How to keep your firm’s know-how when you change AI models

Most companies are choosing an AI model the way people choose a football team, and then building their working lives around it. We think that is the wrong decision to spend energy on. The model you use this year will not be the one you use in two years. What matters is whether the way your company works survives the change.

At Joan we call that your firm's sovereignty. Sovereignty means your firm owns and controls where and how its work gets done: its know-how, its memory, its approvals, its record, its data and its choice of AI. This article explains what that covers, why it belongs to you, and four checks you can run this month to see how much of it you hold today.

The loyalty problem

There are ChatGPT companies, Claude companies and Gemini companies now, and their people defend the choice like a hometown team. That is understandable. When you use a model occasionally, it is easy to stay loyal.

Our founder, Latif Horst, uses these tools all day, every day, for real client work. He never runs just one. Usually two are open and one is checking the other's work. His view is blunt: loyalty is a casual-user habit. When you live in these tools, you see models have a bad month. The one he relied on most "turned into a stranger for a month" this spring, then went back to normal. Sometimes the vendor explains it weeks later; usually you simply notice.

That is not a complaint about any vendor. Models improve, change and get replaced constantly, and prices move with them. The practical conclusion is that a company should be able to change model the way it changes a supplier: with some care, and without losing anything that matters.

What your firm should own

An AI-enabled business has two kinds of things: what you rent, and what should be yours.

• What you rent. The model, the rented intelligence that reads, writes and reasons. It is powerful and replaceable. Most of the software that holds your records, such as the CRM, the accounting system and the document store, is rented too.
• What should be yours. This is what actually makes the work yours:
  • your know-how: the instructions, checklists, skills and steps that turn a model into a useful colleague;
  • your memory: what the AI needs to know about your business, clients and standards, and what has been learned, decided and written down;
  • your approvals: what each person, and each AI tool, may read, draft or do, and which person approves;
  • your record: who asked for what, what was produced, who checked it;
  • your data: where your information is kept, who can see it and whose terms apply;
  • your choice: which AI provider you use, and the freedom to change it without starting over.

Together, these are your firm's sovereignty. Data sovereignty is the part compliance teams ask about first, and our security page sets out how we build for it. The rest matters just as much: the know-how, the memory, the approvals and the record are how your firm actually works.

If any of it lives inside one vendor's product, or inside one employee's personal chat account, your firm does not own it. It is renting it, and the rent can change or the door can close.

Why this matters more than the model choice

A model change should cost nothing. If your instructions, knowledge and checks are held in a form you control, moving to a better or cheaper model is a configuration change and a round of testing. If they are buried in one product's settings and chat history, it is a rebuild.

People leave, and increasingly their know-how leaves with them. It used to be that when a strong person left, the way the work really got done left with them. Now there is a new version of the same problem: the know-how sits in that person's own AI account. Their prompts, their saved instructions, months of useful history. The account is theirs, so when they go, it goes. The following is an illustrative scene, not a client story: a monthly report quietly gets worse after a good analyst leaves, and only then does anyone discover she had been running it through her personal assistant for a year.

We want people using these tools. We just want the context, the instructions and the record of the work to belong to the company.

Outages and policy changes happen. Every AI service has bad days. A company that owns these things can switch to another model, or run in a reduced mode, instead of stopping. Continuity is the real test of sovereignty: if you cannot keep working when one rented piece fails, you do not have it yet.

What we recommend

1. Use more than one model on purpose. Not for fashion, for checking. Having a second model review important work catches mistakes that a single model will confidently repeat, and it keeps you practiced at switching.
2. Move the know-how out of personal accounts. Instructions, templates and reference material that people rely on should live in company-owned storage that any approved tool can read, not in individuals' chat histories.
3. Write down who may do what. For each kind of AI work, record who can ask for it, what it can read, what it can change and who checks the result. This is also the start of your security answer.
4. Keep the record. Keep a simple, company-owned record of important AI work: the request, the output, who reviewed it. It is how you learn what works, and how you answer a client or regulator who asks.
5. Test a switch before you need one. Once a quarter, run one real piece of work on a different model using the same instructions and knowledge. If that is painful, you have found where you are locked in.

Four checks you can run this month

• Where does our know-how live? List the five AI-assisted tasks your company depends on most. For each, find where the instructions and reference material are kept.
• Whose account is it in? If any of those five depend on one person's personal account, you have a continuity risk today.
• Could we switch model next week? Ask your team what it would take. "Rebuild everything" is an answer worth hearing now.
• Can we show the record? Pick one important AI-assisted output from last month. Can you show who asked for it, what was used, and who checked it?

A note on what this is not

This is not an argument against any AI vendor, or against committing to a good platform. Most companies should standardize on a small number of well-governed tools. The point is narrower: the tool is rented; how your company works should not be.

Last reviewed: October 1, 2026