8 September 2026 · Philipp John
AI in the product design process: what stays human
Six steps from understanding to shipping. Where AI speeds up product design with Figma Make, Claude Design and Claude Code, and where people decide.
Use AI for exploring and making, keep people for framing the problem and deciding what ships. In my process the first two steps and the last three are human-led. The middle, where ideas become prototypes, runs as fast as the tools allow. That is why loops take days, not weeks.
This is the process in six steps, with the tools I use for each.
Why keep the start and the end human?
AI is fast at producing options. It is not good at knowing which problem is worth solving. Talking to users, understanding stakeholders and judging what the evidence means need a person who is in the room. The same goes for the end: deciding what ships, and what it costs, is a human call.
So the loop has a human beginning and a human end, and AI works in between.
What are the six steps?
01 Understand (human-led)
Talk to users and stakeholders, on site where possible. AI can cluster the notes. I check every pattern. At Bosch Security I mapped stakeholders before the first interview, which is how field technicians, who had no say in the brief, ended up in the sample.
02 Define (human-led)
Turn findings into insights and testable hypotheses, and pick the riskiest assumption to test first. A hypothesis needs a signal that can prove it wrong. At Bosch Security the riskiest one, capacity as a live traffic light, was tested first because it could have invalidated the whole direction.
03 Explore (AI-accelerated)
Generate many directions fast in Figma Make and Claude Design. Exploring is cheap, choosing is the skill. For an internal AI platform at an insurance group, I benchmarked how leading AI tools handle sources, projects, roles and long tasks before exploring directions. More in how to explore fast with AI.
04 Prototype (AI-accelerated)
Build working prototypes with Claude Code on the real design system, with tokens synced via Figma MCP. The clickable prototype of that platform shows one product in four views. Switch the role, and start cards, signals and locked features change. More in design to code.
05 Validate (human-led)
Users test the working prototype. AI agents check accessibility and design system consistency first, so the sessions can focus on real problems. At Bosch Security, two field technicians saw prototypes against the live build in five-day cycles. They chose between two or three real options instead of describing what they wanted.
06 Ship and learn (human-led)
Hand over code-ready components, measure the signal from step 02 and feed it back into the loop. At NURI, components went to development through Storybook with variants and tokens. After the redesign, registrations went up.
What does AI change?
- Speed. Ideas reach users in days.
- Fidelity. A prototype that works answers more questions than a drawing.
- Cost of exploring. More directions can be tried before one is chosen.
It does not change who is responsible for the result.
Key takeaways
- AI speeds up exploring and making.
- Framing the problem, talking to users and deciding what ships stay with people.
- Working prototypes on the real design system make validation and handover faster.
- Measure the signal you defined at the start, and feed it back.
You can see the process in full on How I work, and the results in the case studies. For the bigger picture, read what has changed in product design.