What it means
AI lock-in is subtler than in traditional software, because the obvious layer is the most portable. Basic chat completion is close to a commodity and swapping providers is often a small change.
The real costs sit elsewhere. Prompts tuned against one model's quirks degrade against another. Embeddings are the hardest single dependency — vectors from different models are not comparable, so changing embedding models means re-indexing your entire corpus. Fine-tuned models do not transfer at all. And proprietary features — specific tool-calling formats, caching behavior, vendor-specific agent frameworks — each add friction.
The countervailing force is unusually strong here: capability leadership changes hands frequently, and prices fall fast. The ability to move is worth more in this market than in most.
Why it matters
Being unable to switch is expensive in a market where the best option changes every few months. And because lock-in accumulates through ordinary implementation choices rather than through a contract, most teams don't notice until they try to move.
In practice
Abstract model calls behind your own interface from the start — it is nearly free early and expensive to retrofit. Keep an eval set that lets you compare providers on your own tasks. Treat the embedding model as your most consequential lock-in decision.