Can a company trust its data to Anthropic or OpenAI? Ask where the memory lives
The model providers' contracts answer the training question. The bigger question is who owns your AI's memory and context, because that is where the intellectual property now sits.
Every board I speak to asks a version of the same question: can we trust our data to Anthropic or OpenAI? It is the right instinct and the wrong question. The contracts largely answer it. What they do not answer is who will own the most valuable thing your AI programme creates.
The training question is mostly answered
Both companies say that, by default, they do not train their models on data from their business products: the API and the enterprise and team plans. Consumer products are different. On the free and personal plans, conversations can be used to improve models unless the user opts out. So the first rule is simple: company data goes through business accounts under a business contract, never through someone’s personal login.
Two caveats remain. First, a promise not to train is not a promise not to keep. Providers retain some data for safety monitoring, and in 2025 a court order required OpenAI to preserve some customer content pending litigation. A legal hold overrides a deletion policy. Second, your data is subject to the laws of wherever it is processed. Retention periods, processing regions, subprocessors and deletion on exit all belong in the contract, and most of them can be negotiated by a serious enterprise customer.
Handled properly, a company can trust these providers with its data in the way it already trusts a cloud provider or a bank. That is not where the real risk lies.
Where the real value now sits
The model is becoming a commodity. As I wrote in my last piece, a better one arrives every few months. The value sits in the harness around it, and specifically in its context and memory: everything the system learns about your business while you use it.
That is memory and context:
- Memory is what the system accumulates over time: how your team makes decisions, which deals you passed on and why, which customer prefers which terms, which analyst’s numbers need checking.
- Context is what you assemble for each task: the right documents, the right history and the right rules, retrieved and ordered so a general model can do a specific job well.
Together they are the part of the harness that turns a generic model into your firm’s institutional knowledge, made usable. That knowledge used to live in people’s heads and leave when they did. Now it can be captured, and whoever holds it holds the compounding asset. Context and memory are becoming the core intellectual property of a company. They are IP gold.
The risk is not theft. It is lock-in
If that memory lives inside a provider’s product, in its built-in memory, its hosted projects and its managed agents, you may be building your most valuable asset on someone else’s balance sheet. You cannot easily take it with you. Switching models means starting again. And the provider learns, at least in aggregate, how work in your industry is done, even if it never trains on your files.
None of this needs bad faith from the provider. It is simply what happens when the asset accumulates in the vendor’s system rather than yours.
What I recommend
- Rent the model, own the memory. Keep memory stores, document indexes and knowledge graphs on infrastructure you control. Send the model only the context each task needs. That is how the agent systems I build are designed.
- Own your traces and evaluations. The record of what your agents did, and the tests that tell you whether they did it well, are what let you switch providers in days rather than months.
- Make memory portable. Use open formats and open protocols so that memory built today can be read by next year’s model, whoever makes it.
- Negotiate the exit at the start. Agree no training, retention limits, processing regions, subprocessors, deletion and data export before the first pilot, while you still have leverage.
So yes, a company can trust its data to Anthropic or OpenAI. What it should not do is hand over its memory. The model is rented. The memory is the company.