How to choose a design agency for an AI product: a working checklist
Hiring a design agency is always a bet. Hiring one for an AI product is a bet with worse odds, because most agencies have not shipped AI interfaces and the difference does not show up in portfolios. Every deck looks confident. This checklist is how you find the real signal before you sign. ## The five questions **1. "Show me a shipped AI product where model output affects user decisions."** Not a concept. Not a redesign of a chatbot skin. A live product where a wrong model answer has a cost. Agencies with this work will name it fast. Agencies without it will pivot to vision. **2. "Show me the state matrix."** Real AI interface work produces a map of states: streaming, empty, slow, wrong, corrected, retried. Ask to see one from the project they just named. This single artifact separates practitioners from decorators better than any portfolio review. **3. "What is your revision speed once models change?"** AI products change under you. A model upgrade shifts what the interface has to handle. Firms that ship revisions in days fit that reality. Firms that present in stakeholder cycles do not. We wrote about why iteration speed decides AI design outcomes in [the generative UI workflow piece](/blog/generative-ui-pre-figma-workflow-2026). **4. "What exactly does engineering receive?"** The design is worth what engineers ship. Structured handoff (tokens, components, HTML, specs) plugs into modern AI-assisted engineering. Flat mocks add a translation step and a round of drift. If