How hard it would be to move. Prompts tuned to one model, a fine-tuned model that cannot be exported, embeddings that only work with one provider’s model, and data held in a proprietary format all raise the cost of leaving. Some of that is an acceptable price for capability; the point is to know what you are accepting.
Technical view
Lock-in sources include model-specific prompt tuning, provider-specific tool and structured output formats, fine-tuned weights that cannot be extracted, embeddings requiring full re-indexing on model change, and proprietary orchestration. Mitigations include an internal abstraction over providers, maintaining a portable evaluation suite to qualify alternatives quickly, exportable data formats, and periodically testing a secondary provider rather than assuming portability.
Related
Part of AI Terminology Index, section: Product and Adoption.