Tristan Clarkson Contact

All workCase study 01 of 03

Darwin, part of Marsh · Design Operations Lead · 2022 to present

Should this be AI at all?

AI was arriving in chat, summaries, search and research tools, and every team risked inventing it again. I connected product strategy, experience language, governance, reusable components and AI-enabled ways of working, so each new use did not start from nothing.

Darwin benefits page with a contextual AI policy summary, a verified source and follow-up actions.

01

Role and ownership

Context
AI features were arriving across Darwin's products faster than shared standards for them.
My role
Design Operations Lead, 2022 to present.
What I owned
The suitability route, interaction guidance, accessibility requirements, reusable patterns and the AI-enabled design workflow.
Worked with
Product and engineering colleagues, and risk and compliance partners.
Constraints
WCAG 2.2 AA for conversational and generated experiences, and a route to a person when AI cannot answer.
Outcome and evidence
A framework, patterns and a working method. No adoption, satisfaction or commercial measures are claimed.

What I led

  • A suitability route used before an AI feature is designed, so the choice of model is assessed and recorded.
  • Interaction guidance for uncertainty, escalation, reversibility and thresholds for acting on a person's behalf.
  • Accessibility requirements for conversational and generated experiences, against WCAG 2.2 AA.
  • Reusable AI patterns and states, including citation, caveat and human-handoff behaviour.
  • Coded and AI-assisted prototypes that test latency, streaming, failure, keyboard and screen-reader behaviour.

Alongside

Product and engineering colleagues, and risk and compliance partners. Design-system quality, design-to-code parity and defect learning were shared with product and engineering.

What this case does not claim

Shipped feature results, adoption, satisfaction or commercial outcomes. The screens are design explorations, and the diagrams are representative views of the framework.

The call

Ask whether AI belongs at all before designing any AI interface.

Rather than

Treating each AI feature as its own design problem, and adding guardrails afterwards.

Why

A model brings uncertainty, responsibility and failure modes that a deterministic pattern may avoid. Deciding first reduces risk before it creates capability.

02

AI was arriving everywhere. The risk was that every team invented it again.

Chat, summaries, search, research and prototyping were all useful. Without a common system, they also created inconsistent language, duplicated patterns, uneven confidence handling and unclear ownership.

Chat, summaries, semantic search, UX automation, research synthesis and prototype support passing through fragmentation risk into a Darwin AI framework and better outcomes.
The opportunity was not one feature. It was one coherent system for making responsible AI repeatable.
Read this diagram as text

On the left, fragmented AI activity: chat assistant, AI summaries, semantic search, UX automation, research synthesis and prototype support. Together they create a fragmentation risk: inconsistent language, duplicated patterns, accessibility gaps, unclear ownership, uneven confidence handling and tool-by-tool behaviour.

In the centre, the Darwin AI framework, described as a coherent system for responsible, consistent and scalable AI. It rests on six parts: principles, language, components, governance, validation and ways of working.

On the right, the better outcomes: reusable patterns, safer adoption, platform consistency, faster delivery and clearer ownership.

  • Product risk

    Different features could explain confidence, uncertainty and sources in different ways.

  • Delivery risk

    Teams could solve the same interaction and accessibility problems repeatedly.

  • Practice risk

    AI confidence could cluster around a few individuals instead of becoming a shared capability.

03

The first AI design decision was whether AI belonged at all.

A model introduces uncertainty, responsibility and failure modes that a deterministic pattern may avoid. The framework made suitability, consequence, autonomy and human verification design inputs before interface work began.

Step through the route

Could a well-designed non-AI solution meet the need?

The route starts from a clear user or business need. If a standard product pattern can meet it effectively, that is the answer, and the work ends in a deterministic interface.

Possible outcome: standard product pattern.

  • Standard product patternA deterministic interface
  • Assistive AIGuided support
  • AI with human verificationA reviewed recommendation
  • Not appropriate for AIDo not automate

An illustrative walk through the representative decision model below. It is not a live tool.

