The product-design operating model I architected and put into practice.

I designed and implemented this model to integrate AI across research, design systems, prototyping, delivery, quality assurance and performance measurement. I am its architect and advocate; I also enabled the team to use it through structured onboarding, office hours, practical training materials and support on live work.
On this page
From isolated AI tasks to a system teams can actually use.
I created the model to connect six stages that were often treated as separate experiments: agentic research, rapid design, design-to-story and code, scalable content generation, automated quality assurance, and performance insight fed back into the next cycle.
The model is already in use and continues to evolve; the artefacts below show the implemented research workflow, design-system grounding, generated prototypes, custom tooling, code-ready output and controlled templates behind it.
I led adoption alongside implementation. I onboarded the team, ran office hours to resolve questions on live projects and created training materials that translated the model into repeatable steps, examples and review points.
Accountable ownership
I position AI as an accelerator; researchers, designers and product leaders still own the decisions.
Controlled context
I ground each workflow in approved research, brand guidance, components, tokens and product data.
Traceable output
I keep assumptions, findings, specifications, test results and KPI signals visible and reviewable.
Governed adoption
I make privacy, accessibility, security, brand and code review explicit gates in the workflow.
I made adoption part of the implementation.
I introduced the model through active delivery, rather than handing over a process diagram. The support structure gave the team a place to learn the workflow, apply it to current work and improve it through use.
Introduced the six stages, roles, source material, outputs and review responsibilities.
Provided regular access for live questions, workflow coaching and practical problem solving.
Created reusable guidance, worked examples and guardrails that teams could follow independently.
A working research workflow with researcher judgement built in.
I have integrated agentic support into the research workflow for surveys, discussion guides, recruitment operations, scheduling, transcription and first-pass thematic analysis. The researcher still interrogates the evidence, separates severity from confidence and decides what must be validated next.
- Draft and refine research instruments
- Support participant operations and session records
- Cluster evidence and surface themes for review
- Turn findings into prioritised, testable hypotheses


I made the design system the starting point for generation.
I configured the Figma Make workflow around approved brand libraries, tokens and existing components. That lets teams produce responsive variations and working prototypes quickly without inventing a parallel visual language. New patterns only enter the shared system after review.
- Explore states, flows and breakpoints rapidly
- Reuse tokens and components before creating new patterns
- Prototype interactions with realistic content and behaviour
- Review accessibility, brand and system fit before adoption



I use the model to remove specific workflow bottlenecks.
When I identified motion specification as a recurring hand-off gap, I used an agentic development workflow to create a working Figma plugin. Prompting accelerated the build; I retained code review, permission control, documentation and QA. The result converts motion decisions into reusable annotations and exportable behaviour.




I designed the hand-off as structured delivery output.
I structured the Figma source so it can generate first-pass Azure DevOps stories, acceptance criteria and specifications while exposing code-ready tokens, semantics and layout rules. Product, design and engineering review the generated output together before it enters delivery.
- Translate named components and states into story structure
- Carry behaviour, accessibility and edge cases into acceptance criteria
- Expose tokens and reusable implementation patterns
- Keep a reviewable line between design source and shipped code


I turned approved templates into controlled generation tools.
I introduced controlled Figma Make and Firefly workflows around approved templates. Locked structures, defined editable fields and curated media let teams produce page, content and asset variations without rebuilding the composition or drifting away from the brand system.
- Generate layouts from approved page templates
- Swap structured content without rebuilding the composition
- Create on-brand image and asset variations for review
- Preserve provenance, rights and human approval before release

I connected quality assurance back to the approved design source.
The model uses approved Figma output as the reference for automated regression and visual QA. It flags layout, component, content and accessibility differences for review; the team then records whether each difference is a build defect, an accepted variation or a required update to the design source.
Approved Figma output
Components, tokens, states, breakpoints and expected behaviour.
Build & regression run
Visual diffs, interaction checks and accessibility signals across key journeys.
Human triage
Fix the build, accept the variation or update the source of truth.
I designed measurement as the input to the next cycle.
I use AI-assisted analysis across Copilot, Userlytics and customer-data platforms to connect behavioural signals, qualitative feedback and business outcomes. The team reviews the patterns, tests attribution and converts prioritised opportunities into the next research questions, experiments and design-system updates.
Behaviour & journeys
Completion, conversion, abandonment, pathing, errors and performance.
Feedback & meaning
Usability evidence, support themes, sentiment and emerging needs.
Opportunity & confidence
Combine scale, severity, value and evidence quality before acting.
Feed the next cycle
Turn live evidence into new questions, experiments and system updates.


