Differentiation (does this workflow win customers?), proprietary data (do you have what others do not?), capability (can you staff and keep it?), total cost (3-year, all-in). Most AI needs are "wrap" — a thin layer over a vendor model — and pretending they are "build" is how budgets die.

Per-seat, per-token, per-outcome, platform fee — each model hides a different trap at scale. Model the 3-year cost at YOUR volume, add the integration and change-management lines the quote omits, and ask the scaling question before signing: what does 10× usage cost?

A PoV is not a demo with your data; it is a bounded experiment: one workflow, a baseline, success criteria the vendor sees BEFORE the demo, a kill criterion, and a decision date. The discipline that turns "impressive" into "purchased" or "politely declined".

Data usage (does the vendor train on yours? can you opt out?), portability (can you leave with your fine-tuned artifacts and logs?), liability (who pays when the model errs?), and SLA honesty (what "availability" means for a probabilistic product). Write the exit plan before signing; the exit you did not plan is the price they did not quote.