Decision model: a user or business need passes through five questions about standard UX, added value, acceptable confidence, appropriate autonomy and handling low certainty, ending in a standard product pattern, assistive AI, AI with human verification or not appropriate for AI.
Representative decision model. Good AI decisions reduce risk before they create capability.
Read this diagram as text

The model starts with a user or business need: a clear problem, opportunity or outcome to solve. It then asks five questions in order. First, could a well-designed non-AI solution meet the need? Second, will AI create a material improvement over standard approaches? Third, how accurate and reliable does the output need to be? Fourth, should AI inform, assist or act autonomously? Fifth, how will we detect, mitigate and respond to uncertainty or errors?

There are four possible outcomes: a standard product pattern (a deterministic interface), assistive AI (guided support), AI with human verification (a reviewed recommendation) or not appropriate for AI (do not automate).

The key insight on the diagram reads: good AI decisions reduce risk before they create capability. Being disciplined about when not to use AI builds trust, controls risk and creates the conditions for AI to deliver real value where it matters most.

  • Match autonomy to consequence

    AI could inform, assist or act only when the impact, reversibility and review route supported that level of control.

  • Design uncertainty as a state

    Low confidence needed a visible explanation, a safe fallback and a route to verified information.

  • Build accessibility into the pattern

    Reading order, loading announcements, keyboard behaviour and recovery changed the component before handover.

04

One AI language made separate experiences accountable in the same way.

Chat, semantic search, summaries, explanations and recommendations needed a common contract: explain what AI is doing, show what it knows, admit what it does not and give the person a clear next action.

Chat assistant, benefit summary and semantic search feeding one Darwin AI language, used by recommendations, contextual help and AI-generated explanations, with six core principles.
One language model carried explanation, uncertainty and guidance across the platform.
Read this diagram as text

Three inputs, the chat assistant, benefit summary and semantic search, feed a single Darwin AI language. It serves three outputs: recommendations, contextual help and AI-generated explanations.

Six core principles sit beneath it: explain what AI is doing, show what it knows, admit what it does not know, preserve uncertainty, avoid false confidence and give the user a next action.

A panel describes what good looks like: plain English, clear confidence, helpful tone, visible reasoning and safe fallback. The diagram notes it was used across the Reward Centre, the Control Centre and internal UX workflows.

Principles only scaled when they became buildable components.

Language rules became confidence, rationale, verification, feedback and fallback patterns. Those patterns became reusable components, then moved into product work without reopening every decision.

Four AI principles passing through interaction patterns and reusable components into product use.
Principle became pattern, pattern became component, and the component made adoption safer and faster.
Read this diagram as text

Four rows trace a principle to its use in product. Preserve uncertainty became the confidence pattern, then the confidence and verification component, used in chat and recommendation experiences. Explain the reason became the why-this-answer pattern, then the evidence and rationale panel, used in summary and assistant experiences.

Give the user control became the review and confirm pattern, then feedback and action controls, used in assistant and AI-generated content flows. Fail safely became the fallback and escalation pattern, then the low-confidence state, used in search and support journeys.

The closing line reads: principles became repeatable patterns, then reusable components, then scalable product behaviour.

Focused design-system section showing a reusable AI summary component inside a live benefit card.
The summary component carried attribution, follow-up prompts and a clear separation between generated and source content.
Focused design-system section showing AI action buttons, explanatory tooltips and reusable prompt actions.
Prompt design became an interaction system: useful starting points without hiding what the action would do.

05

The framework had to survive real product moments, not just a standards document.

AI summaries, comparisons and assistants put the system under pressure. Each example needed to preserve the source task, show where generated content began and make uncertainty or escalation actionable. These are design explorations, not shipped features.

