AI Policy
Version 1.0, effective September 5, 2026
What Katching sends to a language model, what for, what a model is never allowed to decide on its own, and why none of it is used for training.
The model provider
Anthropic, in the United States, is the only model provider the engines of Katching call. Its models draft text and read images. What Anthropic keeps of what it receives, and for how long, is described in its own privacy policy. It is named, with its role, on the subprocessor page.
What is sent, and what for
- Customer support: the message of your customer and the facts of the order it concerns, to draft a reply that waits for your click.
- Cart recovery and post purchase selling: the contents of the cart or of the order, to draft the message that goes with it.
- Disputes: the order, the delivery and the messages of the case, to draft the reply you send to your payment provider.
- The morning report: the numbers of the day of your store, to write them in words.
- Theme generation: your brief, your product photos and the page you asked us to replicate, to propose a theme you then edit.
- Risk assessment for Katching Payments: the free text you wrote about your business, to extract facts from it. The model returns facts, never a rate and never a decision.
What a model never does
- A model never decides an amount: not a refund, not a price, not a discount, not a payout.
- A model never sets your reserve rate. It reads text and returns facts; the rate is computed from those facts by code that gives the same answer to the same facts.
- A model never suspends a store and never closes an account.
- A model never sends anything to your customer on its own. Everything it writes waits for a click.
- No scheduled task in Katching calls a model.
Training
Katching does not train models, and does not use your data or the data of your customers to train one. What is sent to the model provider is sent to obtain an answer, under the agreement Katching has with it.
Cost, ceilings and traceability
Every call to a model is recorded with its price, and every store has a daily ceiling. When the ceiling is reached, the engines stop calling and the work waits. Nothing is sent to a model without a trace.
The list of models the product may call lives in one place in the code, and no other place can name one.
When a model is wrong
A model can be wrong. What it writes is a proposal, and you are the one who publishes it, sends it or acts on it. You are responsible for what leaves your store, as the terms of service say.
Version history
- Version 1.0, effective September 5, 2026First publication.