AI Workflow Automation in London, United Kingdom

Insurance / Lloyd's Market

A Lloyd's Market managing agent (operating 3 syndicates, £1.8B combined gross written premium, specialising in property, marine, and energy risks) processed 28,000 claims annually through a team of 62 claims handlers and 14 claims managers. The claims workflow involved: (1) Loss notification receipt (email from broker with loss details, policy reference, and supporting documentation — often 15-40 pages including loss adjuster reports, surveyor reports, and coverage opinions). (2) Claims handler opens a new claim in the claims management system, manually entering details from the notification. (3) Policy verification — handler confirms coverage applies, checking policy terms, endorsements, exclusions, and territorial scope. (4) Reserve setting — handler estimates the likely claim cost based on loss details and historical claims data. (5) For claims above £50,000: referral to claims manager for review and approval. (6) For co-insured risks: coordination with following market underwriters (typically 5-12 participants), each requiring notification and agreement on claims handling approach. (7) Adjuster appointment (for property and energy claims requiring investigation). (8) Settlement calculation, negotiation, and payment. The problems: (1) Processing time: average 4.2 days from loss notification to claim acknowledgement. Brokers expected same-day acknowledgement for major losses. (2) Data re-keying: handlers manually entered claim details from broker notifications into the claims system — an average of 22 minutes per claim, with a 6.2% error rate in data entry. (3) Coverage verification: handlers spent 35-55 minutes per claim verifying coverage against policy terms — reviewing policy wordings, endorsements, and territorial limitations. Errors in coverage verification (missing exclusions, incorrect policy attachment) occurred in 2.8% of claims, leading to subsequent coverage disputes. (4) Reserve accuracy: initial reserves were set based on handler judgment and varied significantly between handlers. Reserve accuracy (initial reserve vs. actual paid) was ±38% — requiring frequent reserve adjustments that consumed management time and affected syndicate profitability reporting. (5) Following market coordination: for co-insured risks, the managing agent sent notifications to following markets manually via email. Following market responses averaged 12 days. Several following markets had migrated to electronic claims agreement platforms, but coordination across platforms and email remained manual. (6) Regulatory compliance: Lloyd's Minimum Standards and FCA ICOBS (Insurance: Conduct of Business Sourcebook) require fair and prompt claims handling. The 4.2-day acknowledgement time and inconsistent coverage verification were compliance risks.

We built a 4-component AI claims workflow over 11 weeks. (1) Claims Intake Engine: automating the processing of broker loss notifications. The engine: receiving notifications across channels (email, Lloyd's CLASS, broker portals), classifying document types within each notification (loss notice, adjuster report, surveyor report, coverage opinion, photographs, invoices), extracting structured claim data from unstructured documents (loss date, loss description, affected location, estimated loss amount, policy reference, broker reference, claimant details), cross-referencing extracted policy references against the syndicate's policy database to automatically link claims to policies, and creating claims records in the claims management system with all extracted data pre-populated. Extraction accuracy: 97.8% across key fields. Processing time: 45 seconds per claim intake (from 22 minutes manual data entry). Data entry errors: eliminated (0% vs. 6.2%). (2) Coverage Verification Engine: automating policy coverage analysis. The engine: retrieving the full policy wording (including all endorsements, clauses, and territorial limitations) for the linked policy, analysing the claim against policy terms — checking: peril coverage (is the loss cause covered?), territorial scope (did the loss occur within covered territory?), policy period (did the loss occur during the coverage period?), exclusions (do any policy exclusions apply?), deductibles and retentions (what is the insured's retention?), sub-limits (are any sub-limits applicable to this loss type?), generating a coverage opinion with confidence score — clearly identifying which terms apply, which exclusions might be relevant, and what the deductible/retention structure is, and flagging ambiguous coverage situations for human review (overlapping exclusions, aggregation questions, complex territorial issues). Coverage verification time: 3 minutes per claim (from 35-55 minutes). Coverage errors: reduced from 2.8% to 0.3% (the remaining 0.3% involved genuinely ambiguous coverage situations that the system correctly escalated for human review). (3) Reserve Intelligence Engine: automating initial reserve setting with actuarial-quality estimates. The engine: analysing the claim against historical claims data (15 years of claims history, 340,000+ closed claims with known ultimate costs), matching the current claim to similar historical claims based on: loss type, property type, geographic location, policy year, loss size indicators, and complexity indicators, applying development factors based on claims maturity and class of business, and producing initial reserve recommendations with confidence intervals. Reserve accuracy: ±18% (from ±38%). Reserve volatility: significant reduction in reserve adjustments — 40% fewer reserve movements in the first 6 months. (4) Workflow Orchestration Engine: automating the end-to-end claims workflow. The engine: routing claims based on complexity and value (routine claims under £25,000 fast-tracked with minimal human touchpoints, complex claims routed to specialist handlers with pre-compiled dossiers), generating following market notifications automatically (formatted per Lloyd's Market standards, sent via ECF — Electronic Claims File — where possible, email for markets not on ECF), tracking following market responses and escalating non-responses after defined periods, generating claims reports for management (reserve summaries, open claims aging, handler workload distribution), and producing Lloyd's Minimum Standards compliance documentation automatically.

Claims acknowledgement time: reduced from 4.2 days to same-day for 94% of claims (target met). Claim intake processing: 45 seconds per claim (from 22 minutes) — 3,080 hours of annual handler time redirected to complex claims management. Data entry errors: eliminated (from 6.2%). Coverage verification: 3 minutes per claim (from 35-55 minutes) with 0.3% error rate (from 2.8%). Reserve accuracy: ±18% (from ±38%) — significantly reducing reserve volatility and improving syndicate profitability reporting accuracy. Following market response time: reduced from 12 days to 5 days average (automated follow-ups and ECF integration). Claims team: reduced from 62 handlers to 45 through attrition — remaining handlers focus on complex claims requiring judgment, negotiation, and relationship management. Claims manager approval workflow: automated pre-screening reduced claims manager review time by 60% — managers review AI-generated coverage opinions and reserve recommendations rather than building analysis from scratch. Lloyd's and FCA compliance: positive feedback from Lloyd's Performance Management Directorate on claims handling standards. Zero FCA conduct complaints related to claims processing delays in the 12 months following deployment. Annual value: £3.4M (operational efficiency — handler time savings, reduced rework, faster processing) + £1.8M (reserve accuracy improvement — reduced reserve volatility, better capital allocation) + £620K (compliance risk reduction — faster acknowledgement, consistent coverage verification) = £5.82M. Development cost: £245,000. Payback: under 16 days.