Model Deprecation Clauses: The AI Contract Risk Nobody's Reading For
Traditional SaaS assumes the product you bought still exists at the end of the term. An underlying model can be retired mid-term, sometimes on weeks of notice.
Traditional SaaS contracts have a fairly stable core assumption: the product you're buying today will basically still exist, in roughly the same form, for the length of your contract. AI vendor contracts complicate that assumption in a specific way. The underlying model your workflow depends on can be deprecated, replaced, or significantly changed during your contract term, sometimes with only weeks of notice.
Why this risk is genuinely different from typical SaaS
A traditional SaaS vendor might change a feature or a UI, but the core functionality you signed up for generally persists. AI model providers, by contrast, regularly retire older models in favour of newer ones, sometimes with materially different behaviour, cost structure, or output quality. If your workflow or integration was built around a specific model's particular behaviour, a deprecation can mean real rework, not just an adjustment period.
Your term, and the model lifecycle running underneath it
Notice period, sometimes weeks
Traditional SaaS review asks what happens if the price changes. This asks what happens if the product does.
What to check in AI vendor contracts specifically
- Whether the contract or terms of service specify a minimum notice period before a model your usage depends on is deprecated
- Whether deprecated models are typically replaced with a directly comparable option, or whether migration requires meaningful rework on your side
- Whether pricing is guaranteed to stay stable through a model transition, or whether a forced migration can also mean a pricing change
- The historical pattern, if available, of how much notice this vendor has given for past model deprecations
Why this deserves specific attention in negotiation
Most standard SaaS contract review doesn't include a question like "what happens if the underlying technology changes mid-term," because for traditional software, that scenario rarely comes up. For AI vendor contracts, it's a real and recurring possibility, worth explicitly asking about rather than assuming standard contract review already covers it.
What to do if your workflow depends heavily on model-specific behaviour
Build in some resilience deliberately. Avoid over-optimising a workflow around quirks of one specific model version if you can help it, and ask the vendor directly what migration support, if any, comes with a forced deprecation.