Why we do not train on your customer conversations
Published on March 5, 2025
A support inbox is an uncomfortably honest dataset. Addresses, order values, complaints, sometimes the medical reason behind a return. A vendor who turns that into a better product for everyone is quietly selling your customers to your competitors. We do not do that.
What the policy rules out
Your conversations do not feed models that serve other merchants. What your agent learns about your tone, your goodwill rules and your standard cases stays in your account. If you cancel tomorrow, that knowledge does not travel into the next shop's account as a small head start.
The infrastructure carries the same idea. We encrypt with AES-256 and host on GDPR-compliant German servers. A data processing agreement is available without you having to ask for it.
Why we think this is an advantage
The usual objection is that a model gets better with more data from more places. For generic answers that is true. For support it is not, because the valuable part is not the average of a thousand shops, it is your exception. Your shipping windows, your goodwill limit, how you handle someone writing in for the third time.
An agent trained on a thousand other shops sounds like a thousand other shops. We would rather it sounded like you.