Choice Is the Best Protection Against AI Lock-In

A global manufacturer signs a multi-year deal with a single AI vendor. The pricing is great. The integration is fast. The pilots ship in weeks.

Eighteen months later, a new regulation requires that certain workloads run in-country. The vendor has no region there. A better model is now available from a different provider, but it would take six months to migrate because all the prompts, agents, and semantic mappings are inside the vendor's UI.

The decision that felt smart in month one is now the decision that gates every roadmap conversation in month eighteen.

The fastest path to a working pilot usually goes through one vendor's full stack. That is fine for a prototype. It becomes painful the moment a better model appears, a cheaper provider emerges, a regulator asks a hard question, or a customer requires data residency you cannot offer.

Lock-in is rarely a single bad decision. It is the accumulated weight of many small conveniences.

Choice is the fourth of the four C's, and it is the one most teams underweight until the day they need it. It is also the one you cannot retroactively add — by the time you need optionality, the migration is already too expensive.

Buy the insurance early, while it is still cheap.

Choice means your business meaning, your governance rules, and your agent contracts live in formats you own — not inside a vendor's product. The compute, the models, and the surface UI can be swapped. The brain of the system cannot be held hostage.

Open table formats. Portable semantic layers. Model-agnostic abstractions. Clean interfaces between agents and tools. None of these are architectural purity. They are migration insurance.

Manufacturing has known this game for a century. No serious supply chain depends on a single supplier for a critical part. A 'second source' is required for every component that matters.

AI architecture is now in the same category. The model is a part. The vector store is a part. The agent framework is a part. Every critical part needs a credible second source, even if it is never used. The day you need it, it is already too late to build it.

Pick your most important AI feature. Ask: if we had to swap the model tomorrow, how long would it take? If we had to swap the cloud, how long? If the answer is 'months,' you have a lock-in problem in disguise.

Then identify the smallest change that would move that answer to 'days.' Usually it is moving one abstraction — the prompt template, the retrieval logic, the agent loop — out of the vendor's UI and into your own repo.

Data architects: portability seams are now a design responsibility, not a wishlist item. Platform engineers: maintain at least one second-source path for every critical AI component. Data leaders: include a 'portability budget' — a small percentage of every project's time — to keep the seams clean.

Optimizing every decision for short-term velocity. Picking the path of least resistance for prompts, retrieval, evaluation, and observability — all inside one vendor — because each individual decision was reasonable.

Optimize for speed inside a vendor. Optimize for portability at the seams between vendors. Use open formats wherever they exist. Wrap proprietary services behind interfaces you own. Test the second source on a quiet Friday once a quarter — not on the day you discover you need it.