. If a vendor thinks we have read them wrong and can point at a URL, that correction gets published the same way.
Update, 19 August 2026: Freshdesk has since been scored on this criterion and came in at 50, below all three above. Freddy AI trains Collaborative Models on de-identified data drawn from multiple customers, so one company’s conversations improve a model serving others, and opting out means emailing support rather than changing a setting. The Freshdesk scorecard is here. We also compared what all four charge for AI, and found they do not use the same billing unit: four tools, four units.