Focused Darwin benefit page showing an AI policy summary grounded in a verified policy source.
AI summary in context. Source verification, useful questions and a direct route back into the benefit task stayed together.
Close-up of the Darwin assistant comparing cost, cover and dependant differences between two health benefits.
Expose the material differences. Cost, cover, dependants and excess were structured before any recommendation.
AI-assisted benefit comparison modal with five cover options and a clear generated-content disclosure.
Structure before recommendation. Comparison made evidence inspectable before asking for a decision.

Loading, uncertainty and human takeover were part of the component contract.

A polished answer is only one state. The pattern also defined what the person hears while a result is generated, how low confidence is stated and which context follows them when a human needs to take over.

AI summary generation state with progress skeletons and a cancel action.
Generating. Remained cancellable and understandable.
Focused Darwin assistant response that clearly states it cannot confirm the answer and explains what the available evidence does and does not cover.
Uncertain. Explicit, and grounded in the available evidence.
Focused Darwin human handoff screen showing why specialist support is appropriate, which context can be shared and three contact routes.
Human verification. Escalation as a transition, not an apology. The handoff explained why judgement was needed, let the person control which context travelled and made ownership, response time and return routes explicit.

06

AI changed the product. It also changed how UX discovered, tested and delivered the work.

I helped designers move from occasional tool use towards practical, repeatable workflows. The aim was not adoption for its own sake. It was better judgement, richer prototypes and a team able to work confidently alongside AI-enabled engineering.

Traditional UX delivery compared with an AI-enabled operating model spanning problem framing, exploration, design-system application, interactive prototyping, autonomous usability testing, synthesis and handover.
The operating model joined everyday AI assistance to accessibility, governance, design-system ownership and quality assurance.
Read this diagram as text

The top row shows the earlier linear process: requirements, design, prototype, handover, build and QA, with limited feedback and scale.

Beneath it, the AI-enabled model runs problem framing, AI-supported exploration, reusable design system, interactive prototype, autonomous usability testing, AI insight synthesis, refinement and implementation handover. Supporting layers sit around it: AI assistants for everyday support, design-system ownership, prompt optimisation, accessibility checks, governance guidance and QA support.

The stated outcomes are faster iteration, clearer testing, reusable knowledge, stronger consistency and better support for teams.

  • Research and content

    AI supported insight gathering, data analysis, usability-test synthesis, UX writing and copy exploration. Autonomous test runs accelerated feedback while designers remained accountable for interpretation.

  • Prototype and learn

    Prompt design, reusable assistants, agents and agentic workflows supported richer exploration. Vibe coding helped UX build interactive prototypes that could test states static frames concealed.

  • Enable and govern

    I translated technical concepts into usable guidance, supported less-confident designers and connected new tools to accessibility, ownership and repeatable quality checks.

07

A capability that could travel further than one feature or one designer.

The framework connected customer-facing uses, system foundations and team enablement. It gave Darwin a route from principle to product implementation, then back into shared learning.

Darwin AI experience framework connecting product uses, principles, language, components, governance, accessibility, validation and team enablement.
One framework could support product experience, governance and everyday design work without fragmenting.
Read this diagram as text

The Darwin AI experience framework sits at the centre as one coherent system for language, behaviour, governance and reuse. Three groups feed it.

Customer-facing product uses: chat assistant, benefit summary, contextual guidance, semantic support and recommendations. System and governance foundations: principles, language, components, governance, accessibility and validation. Team and workflow enablement: an AI mentor for designers, prompt optimisation, research synthesis, AI suitability checks, prototype support and UX QA assistance.

Below, the framework is applied across the Reward Centre, the Control Centre and the design workflow. The closing line reads: from isolated AI experiments to one scalable UX capability.

  • Product

    A consistent experience language for summaries, search, recommendations, assistants and generated explanations.

  • System

    Reusable interaction patterns and components carrying confidence, sources, fallback and accessibility into delivery.

  • Practice

    Practical workflows, prompting guidance, assistants, prototypes and research support that helped more designers work confidently with AI.