Enter password to view case study
Product Design
/
AI
/
Machine Learning
Aurora Copilot
Embedding AI into a complex multi-stakeholder platform
Overview
The copilot is not a chatbot. It's the AI layer embedded across Aurora's platform. It lives inside the workflows of brokers and underwriters, surfacing intelligence where decisions happen.
The constraint from day one: under FCA regulation, the AI can surface information but cannot give advice. It can tell an underwriter a rebuild cost is 40% below comparable properties. It cannot say to increase it. Every piece of AI-generated copy in the platform had to respect that boundary.
the design approach
Ambient & contextual intelligence. The copilot model.
We deliberately chose "copilot" over "autopilot." The AI does not make decisions, it prepares the underwriter to make better ones. We evaluated three interaction patterns: contextual suggestions, ambient insight cards, and blank-canvas chat. We shipped the first two.
year
2024 – Current
company
Aurora
role
Lead Product Designer
focus
Product Design
UX Strategy
Rapid prototyping
User testing
What was the AI impact so far?
field correction rate
< 5%
Cases returning for missing data
~40%
Quote rate
75–81% maintained as volume scaled 5×
the first challenge
Different AI outputs need different trust mechanisms
Senior underwriters trust their own judgment more than any algorithm.
Interviews revealed that underwriters' attitude toward data is intuitive. They rely on experience; data plays a supporting role. Blank-canvas conversational interfaces do not help because underwriters work alongside submissions, evidence packs, and ongoing conversations, not in isolation.
the first challenge
Where should AI intelligence appear?
Senior underwriters trust their own judgment more than any algorithm.
Interviews revealed that underwriters' attitude toward data is intuitive. They rely on experience; data plays a supporting role. Blank-canvas conversational interfaces do not help because underwriters work alongside submissions, evidence packs, and ongoing conversations, not in isolation.
Ambient AI cards
As an underwriter, I want risk signals surfaced in context, and not buried in a separate dashboard I have to navigate to.
Ambient cards surface proactive intelligence alongside the current view, risk signals on a case, market trends across a portfolio, regulatory changes relevant to a product line. What appears depends on context. The pattern is the same: cited sources, no user prompt, relevant to what the underwriter is looking at right now.
I replaced the chat with a contextual, more subtle link. The card signals what needs attention. The deep-dive happens in the case view, where the user has committed to investigating.
this lead to our solution
Contextual intelligence
As an underwriter, I want the AI to have already identified the fields that need my attention so that I'm making decisions, not doing triage again.
A blank-canvas chat would have asked the underwriter to formulate a question before they had read the case. Wrong cognitive sequence. Instead, the sidebar ships pre-populated fields confirmed, fields missing, fields needing review.
the second challenge
The confidence score nobody could explain.
The earlier tests surfaced a specific problem. Users asked: "What does 65% mean?" and "What is high confidence?" A percentage puts interpretive work on the user. Something like "Check source" does the same work but makes it actionable.
We also agreed internally that a percentage implies precision the model does not have.
For underwriters who need auditable, defensible actions, the answer is not better confidence scores. It is specificity.
Different output types, different trust affordances
After discussions at length with our engineer and head of product, we managed to put together a simple table that would dictate how we approach implementing trust.
I then utilised Claude to help codify these rules into a trust language framework md doc. A reference document governing copy patterns, anti-patterns, and CTA structures across all four output types. It is fed directly into the AI pipeline as a system-level constraint, ensuring every piece of AI-generated text in the product follows the same rules without manual review.
As an underwriter, I want to know whether the AI found a value or inferred one, because those two states require completely different actions from me.
A percentage tells the underwriter how confident the model is. A label tells them what to do next. "Found" means verify. "Not found" means chase the broker. The action matters more than the score.
Information, not advisement
The FCA constraint shaped the AI's entire vocabulary. Observational language only: "detected," "below median," "may trigger." Never directive: "should," "must," "recommend."
Outcome
The copilot shipped a month ago as an integrated feature within the Aurora platform and is used by underwriters in their workflow.
Three underwriters handling 480 quotes per month. Insight and pricing adjustment controls at 5% engagement in the first month. The signal to watch is whether this trends upward as usage patterns settle.
Referral rates by product type: Unoccupied 10%, Residential 25%, Commercial 30%, Mixed 40%. Quote rate held while volume scaled. Fewer cases bouncing back for missing data over time.
Metrics in active measurement: Field-level correction rate, First-time completeness rate, Insight engagement rate, Time-to-decision with vs. without insights, Preventable referral rate…etc
What I learnt
Give a human reviewer enough context to make a fast, defensible decision without starting over.
AI confidence is not the same as user confidence. Source provenance builds trust faster than any model accuracy claim.
The interface patterns for showing "why" (progressive disclosure, source links, confidence indicators) were as important as the AI model accuracy itself.
Match AI to cognitive mode. Scanning needs ambient information and links, not inputs. Case investigation needs a pre-populated briefing.
The copilot's value isn't replacing those heuristics but helping underwriters apply them to more structured, complete information faster.