The exposure usually becomes visible when a pricing change arrives or a model gets deprecated. By then, the application has months of prompt engineering optimized for one model’s behavior and an embedding index that only makes sense in one vector space. The cost of switching is now the cost of rebuilding those investments, not just updating an API key.
Four places AI lock-in accumulates
The most visible form is API lock-in: different providers use different endpoints and incompatible request schemas. An application built against one provider’s function-calling schema requires code changes to run against another. Cloud computing research identifies this as a well-documented pattern: customers are often unaware of the proprietary standards that block interoperability until they attempt migration (Opara-Martins et al., 2016).
Less visible, and stickier:
Embedding lock-in. Embedding models map text into high-dimensional vectors where similar meanings cluster together. Those clusters are specific to the model that generated them. A retrieval system built on one provider’s embeddings cannot be compared against vectors from a different model. They occupy incomparable mathematical spaces. Switching embedding providers means re-indexing every document in the corpus.
Fine-tuned weights. Fine-tuning a proprietary base model produces weights that adapt an architecture you don’t own. The fine-tuned model is yours, but the base it builds on is not portable to another provider’s infrastructure. Open-weight models address this specifically: the base parameters are public, so any adaptation you build on top remains independently deployable.
Prompt engineering accumulation. Prompts are written and tested against one model’s specific behavior: its tendency to hedge and its failure modes under particular phrasings. Sustained iteration produces prompts calibrated to one model’s quirks. Porting them to a different model typically requires re-testing and often partial rewrites, particularly at the system-prompt level.
Why the risk compounds
Standard software lock-in is mostly static: the cost is fixed at adoption. AI lock-in grows with use. Each prompt refinement tuned to one model and every document added to the embedding index increases the migration cost. A system that has been in production for a year is harder to move than the same system was at launch. Not because the API became harder to replace, but because the behavioral dependencies have accumulated.
Open-weight models as the main countermeasure
Running an open-weight model (one whose parameters are publicly released) removes API dependency entirely. The model runs on your own infrastructure; the provider cannot unilaterally change pricing or retire the model. Research comparing open-weight alternatives against proprietary API services found production cost reductions of 5x to 29x, with competitive accuracy on narrow, well-defined tasks (Irugalbandara et al., 2024).
Open-weight models don’t eliminate all lock-in vectors. You still accumulate prompt engineering specific to a given model’s behavior, and you take on the operational cost of running inference infrastructure. They do remove the two most acute risks: pricing dependency and unilateral model changes by the provider.
For the trade-offs between self-hosted open-weight models and cloud-hosted APIs across cost and operational overhead, see how to think about choosing an AI model.
The regulatory response
The EU Data Act (Regulation (EU) 2023/2854, entered into force January 2024, applicable from September 2025) targets switching barriers directly for cloud and data processing providers. It requires providers to support data portability in interoperable formats and to disclose switching costs transparently in contracts; unreasonable egress fees must be eliminated by January 2027. AI infrastructure providers (cloud platforms hosting training pipelines and inference APIs) fall within its scope where they meet the regulation’s service definitions.
What the regulation does not reach: model-behavioral lock-in and prompt/embedding incompatibility. Those are architectural constraints, not contractual ones.
The stickiest layer isn’t the one you are watching
Most lock-in analysis focuses on API pricing and contract terms: the visible exposure. The harder cost is behavioral: the embedding index built against one model’s vector space and the prompt system refined over months for one model’s quirks. None of it appears on an invoice. It appears when you try to leave.
Sources
- Opara-Martins, J. et al. Critical analysis of vendor lock-in and its impact on cloud computing migration: a business perspective. Journal of Cloud Computing 5, 4 (2016).
- Irugalbandara, C. et al. Scaling Down to Scale Up: A Cost-Benefit Analysis of Replacing OpenAI’s LLM with Open Source SLMs in Production. arXiv:2312.14972, April 2024.
- European Parliament and Council. Regulation (EU) 2023/2854 of 13 December 2023 on harmonised rules on fair access to and use of data (Data Act). Official Journal of the European Union, 22 December 2023.
Questions, answered
Is AI vendor lock-in different from regular software vendor lock-in?
In one important way: beyond API and format dependencies, AI systems accumulate behavioral dependencies. Prompts optimized for one model's quirks and fine-tuned weights built on a base architecture you don't own both grow with use, unlike most software lock-in which is fixed at adoption.
Does using open-weight AI models eliminate vendor lock-in?
Mostly. Open-weight models, whose parameters you can download and self-host, remove API dependency and per-token pricing exposure. Research found cost reductions of 5x to 29x when replacing proprietary LLM APIs with self-hosted open-weight alternatives. You still accumulate prompt-behavioral dependencies, and you take on inference infrastructure costs.
What is embedding lock-in in AI?
When you build a retrieval system using one provider's embedding model, the resulting vectors are only meaningful for similarity search within that model's dimensional space. Switching to a different embedding model requires re-indexing your entire document corpus. The vectors are mathematically incomparable across providers.
Does any regulation address AI vendor lock-in?
The EU Data Act (Regulation (EU) 2023/2854, applicable September 2025) requires cloud and data processing providers to support data portability in interoperable formats and to disclose switching costs in contracts; unreasonable egress fees must be eliminated by January 2027. It applies to AI infrastructure providers within its scope, though model-behavioral lock-in remains an architectural, not a contractual, problem.
Brief is a team of AI associates you direct in plain language. Opening to a small group at a time.
Request access