AI concepts
Vendor lock-in (dependence on an AI vendor)
A state in which migrating from one AI model vendor to another becomes economically irrational. In the AI context it stems from three sources: an incompatible API, incompatible prompt semantics (model A and model B interpret the same instructions differently), and fine-tuned customisation that cannot be transferred. The cost of migration typically equals the cost of the original deployment.
Primary source: MIT Sloan Management Review 2025, Gartner Cloud AI Services Magic Quadrant 2026
In classic software lock-in (SAP, Oracle) the main mechanism was custom development of the application layer. In AI the lock-in mechanism is more subtle: every LLM understands language differently, a prompt that works on GPT-4o does not work on Claude 3.7 Sonnet without rewriting, and fine-tuned models are non-migratable by definition.
Three levels of lock-in
API level. The shallowest. OpenAI, Anthropic and Google use different call schemas and different response formats. A wrapper library (LangChain, LiteLLM) solves this problem in a couple of hours of work.
Prompt level. Medium depth. A prompt optimised for GPT-4o, giving 92 percent accuracy, may give 71 percent on Claude without modification. Prompt engineering is a capital investment tied to a specific model. Migration means re-engineering the prompts for the entire application.
Fine-tuning and RAG index level. The deepest. Fine-tuned models (OpenAI Custom Models, Anthropic Constitutional AI) are non-migratable. RAG indexes embedded with a specific vendor's embeddings require re-embedding the entire knowledge base.
The scale of migration cost
MIT Sloan Management Review, in its analysis of 47 enterprises from 2025, reports an average cost of migration between LLM vendors of 60 to 110 percent of the original deployment cost. The median is 87 percent. For a deployment with a budget of 2 million zloty, migration costs 1.7 million zloty plus 4 to 8 months of downtime.
Mitigation strategies
Three approaches adopted by experienced CTOs. First, an abstraction layer in the architecture: a layer in which the specific LLM vendor is a configuration variable rather than hardcoded. Second, model routing: different queries directed to different models (cost-sensitive to the cheaper one, accuracy-sensitive to the more expensive one). Third, a local LLM as a fallback: Llama, Mistral or another open-weights model as a safety net in case of a change in the vendor's pricing or policy.
Gartner Cloud AI Services Magic Quadrant 2026 notes that 67 percent of mature enterprise deployments use at least two LLM vendors in parallel, precisely for this reason.
The Polish context
The Polish enterprise scene in 2026 is 85 percent single-vendor (most often Azure OpenAI Service). Migration to multi-vendor would be capital-intensive, but single-vendor is becoming a strategic risk, especially given Microsoft's dynamic pricing changes in Q1 2026.
We decide whether multi-vendor makes sense for your company as part of the AI Readiness Audit.