Map the Exit Before the Tool

The frontier labs are arguing about whether AI should slow down.

That debate matters, but it is not the part most companies can control.

The useful question is smaller: if an AI vendor became too risky, too expensive, or too misaligned next month, could your team move without breaking the work?

Markets only discipline vendors when customers can leave. AI is making that harder, not easier. The more useful the tool becomes, the more it turns into workflow infrastructure. Prompts live inside it. Review habits form around it. Automations depend on it. People stop documenting the process because the interface feels like the process.

Then the vendor changes price, policy, model behavior, roadmap, or risk posture, and the company discovers the exit was never designed.

That is not a reason to avoid AI. It is a reason to map the dependency before it becomes load-bearing.

List the real AI systems in use. Name the workflow, data touched, output affected, owner, and replacement difficulty. One table. No theater.

If a tool would be hard to replace in 30 days, it needs stronger rules now.

The market will not rescue an AI strategy with no exits.