“We do not train on your data” still means the file left

Training means the vendor uses your documents to make the model better for the next customer. “We do not train” means your draft is not supposed to become part of a product sold to someone else.

OpenAI’s enterprise privacy page says this is the default, and that an enterprise customer can decide how long its data is kept. For some API customers, OpenAI goes further. It offers what it calls zero data retention: the prompt is not stored after the answer is produced.

Read the three promises in the order they happen. The file is sent. The vendor says it will not use the file to train. The vendor may also say it will not keep a copy. The computer that read the advice still belongs to someone else.

A draft advice, a witness proof, or the strategy for a hearing is useful to the other side whether or not it was used to train a model. The risk is that it left. A copy that is kept can be seen by the vendor’s staff, produced under a lawful demand, or lost in an incident at the vendor. A copy that is deleted was still received.

Zero retention is a storage setting. It is not a statement that the processing happened inside the firm. If the product you are shown in the meeting does not even include that setting, you are being offered the weaker promise and asked to treat it as the stronger one.

Makebell runs on the firm’s own machines. The draft is read there. The chat is stored there. We do not receive the file, so there is nothing on our side to train on, nothing to retain, and nothing for us to be compelled to produce.

The company that trained the model does not receive the file either. You do not have to negotiate a no-training clause, a retention period, or a zero-retention add-on, because those clauses exist to control a copy that was sent. In our system the copy is not sent.

If a firm has already decided that a particular document can leave, a contractual promise is better than no promise. For client work, we do not think the firm should have to decide that file by file. The safer design is that the tool never needed the file to leave.