AI Vendor Lock-In: An Exit Plan for Your AI Factory

Paweł Szczepanik
Paweł Szczepanik
September 25, 2026
•
9 min read
Loading the Elevenlabs Text to Speech AudioNative Player...

People ask how to avoid AI vendor lock-in. You cannot remove it; you can price it and rehearse the exit. Cloud shows how sticky a first choice gets: fewer than 1% of UK cloud customers switch provider in a year, according to the July 2025 final decision of the UK Competition and Markets Authority (CMA). And your model vendor will make you move anyway. At the time of writing, OpenAI gives at least 6 months of notice before retiring a generally available model, and Anthropic at least 60 days for publicly released models.

A model change is routine, and leaving a vendor is the same operation at a larger scale. AI vendor lock-in is the exit you have never priced.

What AI Vendor Lock-In Actually Means

AI vendor lock-in is a dependency on one provider of models, AI services or cloud infrastructure that makes leaving slower or more expensive than the business can accept. It is measured by the cost and time of an exit, not by the number of vendors you use.

The European Central Bank's Guide on outsourcing cloud services to cloud service providers (July 2025), a non-binding guide for the banks it supervises, defines ICT vendor lock-in as a dependency "that endangers the ability of the institution to effectively exit from some or all the services provided." The risk comes from an exit whose price and duration nobody has estimated.

This article is about pricing and planning the exit from an AI vendor: where an AI factory gets locked in, what leaving would cost and how long it would take, which assets you must keep to make a switch possible, and what the contract and the EU Data Act give you. For how the layers are drawn so one component can change without touching the rest, see our reference architecture guide.

Why AI Lock-In Sticks Harder Than Cloud Lock-In

Cloud already showed what switching costs look like. The CMA names egress fees as "A key commercial barrier" and traces technical barriers to "the differentiation of features and interfaces in cloud services." Those findings cover UK cloud infrastructure, not model providers, but every AI factory inherits them, because models, vectors and logs sit on cloud services.

AI adds something the cloud never had to move: behavior.

  • Prompt format is model-specific. In a 2023 study by Sclar and colleagues, formatting alone moved accuracy by up to 76 points on one model, LLaMA-2-13B, and "format performance only weakly correlates between models."
  • Behavior shifts under a fixed name. Chen, Zaharia and Zou found that GPT-4 told prime from composite numbers with 84% accuracy in March 2023 and 51% in June 2023. A model name is not a behavior contract.
  • Providers admit it. One writes of its own compatibility layer: "If you've done lots of tweaking to your prompt, it's likely to be well-tuned to OpenAI specifically."

Vectors from one embedding model cannot be searched with another's, so a new embedding model means re-embedding the corpus and rebuilding the index. In cloud, you move data and code. In AI, you also move behavior, and behavior does not export. That is why LLM vendor lock-in costs more than the API integration suggests, and why AI vendor lock-in tends to surface only when a switch is forced.

Seven Places an AI Factory Gets Locked In

An AI vendor lock-in review can walk this list of dependencies row by row.

Lock-in pointWhat makes it stickyWhat drives the exit costThe hedge that pays
Model API and SDKProvider-only features wired into application codeCall sites and features with no equivalent elsewhereOne internal client interface; provider-only features behind flags
PromptsTuned to one model's behavior; small format changes move resultsNumber of prompts times the effort to re-tune and re-test eachPrompts versioned in your repository, next to the eval set
Evaluation setsMissing, or stored inside a vendor toolWithout them there is no evidence the replacement worksEval sets and scoring scripts in your repository, runnable against any model
Embeddings and vector indexVectors are comparable only within one embedding modelCorpus size, re-embedding time and index rebuildSource documents and a reproducible embedding pipeline, model version recorded
Fine-tuned modelsTuned weights may stay on the provider's platformRetraining on a new base model and re-evaluatingTraining data and eval results kept by you; export terms checked before tuning
Orchestration and agent runtimeWorkflow definitions in a vendor-native formatNumber of workflows and tool integrations to rebuildBusiness logic in code you own; the runtime treated as a replaceable host
Data, logs and feedbackStored in the provider's format or region; egress chargesData volume, egress price and export toolingData layer in your own cloud account, exports in open formats

Our map, not a standard. For each row, write down how long a switch would take and what it would cost. A row without those two numbers is where the lock-in is.

Own the Assets That Make a Switch Possible, Price the Rest

Our recommendation: own four things before any switch.

  • An evaluation set and scoring script in your repository, runnable against any model. The eval set turns a switch from a guess into a test.
  • Prompts in your repository with version history, next to the eval set, not in a vendor console.
  • Source documents and a reproducible chunking and embedding pipeline with the model version recorded, so re-embedding is a job you schedule, not a project you staff.
  • Data, logs and user feedback in your own cloud account, in an open format such as Apache Parquet. Where state lives decides who controls the exit, which is why we recommend a data platform built in your own cloud account, with the code handed over.

