Insurance Insights

There is currently £14bn outstanding debt on credit cards in the UK, with 1 in 10 reporting concern about repayments and 1 in 4 reporting serious impact to their mental wellbeing. Credit Card Insights set out to solve a straightforward problem, helping people understand which credit card actually makes sense for them to accelerate their debt repayment and provide a route to removing the financial distress. Most customers know a balance transfer or rewards card could save them money, but the comparison process is overwhelming. Pages of APRs, fees, eligibility criteria and promotional periods leave people paralysed rather than empowered.

We wanted to use their own financial data to show them what mattered specifically to them, explain it clearly and give them confidence to act, by understanding the positive impact this product can have to somebody already in a financially distressed state of mind. Critically, this had to stay firmly within non-advised guidance. The FCA distinguishes sharply between generic information and personal recommendations, and we needed to walk that line precisely.

Success meant measurably increasing customer confidence and conversion while maintaining absolute regulatory discipline. We aimed to prove that generative AI could be deployed responsibly in consumer credit without crossing into advice.

The results validated both the technical approach and the business value. We saw a 6% engagement rate with the insights feature, and customers who engaged converted at 24% higher rates than the control group. Over three months, this translated to 178,000 additional customers finding and applying for credit cards through our platform.

Beyond the numbers, qualitative feedback confirmed we'd reduced cognitive load. Customers reported feeling clearer about their options and more confident in their choices. Our compliance team and the FCA have seen no issues with the guardrails holding.

We've also proved the commercial case for responsible AI in regulated financial services. The technology works, the controls hold and customers respond positively when you give them genuine insight rather than generic comparison.

The technical complexity was high but manageable. Building reliable LLM orchestration with safety controls required careful prompt engineering and testing.

The real difficulty was navigating regulatory uncertainty. There's no FCA playbook for using generative AI in non-advised credit journeys. We had to define our own interpretation of where information ends and advice begins, document our controls rigorously and build internal confidence with compliance and legal teams who were understandably cautious.

We also had to move quickly. Three months from concept to production meant we couldn't afford extended review cycles. That required very tight collaboration between engineering, product, compliance, and executive leadership to make rapid decisions with appropriate risk consideration.

Getting all of that aligned while maintaining quality and safety standards was the hardest part.

Credit Card Insights addresses a persistent friction point in consumer credit comparison. Customers visit our site knowing they want a better credit card deal, but the journey traditionally dumps them into tables of products with dozens of variables to decode. Even when we surface eligibility likelihood, customers still struggle to interpret what a 0% period or balance transfer fee actually means for their situation.

We realised we had the data to answer their real questions. We know their credit profile, their current card usage patterns, and the products they're eligible for. The missing piece was translating that into plain language that builds understanding and confidence.

The solution uses a secure generative AI stack running on AWS Bedrock. When a customer reaches the comparison page, we pass anonymised context about their financial profile and the products they're viewing into a tightly controlled prompt structure. The LLM generates short, personalised explanations that surface the most relevant details for that customer. For example, if someone has existing card debt, we highlight balance transfer savings potential and break down how long their 0% period gives them to clear it.

The critical technical work was in the orchestration layer. We built a prompt control system with golden examples to ensure output consistency, safety checks to catch any language that could be construed as advice, and fallback logic if the model produces anything outside acceptable bounds. Every response goes through guardrails before reaching the customer.

On the regulatory side, we worked closely with our compliance team to define precisely what constitutes information versus recommendation. The Consumer Duty rules require that we help customers make informed decisions, but we cannot tell them which decision to make. Our messaging was carefully tested to ensure it explained product features and potential outcomes without crossing that line.

The design decision to avoid a chatbot interface was deliberate. Chatbots encourage customers to ask open-ended questions, which increases the risk of expecting advice. Our structured overlay format keeps the experience bounded. Customers see insights generated for them, but they don't have a text box inviting "which card should I choose?"

We instrumented everything for testing. Multiple A/B experiments validated messaging approaches, UI treatments, and call-to-action language. We also ran qualitative research to confirm customers understood they were receiving guidance, not advice and felt more confident after seeing the insights. 96% of users who engaged reported a positive interaction.

The project ran for three months with a team of four, including backend engineering, frontend development, product management and my direct oversight on architecture and regulatory controls.