The feature that got un-shipped, and the contract clause it argues for
OpenAI is winding down its fine-tuning platform on a published schedule. The procurement lesson applies to every AI capability you rent.
A quiet case study in AI procurement has been unfolding on OpenAI's deprecations page this year: the self-serve fine-tuning platform is winding down in stages, with new training jobs ending for everyone by January 2027. Models already trained keep running, for now. A capability that enterprises built workflows on is being un-shipped, politely, on the vendor's schedule.
I have no quarrel with the business decision. I have a memo for everyone whose AI contracts assume features are forever.
Traditional software deprecations gave you years and an escape: the old version kept working on your servers. Rented AI capabilities have no your-servers. When the vendor retires the feature, the workflow retires with it, on their timeline, and your only variables are how early you knew and what your contract says about it.
So, three clauses worth the negotiation time on any AI agreement of size. Deprecation notice: a defined minimum (twelve months is gettable) for any capability your workflows depend on, in writing, not in a blog post. Sunset assistance: if the vendor retires a capability, they owe migration support or credits, which converts their roadmap freedom into a priced tradeoff instead of your surprise. And export rights: whatever the capability produced (tuned weights where applicable, datasets, evaluation results) leaves with you, because artifacts you cannot take are rent, not assets.
And one internal habit that costs nothing: keep a register of which workflows depend on which rented AI features, so a deprecations-page update reads as a work order instead of an investigation. The fine-tuning wind-down was announced with dates. The companies scrambling anyway were not short of notice. They were short of a list.