Most design systems have no answer for streaming text, confidence states, agent actions, or citations.

So teams building AI products improvise. Every surface handles generation slightly differently, each engineer invents their own loading state, and the product starts to feel unreliable even when the model underneath is good. Inconsistency reads as instability, and users do not separate the two.

Four weeks to build a system that covers the AI patterns, documented well enough that your team uses it instead of working around it.

Week 1. Audit. What already exists, where it is inconsistent, and which AI patterns your roadmap is going to demand in the next two quarters.

Weeks 2 to 3. Build. Foundations and a component library in Figma with proper variants and states, built for real product density rather than a showcase page. Alongside it, the AI specific layer: streaming and loading, confidence and provenance, agent actions, corrections, and error recovery.

Week 4. Document and hand off. Usage guidance, contribution rules, and handoff into Storybook or zeroheight with your engineers in the room.

A normal component has known states. An AI component has uncertain ones: partial output, low confidence, a revised answer, an action already taken on the user's behalf. Those need designed treatments and explicit rules, and a system that does not provide them pushes the decision onto whoever is closest to the deadline.

AI product teams whose UI is drifting as they ship, and teams standing up a system for the first time who would rather not discover the AI gaps six months in.