Technology buyers evaluating an AI product typically start with capability: what can it do, how well, compared to what alternative. A question that matters just as much and gets asked far less often is what happens to the product when the underlying model it's built on changes - because for most AI products, that underlying model is not fully within the vendor's control.
Why this question is easy to skip
A capability demo is concrete and immediate; a question about future model changes is abstract and speculative, which makes it easy to defer past the point where it should have been asked. But foundation models get updated, deprecated, and replaced on a timeline set by their providers, not by the application vendors built on top of them, and those changes can materially affect a product's behavior overnight.
What can actually go wrong
A model update can change response style, degrade a fine-tuned behavior the vendor spent months tuning, or in the case of a deprecated model, disappear entirely on a forced migration deadline. A buyer who's already built workflows and trained staff around specific product behavior can find that behavior shifting for reasons entirely outside the vendor relationship they thought they had.
Questions worth asking during evaluation
Which specific model or models does this product depend on, and does the vendor control the update cadence? What's the vendor's track record when their underlying model provider has made a breaking change? Is there a contractual commitment to notify customers before a dependency change that could affect their workflow?
Why this belongs earlier in the evaluation
A product's current capability tells you about today. Its dependency structure tells you about how stable that capability is likely to remain. Buyers making a multi-year commitment are underwriting more of the second question than the first, even though most evaluations spend nearly all their time on capability alone.