Agentic AI in design workflows

Five participants took part in semi-structured interviews about their experiences with agentic AI tools in product design. Interviews lasted approximately 16 to 31 minutes and were conducted remotely. Participants included students, early-career designers and professional product designers. Their experience ranged from limited startup design experience to several years of product design and engineering work. To protect anonymity, participants are referred to as P1 to P5.

Participants were recruited through convenience sampling from my academic and professional network. All participants gave consent for recording and transcription. They were informed that there were no right or wrong answers, that they could skip questions, and that they could stop the interview at any point.

The interview protocol covered four areas: participants’ background and design experience, how they first encountered agentic AI tools, a recent project or workflow involving these tools, and perceived benefits, barriers, privacy concerns and future role changes.

Questions focused on concrete examples rather than only abstract opinions. For example, participants were asked, “Can you walk me through a recent project where you used these tools?”

Analysis used Braun and Clarke’s reflexive thematic analysis. I first read all transcripts to become familiar with the data, then created initial codes from participants’ own language and workflow descriptions.

I then grouped related codes into candidate themes and checked whether each theme captured a meaningful pattern across participants. The final themes were not treated as objective facts emerging from the data, but as researcher interpretations shaped by the research question, interview context and my own position as an HCI student studying human-AI interaction.

The interviews focused on professional and educational experiences rather than sensitive personal issues. Transcripts were anonymised and participant names were removed.

As an MSc HCI student interested in AI design tools, I may have been more attentive to productivity gains. To balance this, I actively coded negative cases, including tool abandonment, privacy concerns, token costs, design-system problems and frustration with prompt-based editing. My peer role may also have shaped the interviews, as participants sometimes assumed shared knowledge of tools such as Figma, Claude or MCP.

Planning and early design work often started before participants opened Figma. Instead of beginning from a blank Figma canvas, several participants described starting with Claude, Gemini, paper sketches, a PRD, or even a code repository.

P2 described Figma as something they use after the initial thinking has already been structured, relying instead on paper and Claude to formalise and structure their thoughts. P5 described a similar AI-first workflow, starting with a PRD, generating a prototype from it, and using that prototype for team discussion.

This shows that agentic AI is reorganising the order of design activity. Early design becomes conversational and code-adjacent, while Figma becomes a later-stage space for polish, correction and design-system alignment. This reflects Suchman’s argument that human action is shaped by available artefacts and environments.

Participants valued agentic AI mainly for speeding up exploration, prototyping and repetitive checking. However, they did not describe AI as replacing their own design judgement. Instead, they used AI to generate options and then positioned themselves as reviewers, editors and decision-makers.

P5 explained that AI had compressed early product work: what used to take a week was now “just an hour’s job”. However, the participant did not treat this as full automation. They still brought the work back into Figma because “I want clean design, I want version history”.

P4 made this boundary especially clear when discussing what should remain human-led. They used agents for more deterministic tasks, such as design audits and PR reviews, but for work requiring judgement the designer still had to direct the tool.

This challenges simple claims that AI will “do design”. In practice, participants used agentic AI as a generator of possibilities, not as a final authority. This also aligns with Schön’s view of design as reflective practice, where practitioners make judgements through cycles of seeing, framing and reframing.

Participants repeatedly described agentic AI as useful because it connects design and code. For some, this reduced the traditional handoff gap between designers and engineers. P4 explained that previously their role ended at final Figma screens and handoff.

P5 also described a workflow where Claude, Figma MCP and GitHub were connected, taking the repository of their front end, giving prompts, connecting Figma through MCP and pushing the results back to GitHub. This shows how agentic AI tools can create a two-way workflow between design files and implementation environments.

However, this connection also created a loss of direct control. Participants often wanted to edit AI-generated designs manually rather than communicate every change through prompts. P1 liked Claude Design, but editing was frustrating, and every suggested edit took up a lot of tokens.

P2 described this as cognitive load. The reason was not just speed, but the effort of explaining.

P3 raised a related issue around design-system adherence. AI prototypes were useful at a high level, but “once we start getting into the nitty details of design system, then it struggles a lot”. P3 said they sometimes created “30, 40 iterations to arrive at something which I liked”, and that this “might have taken me a similar time if I had created all of this manually”.

Speed at the macro level can become slowness at the micro level if small visual corrections require repeated prompting. Norman argues that good design should make actions visible, understandable and controllable; participants’ frustrations suggest that many agentic design tools still lack this kind of direct controllability.

Participants did not see agentic AI as only a tool-level change. They also described it as changing expectations around product roles. P1 suggested that design roles may become more generalised, with UI, UX and UI engineering combining into one field, because with AI the coding becomes accessible.

This role expansion was already visible in workplaces. P3 explained that in their company, product managers were being pushed to explore AI design tools, and had run internal demos to support that change.

P5 described a broader organisational expectation, that everybody in the team, technical and non-technical, has to start with AI tools. P2 connected role-blurring to organisational expectations, arguing that companies increasingly want people who can own an entire vertical.

Participants framed this role expansion both positively and critically. On one hand, AI enabled designers to build more, test more and contribute closer to production. On the other hand, this can increase pressure on designers to cover more areas while still being responsible for craft and quality. Jarrahi argues that AI can support human decision-making, but also changes the distribution of expertise between humans and machines.

The sample was small and recruited through convenience sampling, so findings cannot be generalised to all designers. Participants also had different levels of experience, from students to professionals, which created useful variation but limited direct comparison. Finally, the study relied on retrospective interviews rather than live observation, so participants may have forgotten details or described idealised workflows.

Future work could use screen-recorded walkthroughs, diary studies or observations of designers using agentic AI in real projects. A follow-up study could also compare designers, PMs and engineers to examine how agentic AI changes collaboration across product teams.