InsureFlow: How an InsurTech Reduced Claim Processing Time by 70% and Fraud Detection to 96%

A Global Insurance Provider processed over 4,200 claims per month across their UK health and motor portfolios. Their legacy system was not a system in any modern sense — it was a patchwork of Access databases, emailed PDFs, and manual investigator notes accumulated over 15 years. Every claim required in-person or postal submission, took 8–12 business days to process, and involved at minimum three handoffs between departments before a decision was reached.

The fraud problem was quietly catastrophic. The client's rule-based fraud detection system — essentially a checklist of red flags built in 2009 — was catching 71% of fraudulent claims. The other 29% were slipping through. Internal estimates put the annual cost of undetected fraud at £2.3 million. The deeper problem was that the fraud they were missing wasn't simple individual fraud — it was collusive fraud: organised rings of claimants, medical professionals, and repair shops filing coordinated, individually plausible claims that no rule-based system could correlate. Humans were missing it too.

The FCA's evolving explainability requirements added a third layer of complexity. Any AI decision that affected a customer claim could be challenged. A black-box model — even a highly accurate one — was off the table. Whatever we built had to be auditable at the decision level. Investigators needed to be able to explain, in writing, why a claim was flagged. The solution couldn't just be accurate; it had to be trustworthy enough to stand up in a regulatory dispute.

We began with a 3-week discovery and data audit. The client's data science team had never attempted ML on their claims data before, so the first job was understanding what they actually had. We ingested 28 months of historical claims — structured fields, free-text medical narratives, vehicle damage descriptions, and adjuster notes — and ran exploratory analysis to map quality issues, class imbalance (fraud represented 6.4% of claims), and feature distribution across policy types.

The AI architecture we designed was deliberately sequential rather than a single model. First, an Autoencoder identifies statistical anomalies — claims that deviate significantly from historical patterns in ways that are hard to specify in rules. Second, a fine-tuned Bio_ClinicalBERT model processes the free-text medical narratives in health claims, flagging language patterns associated with fraudulent or exaggerated injury descriptions that human investigators typically miss on a first read. Third — and most importantly for collusive fraud — a Graph Attention Network maps the relationships between claimants, medical providers, repair facilities, and legal representatives across all claims. Claims that appear individually legitimate but share suspicious network connections are surfaced with high confidence.

Every flagged claim is accompanied by a SHAP (SHapley Additive exPlanations) report generated in plain English: which features contributed to the fraud score, how much each contributed, and what similar legitimate claims look like. This was not optional — it was the design requirement that unlocked FCA compliance. Investigators received training on reading SHAP outputs, and within 3 weeks of go-live they were using the explanations to make faster, more confident decisions than they had before.

InsureFlow replaced The client's fragmented claims operation entirely. Policyholders can now submit claims digitally via the Flutter mobile app — photographing documents, recording damage, and uploading medical certificates from their phone. OCR processing extracts structured data automatically, removing the manual data entry step that had been a source of both delay and error. An automated premium calculation engine cross-references policy terms without human involvement for straightforward claims.

The AI fraud pipeline runs in the background on every submission, returning a fraud probability score, a breakdown of contributing signals, and a recommended action (approve, investigate, escalate) within 30–60 seconds. The React investigator dashboard displays all flagged claims in a prioritised queue, with the full SHAP explanation visible alongside the claim history and a one-click audit trail that records every investigator action for FCA reporting purposes. Claims below the fraud threshold that meet all automated eligibility criteria can now be approved without human review — a straight-through processing rate of 34% within 60 days of launch.

Integration with The client's existing policy management system was handled via REST APIs, meaning there was no forced migration of the policy database — a requirement the client had emphasised from the start. The deployment was rolled out in two phases: first the digital submission and case management layer (weeks 1–10), then the AI fraud pipeline (weeks 11–16), allowing investigators to build trust with the platform before relying on AI-generated scores in their decisions.