Build vs buy is the wrong AI question. Ask what it costs to leave
For AI, the more useful question is how much it costs you to change your mind later. The leading model changes every few weeks, the cost of the same task varies widely between providers, and your vendor's pricing power equals your migration cost. Build an abstraction layer early, hold your own API keys where you can, and keep your exit cheap.
Key takeaways
- The best model has a shelf life measured in weeks.
- The same agentic task can cost several times more on one model than on another.
- A vendor's pricing power over you equals your cost of migrating away.
- Bring-your-own-key is a procurement position, not a developer convenience.
- The moat to protect is your ability to walk away.
Why is build vs buy the wrong question?
I am building a multi-agent legal pipeline. Almost every architecture decision I've made this year came down to one thing. Not which model is best, but how expensive it is to change my mind later.
Why does switching cost matter more than model choice?
- The best model has a shelf life measured in weeks. In July 2026 alone, OpenAI shipped a new model family and a Singapore lab released a free model with frontier-level benchmark results. Anything hard-coded to one provider becomes technical debt within months.
- The cost spread is wide and moving. The same agentic task can cost several times more on a frontier model than on a cheaper one. Independent benchmarks keep finding wide multiples between comparable systems. If you can't route between models, you can't capture any of that.
- Switching cost is the pricing power you hand over. A vendor's power over you equals your cost of migrating. It isn't a metaphor. It is the number. If it is high, every future price change is something that happens to you.
Why does bring-your-own-key matter?
Whoever holds the key holds the cost base, the data terms and the exit. That makes bring-your-own-key a procurement position, not a configuration setting. When you buy AI software, ask whether you can use your own model account, and what the vendor's terms say if you do.
How do you keep your exit cheap?
- Put an abstraction layer between your product and the model. Change providers by configuration.
- Keep prompts and evaluation sets provider-neutral. Then you can test a new model in days, not months.
- Keep your data exportable. Make it a contract term, not a hope.
- Run a second provider regularly. An untested fallback isn't a fallback.
Build the abstraction layer before you need it. The moat you're protecting isn't the model. It's your ability to walk.
Frequently asked questions
What is AI vendor lock-in?
The position where moving to another AI provider would cost so much in rework, data migration or retraining that you accept whatever terms the current provider sets.
What does bring-your-own-key mean for AI tools?
The customer holds its own account and API key with the model provider, and the software vendor uses it. The customer keeps control of the cost base, the data terms and the ability to switch.
How do you reduce AI switching costs?
Route model calls through an abstraction layer, keep prompts and evaluation sets provider-neutral, keep your data exportable, and test a second provider regularly.
This article is general information, not legal advice. It reflects the position as at the date of publication. A plain-text version for AI assistants is at /blog/ai-build-vs-buy-switching-costs.md.