The law helps only with that last item: the Commission's Data Act explainer covers "input and output data, including metadata, generated by the customer's use of the service," not the provider's model.

Interface portability is cheap. vLLM, a server for open-weight models, "Exposes an OpenAI-compatible HTTP API," and Anthropic offers an OpenAI SDK compatibility layer as well. A gateway belongs to the same cheap part of the exit. Behavior portability is expensive: Anthropic calls its own layer "not considered a long-term or production-ready solution for most use cases."

Our rule: accept a dependency on a vendor feature when three conditions hold. The feature, such as prompt caching or a managed index, brings measurable value. You know the exit time and cost for its row. And that exit time is shorter than the vendor's notice period. On the four assets above, refuse it.

A second provider in production doubles evaluation and maintenance; a second model tested on your eval set is enough. You cannot avoid AI vendor lock-in everywhere, and you should not try. Refuse it on the assets you own and price it on the features you rent.

Write the Exit Plan Before Go-Live, Then Rehearse It

The vendor sets the clock. At the time of writing, OpenAI promises "At least 6 months" for generally available models and "At least 3 months" for specialized variants, while previews "may be retired with much shorter notice, such as 2 weeks," unless safety or compliance require faster action. Anthropic gives "at least 60 days' notice" for publicly released models on platforms it operates. Check the current pages before planning against them.

If moving to another model takes longer than the notice period, the AI vendor lock-in is already there. The bill just has not arrived.

For the plan, borrow the ECB method; it binds only supervised banks. It wants exit plans with roles and estimated costs before the systems go live, covering "the critical milestones, a description of the tasks and skillsets that are required to perform the exit, and a rough estimate of the time required, and the costs involved." It also asks for tests of the most critical migration steps and review "by someone who is not responsible for drafting the plan."

Our version for an AI factory is an exit card, one page per use case: the fallback model and its eval score, time to re-embed the corpus, how data and logs get exported, the contractual notice period, and an owner. Then run an exit drill, quarterly in our practice, running the eval set against the fallback model and recording the score and switch time.

Contracts and the EU Data Act: What Switching Rights Cover

Chapter VI of the EU Data Act, Regulation (EU) 2023/2854, covers switching between providers of data processing services such as cloud and edge. It has applied since September 12, 2025, and requires that customers can switch "quickly and smoothly, and without losing any data or the functionality of applications." Since January 11, 2024, providers may charge only their actual switching costs, and from January 12, 2027 switching charges, including data egress, disappear.

Check the scope before relying on it for AI vendor lock-in. If your AI runs on a data processing service in the EU, the Data Act gives you switching rights; whether a specific model API counts is a question for your lawyer.

The Commission has also published non-binding standard contractual clauses for cloud contracts. Three of them turn Chapter VI into ready terms: SCC Switching & Exit, SCC Termination, and SCC Security & Business continuity. For the model provider contract, ask:

  • Is the retirement notice period longer than your measured exit time?
  • Can you export fine-tuning data, logs and evaluation results?
  • Can tuned weights leave the platform?
  • What egress charges apply before January 12, 2027?

At DS Stream we build on the cloud the client already pays for, and code, prompts, eval sets and data stay in the client's repositories and account. Every use case ships with an exit card, as part of an AI factory built on your cloud, with your code. For a first number, ask us to price the exit for one use case.

FAQ

What is AI vendor lock-in?

AI vendor lock-in is a dependency on one AI provider that makes leaving slower or more expensive than the business can accept, measured by the time and cost of an exit. In AI, prompts, eval sets and embeddings stick to one model too.

Does an OpenAI-compatible API prevent lock-in?

Only partly. It changes the call in a few lines, but behavior does not come along. Anthropic calls its own compatibility layer a tool for testing and comparing models. The hedge against LLM vendor lock-in is an eval set that proves the replacement works.

How much notice do AI providers give before retiring a model?

At the time of writing, OpenAI gives at least 6 months for generally available models, 3 months for specialized variants and as little as 2 weeks for previews. Anthropic gives at least 60 days for publicly released models.

How do you test an AI exit plan?

Run an exit drill: execute your eval set against the fallback model and time the switch and the re-embedding. Then have someone who did not write the plan review it, as the ECB guide asks of banks.

Share this post
Artificial Intelligence
Paweł Szczepanik
MORE POSTS BY THIS AUTHOR
Paweł Szczepanik

Curious how we can support your business?

TALK TO US

More insights

More news

No items found.