AI vendor lock-in: check the exit before you enter

AI tools earn their keep by absorbing your work: your prompts, your documents, your saved workflows, your conversation history. That absorption is the value, and it is also the trap. Eighteen months in, the question “could we switch?” can have a five-figure answer. The time to buy the exit ramp is before you enter, when it is free.

Why AI lock-in is sneakier than software lock-in

Classic software lock-in is visible: your data sits in their database, and everyone knows migration is a project. AI lock-in accumulates invisibly, in pieces that do not look like data.

The prompts and instructions you refined over months. The custom assistants configured with your business context. The conversation history your team searches like a knowledge base. The integrations wired into your other tools. The habits themselves: a team fluent in one tool resists the next one. None of this appears in an export file by default, and all of it is switching cost.

The trend that sharpened this: agents increasingly run in vendor clouds rather than on your machines, holding work in progress, not just chat history. Our briefing on OpenAI buying Ona is this dynamic in one headline: the working environment itself becomes part of the subscription.

The exit-ramp checklist

Run this on any AI tool before it becomes load-bearing, alongside the selection filters from how to choose AI tools for a small business:

1. Export: Can you get your data out, in a format another tool could read? Conversations, files, configurations. Test the export once, actually, with a real download. An export button that produces an unreadable archive is theater.
2. Prompts outside: Are your prompts, briefing packs, and recipes stored in your own files, pasted in when needed? Or do they exist only inside the vendor’s “custom assistant” where the only copy lives on their servers?
3. Source documents outside: The tool can index your documents, fine. The originals live in your storage, and the tool holds a copy, never the only copy.
4. Plain-language workflows: Could you hand a one-page description of each AI workflow to a different tool tomorrow? The recipe discipline from the automation playbook doubles as portability insurance: readable recipes move, visual flows locked in a vendor editor do not.
5. Standards where they exist: Integrations built on plain APIs and common formats survive a vendor change better than ones built on a vendor’s proprietary plugin system.

If the answer to the first four is yes, your real switching cost drops to retraining habits, which is an afternoon of grumbling rather than a quarter of migration.

Practice the cheap habits

Lock-in resistance is mostly habits, not heroics.

Keep one folder, in your own storage, holding your context pack, your refined prompts, and your workflow recipes as plain text. Update it when they change. This single folder is most of your AI operation’s portable soul.

Once or twice a year, run a small real task through a competitor. Not a migration, a temperature check: you stay aware of what the market offers and what switching would actually involve, and renewal conversations go differently when leaving is visibly possible.

And before adopting any deep integration, write down what leaving would take, in steps, on one page. If you cannot write the page, that is the answer.

The trade is fine if you make it on purpose

None of this says avoid deep adoption. Sometimes the integrated tool is genuinely worth being locked into, the way many businesses happily live inside one accounting platform for a decade. The failure is not dependence. It is discovering dependence, later, with the renewal letter in hand and no file of your own prompts.

Buy convenience with your eyes open, keep the one folder current, and the vendor relationship stays what it should be: a choice you keep making because it keeps being good, not because leaving became unthinkable.