AI data use
Last updated: 13 November 2026
This is a translation for convenience. Only the German version is legally binding.
Stitchomat only uses AI where it is needed: to draw an image draft from a motif description. The stitches are placed by an embroidery engine without AI. Here is which data goes to which model – and which doesn’t.
By function
| Function | AI? | What goes to a model | What doesn’t |
|---|---|---|---|
| Names, text, monograms from the name field | no | nothing – set in embroidery fonts on our servers | – |
| Logo upload | no | nothing – the logo is vectorized without AI | the logo itself |
| Motif wish in your own words (“A dachshund with F and M”) | yes, language model | the entered wish text, so the model can separate picture and text to embroider and detect inadmissible content | no contact details, no order code |
| Image draft (AI preview) | yes, image model | only the image description derived from it, in English, e.g. “a dachshund with long ears”, plus style instructions | the separated names and text, no contact details |
| Embroidery file, stitch preview, traffic light | no | nothing – runs on our servers | – |
Names that someone types into the wish text are therefore seen by the language model. The image model does not get them: the language model separates them first, and they are set in embroidery fonts.
Models in use
- Image model: Google Gemini (Nano Banana 2.1), called via OpenRouter.
- Language model for splitting the input: Anthropic Claude Haiku, called via OpenRouter. Depending on routing it is operated by Anthropic, Amazon Web Services, Google or Microsoft.
- The providers process data in the USA or worldwide. Providers, locations and legal bases are listed under Subprocessors.
Storage and training
We do not train models with your data or your customers’ data. OpenRouter does not use transmitted content for training either.
Whether a model provider stores content or may use it for training is governed by its terms. We have not excluded this contractually so far, and the image model currently has no endpoint without storage (zero data retention). That is why we only send the models what the draft requires – and no selfies at all.
Labelling and control
- Drafts from the image model are labelled “AI draft” in the interface.
- Before every call a blocklist checks the input for clearly inadmissible content (e.g. unconstitutional symbols, pornography, insults); the language model checks as well. Third-party trademarks and protected characters are not reliably detected – that remains the business’s responsibility.
- Every embroidery file goes through the traffic-light check, and no file reaches the machine without the business’s approval.
- There is no selfie or photo mode.
- Daily and session limits cap AI use; AI drafts are only available via the business’s stand QR code or kiosk.
Questions: datenschutz@golle-it.de.