AI-Augmented UX Workflows
Agents, large language models, and coded prototypes as a fixed part of the design process, from research and audits to working prototypes, with more than twenty years of UX experience as the yardstick for what holds up.
What this actually means
For the ELSTER case study, an agent clicked through the VAT filing flow on its own and documented 21 findings. For the fragrance visualization, another one collected and normalized 349 perfume profiles from public sources. In both cases I set the criteria and rules upfront, and I decide at the end whether a result holds up.
That is what AI looks like in my work today. It takes over whole steps of the work, from researching and reviewing to documenting and building prototypes. That shifts my role: I define what to look for, set the guardrails, and judge what comes out.
The foundation is hands-on training in AI, data, and coding, covering large language models, Python, API integration, and data pipelines, combined with two decades of UX strategy. The yardstick is the same as without AI. A result has to fit the specific idea, not the average of every similar product.
When AI-augmented workflows fit
This way of working is most useful when:
- A team is already vibe-coding or prototyping with AI tools but the results feel close but not quite right
- Research transcripts and notes are piling up and synthesis is slow
- Multiple variants of a design need to be explored quickly
- Routine documentation (component docs, accessibility specs, handover notes) is eating time
- A founder or PM wants to move fast but needs someone who can tell the difference between fast and good
Less useful when the engagement is short enough that the setup overhead is not justified, or when the product prohibits AI use for compliance reasons. In sensitive contexts, a locally-run LLM is an option that keeps data inside the client's network.
Where AI is used, concretely
Research with agents: An agent searches forums, databases, and manufacturer sites against fixed criteria and stores the results in a structured way. That is how the pain points for the washing machine concept came together before the first sketch.
Audits with agents: A custom skill checks an interface against usability heuristics, ISO 9241-110, and WCAG while the agent navigates the flow on its own. The output is a structured draft report, which I weight, correct, and extend with my own findings.
Research synthesis: Interview transcripts and notes are clustered into patterns and themes. That saves days of coding work, and interpreting the patterns remains my job.
Prototypes in code: When Figma is not enough, I build working prototypes in HTML, CSS, and JavaScript, with real data and real behavior. The washing machine’s reasoning explorer and the fragrance visualization were built this way.
From design to code: Connecting Figma and Claude Code turns components and design tokens directly into code, without detours through screenshots and long handover documents.
Documentation: Component descriptions, accessibility specifications, and handover notes are written alongside the design. AI drafts them, I finish them.
What you get
AI-augmented workflows are not a standalone deliverable. They are part of how I work in other projects. In practice that means:
- Design projects in a fraction of the time a purely manual approach would take
- More options explored before a direction is set
- Prototypes that behave like real software and can be tested with users
- Documentation that arrives with the design, not weeks later
- Grounded knowledge of where AI helps and where it misleads, available for the team’s AI strategy
If a team wants to learn how to use AI in its own design process, that can be arranged as a workshop or mentoring, see mentoring.