Planning how to leave an AI vendor before you choose one

Vendors change their products even when you stay

An exit plan is also a migration plan for changes you did not choose. On 3 June 2026, OpenAI announced that its hosted reusable prompt objects, its Evals platform and Agent Builder were deprecated, with shutdown scheduled for 30 November 2026, and advised moving prompt content into application code. Model retirements follow published notice periods: OpenAI states minimum notice of six months for generally available models and three months for specialised variants of them, possibly much shorter notice for preview models, and faster retirement where safety or compliance requires it. Anthropic states at least 60 days for publicly released models. Teams whose prompts and tests lived only in a vendor dashboard had migration work to do whether or not they wanted to leave.

What to keep in your own hands

• The test set: cases, expected outcomes, grading logic and past results, stored where any system can be run against them.
• Prompts, instructions and tool definitions, versioned in your repository rather than only in a vendor console.
• The business rules behind thresholds, routing and escalation, written down outside the vendor's interface.
• Your records: inputs sent, outputs received, corrections and review decisions, in a format you can read without the vendor.
• Source documents, indexing settings and the name of the embedding model. Existing vectors can move to another store if the embedding model stays the same and the new store accepts them. A new embedding model means embedding the documents again.
• Fine-tuned models: check whether you receive the weights or adapters or can use them only on the vendor's platform.

A dependency inventory for a hypothetical assistant

The AWS Generative AI Lens recommends a model catalogue that records lineage and dependencies, and NIST's framework asks for contingency processes for failures in third-party data or AI systems deemed high-risk. A practical version is one line per dependency with where it is configured, how it would be exported and how hard it would be to replace. For an internal support assistant, it might look like this.

• Language model: model ID set in your configuration repository. Replacement effort medium, because the test set has to pass on the new model.
• Prompts: stored in the vendor's console. Export is a manual copy, so the effort stays high until they move into your repository.
• Evaluations: hosted by the vendor, exportable as files. Effort high, because graders would have to be rebuilt elsewhere.
• Vector index: a managed database from a second vendor. If the store allows export and the embedding model stays the same, the vectors can be loaded into another store. Changing the embedding model means rebuilding from the source documents. Effort medium.
• Traces and logs: an observability service with 30 days of retention and an export API. Effort medium, and older history is lost unless it is exported regularly.

The two high-effort lines are the ones to fix first, and both are fixed by moving material into systems you control.

Questions to settle in the contract

• In what format and within what time can you export your data, prompts, configurations and evaluation results?
• How much notice is given before a model, feature or endpoint you use is retired, and before prices change? The FinOps Foundation notes that AI prices can move sharply in either direction.
• Who holds fine-tuned artefacts, and what happens to your data and logs after termination, including written confirmation of deletion?

The legal wording belongs to procurement and counsel. Engineering's part is to state what must be exportable and to test that an export actually works before it is needed.

Design choices that keep switching affordable

Put a thin internal interface between your process code and the model provider, and keep the review interface and business rules in your own systems. The AWS lens suggests offering several model options through routing and a catalogue so that new models can be onboarded quickly. Keep the layer thin. Anthropic's 2024 article on agents warns that heavy framework layers can hide prompts and responses, and an abstraction built to cover every provider tends to give up the features you chose a provider for. Once a year, run the test set against an alternative and record how long the switch would take.

Buying an AI product or building around your systemsWho owns an AI system after the pilot