Enter password to view case study
Product Design
/
AI
Aurora
Replacing a century-old workflow with an AI-native underwriting platform
Overview
Commercial insurance still runs on email. Brokers send submissions as PDFs, email text and spreadsheets. Underwriters re-key the same data across systems. Quotes take weeks. Nobody has a single source of truth.
I joined Aurora as the product design lead to build a solution that replaces this fragmented process with a platform that reads the email, extracts the data, provides data-driven insights, automates outdated processes and lets humans focus on decisions instead of data entry.
The result: what used to take days now takes minutes, with fewer errors, full auditability and the human still in control.
Year
2024 – Current
Company
Aurora
Role
Lead Product Designer
Scope
0 - 1
UX Strategy & Research
Design System creation,
Interaction design
Prototyping
What was the impact?
GWP growth
£221k → £1.72M run-rate in 5 months
Enquiry volume
50 → 259/month (5× in 7 months)
Quote rate
75–81% maintained as volume scaled 5×
STP rate
52% of quotes fully automated
conversion
27–34% enquiry to bind
fundraising
The current flow showing multiple pain points
The problem
Insurance has software. Portals, CRMs, policy admin systems. But here's what actually happens: A broker emails a PDF to an underwriter. The underwriter reads pages, copies data into a spreadsheet, checks rules from memory, prices in another tool, writes a quote in Word, and emails it back. Something missing? Another round trip. This takes days. Sometimes weeks.
Everyone knows it's broken. But it stays this way because brokers won't adopt new convoluted portals, it just adds more work to their side. And underwriters rely on experience and gut feel that's genuinely hard to put into software.
The business cost is real. Every day a quote sits unfinished is premium that might walk to a competitor. In insurance, speed affects gross written premium, combined ratio, and ultimately the insurer's float. Slow quoting doesn't just frustrate brokers. It costs money.
So the big question: How do you build a system that meets everyone's needs without asking anyone to change their behaviour?
during interviews we captured their digital workspaces, and identified legacy programs still being used today, causing high friction
Persona documents mapping broker & underwriter profiles
The audience is multi-sided
The platform sells to insurers, is used by underwriters and brokers, with downstream benefits to businesses (policyholders). Success looks different for each group, and that fundamentally shapes the experience and product goals.
Insurers want profitable growth, governance, and faster iteration on pricing and rules.
Underwriters want structured data on arrival, fewer copy-paste jobs, clear rationale
Brokers want less admin, faster turnaround, no new tool to learn
we did a deep dive by creating a full service journey map just to highlight the multiple stakeholders and processes involved.
post brainstorming I took charge of putting together a table of our features mapped to pain points. We later prioritised based on this.
Research & insights
Across the ecosystem, pain clustered into four repeatable patterns:
Unstructured data everywhere
Submissions arrive as PDFs, spreadsheets, email attachments — no consistent format.
Manual re-keying and formatting
Underwriters spend hours on data entry instead of risk assessment and judgment.
Slow feedback loops
Pricing and rules improve slowly because information is fragmented and outcomes hard to trace.
Low transparency
Appetite and rationale are not consistently visible or explainable to brokers or internally.
When we interviewed users, we found the workflow breaks at every handoff: brokers submit in inconsistent formats via email, underwriters re-key instead of deciding, insurers lack tools for rapid iteration, and quote-to-bind timelines stretch to weeks because information is scattered and incomplete. The practical result: lost momentum and wasted expertise.
Understanding this, I worked together with the Head of Product, our Head of Delivery and our CTO to identify crucial areas within the platform that we should focus on first.
Prototype of a broker submission that has been pre-filled that we tested with 5 brokers to quickly determine usability.
I built out some navigation schemas to study and plan for multi-user needs (crossing broker and underwriter workflows)
UX principles to govern decisions
From the research, I also created a set of UX principles anchored around human outcomes that were important to our users: reducing cognitive load, building trust in AI outputs, and keeping professionals in control.
recognition
over recall
Prefer suggestions, pre-filled answers, and templates over blank forms.
progressive
disclosure
Show the minimum to move forward. Reveal depth when the user asks "prove it."
trust
calibration
For high-stakes steps, always show what we recommend, why, and what evidence supports it.
