How AI changed the way a design team works

Stakes

AI was already entering design workflows one person at a time. The risk was speed without a map: tools without a quality bar, individual prompting mistaken for team capability, and experiments that stop the moment one enthusiast leaves. With a stretched team, the need was to find where AI really helped, make it repeatable, and keep people accountable for what shipped.

My role

What changed in the process

The review tools were one part. The larger change was in how work moved through the team, stage by stage.

• Brief Before, a PRD went straight to a designer and the gaps surfaced in Figma. Now every PRD is reviewed first, and design starts from owned questions.
• Explore Before, a concept started as static screens. Now designers build a working prototype from a prompt and the design system before detailed design.
• Decide Before, a UX assessment was a manual pass through a journey. Now scripts and prompts find and document issues against one framework, and a designer rates them.
• Refine Before, product copy waited on one UX writer. Now a first draft comes from an assistant set up with our writing guide, and the writer edits.
• Ship & learn Before, customer feedback and test results arrived as occasional reports. Now synthesis is a routine that feeds prioritisation.

Decision log

1. 01
   Map the process before choosing tools

   Before buying into "AI everywhere", we mapped five stages — brief, explore, decide, refine, ship — and for each one, what AI could support and where a person had to decide. That map became the reference for every experiment.

Where AI sits in the process

• 01 · Brief — Frame the problem

  AI support: Research and Voice of Customer synthesis · Initial pattern finding

  Human checkpoint: Problem framing · Context · Priorities

• 02 · Explore — Expand the option space

  AI support: Prompt-to-layout · Concept exploration · Prototype support

  Human checkpoint: Relevance · Product judgement · Direction

• 03 · Decide — Evaluate before committing

  AI support: Heuristic review · Pattern comparison · Design-system checks

  Human checkpoint: Trade-offs · Prioritisation · Final decision

• 04 · Refine — Strengthen the solution

  AI support: UX writing · Review assistance · Consistency checks

  Human checkpoint: Craft · Accessibility · Quality

• 05 · Ship & Learn — Connect delivery back to evidence

  AI support: Session analysis · Voice of Customer monitoring · Post-release synthesis

  Human checkpoint: Interpretation · Accountability · Next action

2. 02
   Evaluate the tools before asking for budget

   With my manager, I assessed 26 AI tools across three areas: visual and interaction design, research and concept validation, and UX writing. We recommended a short list, a four-phase rollout from training to standard practice, and a learning plan for the whole team.

   Part of that was approved. The rest of the work ran on tools the team already had, which kept the focus on changing how we worked instead of on collecting licences.

3. 03
   Review every PRD before design starts

   I built a PRD review prompt that runs eight passes on every requirements document, used on every PRD from January 2026: customer journey, feasibility, irreversible decisions, contradictions, open questions ranked by risk, the arithmetic, post-launch measurement and ownership. The output is fixed: one blocker, no more than eight owned questions, an edge-case matrix and a realistic UX scope. Design started from a reviewed brief instead of discovering the gaps in Figma.

   

4. 04
   Check designs against the quality bar

   I set up a review workflow using Claude with the Figma MCP integration, and Cursor for design-quality review, checking work against the UX Quality Bar and design-system rules: missing states, contrast, component misuse, unclear labels. The tools flag; the designer decides.

   

5. 05
   Make customer feedback a routine, not a project

   Voice of Customer used to arrive as occasional reports. I built a recurring app-store review analysis for PaysafeCard across Android and iOS, with a fixed spec, so themes and shifts fed straight into prioritisation and design reviews. The team used the same approach to synthesise usability tests and survey responses, so findings reached a design review in the same week.

6. 06
   Prototype before detailed design

   Designers started building working prototypes with Figma Make, using our Global Design System, at the point where they used to draw static screens. A product manager could click through an idea and react to it before anyone invested in detailed design.

7. 07
   Draft product copy from the writing guide

   The team had one UX writer, and copy was a bottleneck. We set up an assistant with our UX writing guide, glossary and tone rules, so designers got a first draft in the brand voice. The writer moved from writing every line to editing and owning the standard.

8. 08
   Run UX assessments with scripts and prompts

   For the 2026 UX assessments of registration, KYC and P2P, we used Claude with scripts and prompts to walk the recorded flows, find issues and write them up in the standard format. Designers checked each finding and set the severity (UX Assessment Programme →).

9. 09
   Invest in people, not only tools

   I funded AI-for-designers training for one designer, who then ran internal sessions for the team, and proposed a learning plan covering design, research and writing for everyone. Tools change every few months; a team that knows how to test them keeps its advantage.

Operating principle

Speed and quality move together.

A faster step that lowers quality is a failed experiment. AI accelerates; people evaluate intent, fit and the final decision.

Impact

• Every PRDFrom January 2026, reviewed with the 8-pass prompt before design started
• 8 passesJourney, feasibility, irreversibility, contradictions, ranked questions, arithmetic, measurement, ownership
• 1 blocker · ≤8 questionsFixed output, so reviews were comparable across PMs and products
• RecurringVoice of Customer synthesis for PaysafeCard, Android + iOS
• 5 stagesBrief, explore, decide, refine and ship each changed, with a named human checkpoint
• 26 toolsAssessed across design, research and UX writing before a short list was proposed

Reflection

Models will keep changing. What lasts is whether the team gets more capable and stays accountable. I want AI to take the busywork and give designers more time for judgement, exploration and hard product thinking. In financial products there's always a person behind a decision about someone's money, and that responsibility should never become unclear.

What I would do differently is measure it. We changed how the work flowed, but we did not record time saved or rework avoided, so I can describe the change and not yet put a number on it.