Give AI your company's context without giving it away.

The accurate answer is calmer than the worried one. What you type is processed on the vendor’s systems under the terms of your plan. Under most business and enterprise plans, chats are excluded from model training by default, and the major vendors publish real security practices. If training or confidentiality is your concern, the current terms of your agreement are the authority, and they are worth actually reading.

The longer-lived question is not any single chat. It is the accumulated context: the memories the product saves, the custom instructions, the project setups, the thousands of corrections your team has typed in — this client is an exception, we decided against that in March, here is how we phrase it. That accumulation is real operating knowledge, and it builds up inside the platform’s account structure. Under many enterprise licenses, when you leave, it stays behind. Chat logs may export; the working shape of what the platform learned about you generally does not transfer anywhere useful.

The observation this page rests on: the practical risk is not that the vendor sees your context. It is that you cannot take the context with you. That is an ownership problem, not a privacy scandal, and it has an unglamorous fix.

Write the context down once, in plain text, in files the company controls: what has been decided, where the exceptions are, how the work gets done, how the company sounds. Then let whichever AI tool your team uses read those files at the moment of work — pasted, attached, or connected so the tool reads the folder directly.

Everything the platform’s memory was quietly accumulating now exists as documents you can open, correct, version, and hand to the next tool. Nothing about your operation is trapped inside anyone’s platform. The walled garden becomes just a tool you rent, because the garden no longer holds anything you cannot walk out with. Lock-in, in practice, is rarely the contract; it is the accumulated context that makes leaving expensive. Files remove that expense.

The set of files is small — the same five shapes described in context engineering: how-we-work, decisions, exceptions, handoffs, voice. A page or two each, written by people who know the business.

Less than a migration and more than nothing. People keep the assistants they already use. What changes is the first minute of the work: instead of re-explaining the company to a tool that forgot it, they point the tool at the files. Answers start from the company’s actual decisions and exceptions instead of from a general idea of businesses like yours.

Corrections change too. Today, when someone teaches the tool something — that is not how we invoice, this client has special terms — the correction lands in one person’s chat history and dies there. With owned files, the correction goes into the document, once, and every person and every tool reads the corrected version from then on. The company compounds what it learns instead of re-teaching it seat by seat. Why the tools cannot hold this for you is covered in why AI forgets.

Two kinds of context, two owners, written down. Knowledge about the company — its decisions, processes, exceptions, client history — belongs to the company: kept in files the company governs, with named owners, an update rhythm, and plain rules about which roles and which tools may read what. A person’s own working memory — their notes, their drafting habits, the way they think through a problem — stays with the person, in files they keep, exactly as a notebook would. Each side is honored: the company is not mining personal notebooks, and no individual’s departure takes the operating manual with them.

On top of that boundary, the rules are ordinary governance: which sources may enter the shared files, which vendor processes what under which agreement, and what gets logged. None of it requires distrusting a vendor. It requires deciding, in writing, what is yours — the individual version of the same move is a personal AI brain. Taken together, this is owned, portable working memory: the company’s knowledge in the company’s hands, in plain files, readable by any model it chooses to rent.