Choosing between ChatGPT, Claude and the rest, and switching without losing what they've learned
Pick one for the company, let people keep a second for themselves if they want, and decide the boring things first – the contract, the admin controls, the connectors to your systems. The models are closer than the marketing says, and what a tool "learns about you" is smaller and more portable than people fear.

What actually differs
Less than you'd think, and it changes every quarter. The leading assistants are close on writing, summarising and reasoning over your documents, and each has months where it's ahead on something. Choose on model quality alone and you're re-choosing every time a new version ships. The things that differ for a business, and stay different, are around the model:
| What to check | Why it matters | What a good answer looks like |
|---|---|---|
| Business terms | Whether your data trains their models, where it lives, what happens when you leave | A business or enterprise agreement, training off, deletion and export in writing. Protecting IP and company data has the questions to ask. |
| Connections | Whether it can read your mailbox, calendar, files and CRM without copying and pasting | Connectors to the systems your people already use, with permissions you can see and revoke. Standard connectors over proprietary ones where you've got the choice. |
| Admin controls | Who has a seat, what they may connect, what you can see | Company-owned accounts, usage you can read, settings applied to everyone at once. |
| Your own tasks | Whether it's good at what your people actually do | The five-task test below, scored by the people who'll use it. |
The five-task test
Pick 5 real tasks from last week: a difficult customer email, a summary of a long supplier document, a first draft of a proposal section, a plain-English explanation of a contract clause, a plan for a meeting. Give each tool the same brief, written the way you'd brief a person, with the same material attached. Have the people who do those jobs score the results out of 5, without knowing which tool produced which. Half a day, and you'll know more than any review will tell you.
The answer will usually be "they're both fine, and this one's a bit better at our kind of writing". That's a real answer, and enough to decide on.
Pick one tool for the company
Choose one tool as the company's. The library of material it can read, the training sessions, the one-page rule, the agents – all of that gets built once, for one tool, and that's the point. Run two company tools and you're building everything twice and confusing everyone about where to go.
Individuals may well prefer a different one for their own thinking, and there's no harm in that, as long as the never-list applies and company data stays in the company tool. What you shouldn't do is let a department adopt its own company tool – that's how silos start.
What the tool actually knows about you
The worry about switching is that the tool "knows you" and a new one would start from zero. In practice that knowledge is smaller than it feels:
- Saved instructions (custom instructions, system prompts). These are yours, and portable in a document. Keep the master copy in your own files, not only in the tool.
- Projects and attached material. Your library. If it lives in one place in your own systems and is connected to the tool, it moves with you. If it was uploaded piecemeal into the tool, it doesn't.
- Memory features. The tool's own notes about your preferences. Useful, small, and rebuilt in a fortnight of use.
- Chat history. Large, and almost never needed again. Export it if the terms allow and move on.
Kept that way, switching costs an afternoon. Connect the new tool to the same library, paste in the same instructions, re-run the five-task test, and tell everyone. The relationship was never really with the tool. It was with your own material, and you own that.