Sources checked
Inventory the features a replacement must reproduce
Inventory the components your application depends on. Keep prompts, evaluation cases, and source data in portable formats where practical.
Write down provider-specific features that a replacement must reproduce or that you can remove.
API compatibility does not ensure identical behavior
A compatibility layer can hide differences without eliminating them. Test output format, error handling, latency, and tool behavior against the replacement before switching traffic.
What you would have to rebuild
Prompts that depend on one vendor’s tool schema, a fine-tune that exists only as their model id, and an index built with their embedding model are not portable. A chat export is not the system.
The system is the glue that assumes their error codes, their batch API, and their way of truncating a long prompt.
Some of that glue is ordinary software and can be wrapped. Some of it is weights you do not have.
If the contract ends and you cannot download the checkpoint, you cannot reproduce behavior. Write that down as a risk even if you are happy with the vendor today.
How to leave a door open
Keep your eval set in your own repo. It is the only way to tell whether a second vendor is close.
Prefer standard file formats for documents you store. Isolate vendor SDKs behind a small interface so a second implementation is a file, not a rewrite of the product.
You do not have to be multi-cloud on day one. You do have to know which piece would take months to move.
A purchasing decision that ignores that piece is incomplete.
What to record before you rely on it
List the artifacts you do not control: model ids, fine-tunes, embedding spaces, and tool schemas. Mark which ones have an export.
Revisit the list at renewal. Keep the model id or the paper citation, the date you read it, and the inputs you used beside that note, so a later change can be compared with the same test.
When the result is a failure, store the artifact and change the design. Do not relax the check until the model happens to pass.
Name a person who owns that failure.
Common questions
Is using an open-weight model the same as avoiding lock-in?
Only if you can actually run and version that file, and your product does not depend on a host’s private tools. A weight file you have never deployed is not an exit.
Should we abstract every call on the first day?
Abstract the boundary you are likely to cross. A premature wrapper of every parameter can hide the very differences your eval needs to see.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.
Our editorial standards