Automation without ambiguity
Auto-fill only when confidence is high. Otherwise, suggest and require confirmation.
Coming up with these first helped me create prototypes quickly. that we could then rapidly test before handing off to our developers to build.
The new paradigm
Instead of asking users to change their behaviour, we designed a pipeline that transforms existing workflows into structured, auditable processes. Each step solved a specific frustration I heard in research.
Our solution cutting down time to bind significantly
Start where they already are
Brokers live in their emails. That's not going to change even after our first iteration of the product, and multiple interviews. So we made the platform start from their inbox.
They send an email like they always have. The system reads it, pulls out the data, and sends back a smart reply: here's what we found, and here are specific things you're missing in your submission, and what to do next.
This user flow increased our submission rate substantially.
Review over re-keying.
Afterwards, the core experience here was to make sure review, and submission of documents was forefront over editing fields. There were still inevitably going to be some data they might not have documents for, but the key was to ensure it was as little friction as possible in filling the missing questions.
This was core to our submission to quote conversion metric. Every field the broker doesn't have to re-enter, every question they can answer in-line instead of digging through files, is friction removed from top of the funnel.
Best part? We also reduced preventable referrals with better intake design. Fewer cases bouncing back because of missing data that should have been captured upfront.
The underwriter platform
The core design challenge here was designing one system that serves multiple roles with genuinely different needs, risk tolerances, and mental models.
Brokers care about speed and low friction. Underwriters care about accuracy, traceability, and staying within their authority limits. The same data needs to serve both, but the interface can't be the same.
I designed the underwriter platform as the operational core. Where the AI extraction lands, where risk gets assessed, where referrals get resolved, and where decisions get documented.
AI Data extraction
For Underwriters
The real bottleneck isn't underwriting judgment. It's the hour spent reading a PDF and keying in data into forms or spreadsheets.
Underwriters shared early on,"I won't trust a number I can't trace." So I designed source tracing into every data enriched or pre-filled field. Click it, see the source. Edit if wrong (because we know AI can still be wrong) and the audit trail logs who changed what.
Instead of vague medium, high risk scores, I also made sure each AI insight highlighted the reasoning, and the data or document source. Not a black box, but a transparent system.
AI Data extraction
for brokers
For brokers, the experience is slightly different. We had to ensure that brokers had a way to upload any missing documents at every stage of the journey, and that they could reliably track the source source of any extracted data.
The north star metric here is straight-through processing rate, the percentage of cases that go from intake to decision with minimal human intervention.
From email submissions to quote alone, we hit 52% STP at launch.
Aurora CoPilot
From. watching the way underwriters work plus. some surveys, we knew that most underwriters preferred intelligence that augmented their work, as their decisions were intuitive. We also knew they spent a lot of time doing "unnecessary manual work" scattered across tabs, spreadsheets, and their own memory.
So we built a copilot. The last thing an underwriter processing a queue of cases needs is to start from a blank conversational interface. I designed AI that shows up in context: enrichment cards alongside the fields they relate to, risk insights ranked by impact, evidence assembled automatically as referrals get resolved.
The key design decision I made was deciding where AI appears and where it stays quiet. Insights surface only when they're relevant to the field being reviewed. The AI never interrupts the workflow, it augments it.
I also created a trust signals .MD doc. This was a reference doc for AI to follow that defined how the system should communicate confidence, uncertainty, and evidence across every AI interaction surface.
Full case study for aurora copilot WIP
Outcome
£221k to £1.72M GWP run-rate in 5 months.
52% straight-through processing rate at launch, with 3 underwriters handling 480 quotes/month
Launched AI-enabled platform to QBE & Howden.
In parallel, getting multiple insurer interest enquiries, with new product lines in onboarding