If those traces accumulate inside a vendor's memory, agent state and workflow, the firm may retain legal ownership of its data while losing operational ownership of its expertise. To be clear,
@OpenAI and
@AnthropicAI say enterprise customer data is not used to train their models. The immediate danger is not a heist. It is a quieter form of absorption: your institutional memory becomes most useful inside a system whose interface, memory and data portability you do not control. And despite the reassurances on the prompts, the related meta-data, tool use, and interactions with the software are still likely to become an integral part of both labs’ training.