3 min read
Why we will never train on your documents
Pilot reads your files to answer your questions. It never uses them to train a model. What that promise covers, how we enforce it, and what it costs us.
Oskar NilssonHead of Security
On this page
The question I’m asked most in security reviews is not about encryption or certifications. It’s simpler: will our documents be used to train your AI?
The answer is no, and it will stay no. This post explains exactly what that covers, how we make sure it’s true, and why we decided it was worth what it costs.
What the promise covers
When we say we don’t train on customer content, we mean all of it:
- Documents and files in Library
- Prompts, instructions and conversations
- Drafts, edits and comments in Write
- Autopilot agent inputs and outputs
- Brand voice rules and examples
None of this is used to train or fine-tune any model, ours or a provider’s. That applies on every plan, including the free one. It isn’t a setting you have to find and switch off.
What we do use
We’d rather be precise than reassuring, so here is what we do collect.
Usage data. Which features are used, how long requests take, error rates. This tells us what to fix. It never includes the content of your documents or prompts.
Feedback you choose to send. When you rate an answer or report a problem, you can attach the conversation. Those reports are read by a small team, used to investigate the issue, and deleted after 90 days.
Evaluation sets, with consent. Some customers share documents with us specifically to test answer quality, under a separate written agreement. We remove personal data, and we delete the set when the agreement ends.
How we enforce it
A promise is only as good as the systems behind it. Ours rests on three things.
Contracts with model providers. Every provider we use is contractually barred from training on data we send them and from keeping it beyond the request. We review these terms each year and whenever a provider changes them.
Separation in our own systems. Customer content lives in storage that our training and evaluation pipelines cannot read. Access requires a named engineer, a ticket and approval from a second person, and every access is logged.
Outside audits. Our SOC 2 Type II report covers these controls. An independent auditor tests them every year, and Enterprise customers can read the full report in our trust center.
What it costs us
Not training on customer content makes some things harder. Models that learn from real usage improve faster. Some of our competitors are open about doing exactly that, and it’s a reasonable choice for consumer products.
For us it would break the reason teams use Pilot. A consultancy’s proposals, a law firm’s memos and a product team’s roadmap are the most sensitive writing those companies have. They can’t put it in a tool that might learn from it, however small the risk.
So we improve the product in other ways: evaluation sets built with consent, research on public data, and close work with customers who tell us where answers fall short.
Questions to ask any AI vendor
If you’re evaluating AI tools, these are the questions I’d ask, of us or anyone else:
- Is customer content used for training on every plan, or only on paid ones?
- Does the promise cover prompts and outputs, not just uploaded files?
- Are your model providers contractually barred from training on our data?
- How long do you keep our content after we delete it?
- Is any of this covered by an independent audit?
If an answer is vague, ask for it in writing. Ours are in our data processing agreement, and you can request a copy from the security page.

Written by
Oskar Nilsson
Head of Security
Keeps the SOC 2 report current and the permissions model strict.


