A Guide to Choosing Your AI Engagement Model

AI projects fail in the contract stage more often than the code stage — usually because the engagement shape was wrong for the uncertainty involved. There are three honest shapes, and the right one depends on one question: how much do you already know about what you need?

Prototype-first fits high uncertainty. You have a problem and a hypothesis, not a specification. Buy days, not quarters: a working slice, tried by real users, that earns (or kills) the next phase. The deliverable is evidence. Generative AI has collapsed the cost of this shape — refusing to prototype first is now a choice, not a constraint.

Platform fits known problems with known workflows — a builder's growth stack, an education platform. Here you should demand the opposite of a prototype: canonical requirements, formal gap analysis when they change, verified modules, a launch checklist. The deliverable is a system your team operates.

Embedded fits organisations building AI capability in-house. The partner's job is to make themselves unnecessary: pairing, evaluation harnesses, guardrails, and the unglamorous documentation that survives staff turnover. The deliverable is your team, upgraded.

Whichever shape you choose, insist on two clauses: demo modes that work without production keys (so you can evaluate before you integrate), and human approval points wherever AI output leaves the building. Any partner who resists either is selling you their roadmap, not yours.