AI Data Pipelines in London

Insurance / Lloyd's Market

A Lloyd's Market managing agent (GBP 3.5B GWP across 4 syndicates, operating in property, casualty, marine, and specialty lines) faced data infrastructure challenges that limited their AI ambitions. (1) Data fragmentation: underwriting data existed across 14 different systems — 3 underwriting platforms (one per syndicate, plus a legacy system), 2 claims systems, Lloyd's market platforms (PPL, CLASS, ECF), broker messaging systems, actuarial modelling tools, and various spreadsheets. No single pipeline connected these systems into a unified view. Underwriters manually gathered data from 4-6 systems for each risk assessment — spending 35% of their time on data retrieval rather than analysis. (2) Data quality was poor: broker submissions arrived in inconsistent formats — PDFs with varying structures, Excel spreadsheets with different column layouts, and unstructured email attachments. The same data field (e.g., "sum insured") appeared with different names, formats, and currencies across submissions. Actuarial models were fed with manually cleaned data — a process that introduced errors and consumed 40% of actuarial analyst time. (3) Historical data was inaccessible: 15 years of underwriting decisions, claims outcomes, and portfolio performance existed in legacy systems and archived files — but was not queryable for ML model training. Building predictive models (pricing, claims frequency, portfolio optimisation) required historical data that took weeks to extract and prepare. (4) Real-time exposure monitoring was impossible: aggregate exposure data (the syndicate's total exposure to catastrophe events, industry concentrations, and geographic concentrations) was calculated monthly from manual data extracts — meaning underwriting decisions were made without current aggregate context. (5) Lloyd's performance management required data submissions in specific formats — preparing these submissions consumed 3 weeks of analyst time per quarter.

We built an end-to-end underwriting analytics pipeline over 16 weeks. (1) Universal ingestion: connectors for all 14 source systems — Lloyd's platforms (PPL API, CLASS data feeds, ECF messaging), underwriting systems (API integration with all 3 platforms), claims systems (event-driven ingestion — claims updates flowing into the pipeline in near-real-time), broker submissions (automated PDF parsing using document AI — extracting structured data from unstructured broker submissions with 94% accuracy), and actuarial models (output ingestion from RMS, AIR, and internal models). (2) Data normalisation and enrichment: canonical data model (all data mapped to a unified schema — standardised policy, claim, exposure, and financial data models), currency normalisation (all monetary values converted to GBP base with original currency preserved — using Bank of England daily rates), entity resolution (matching entities across systems — the same broker, insured, or intermediary identified consistently across all data sources), and enrichment (augmenting internal data with external sources — company financials from Companies House, credit ratings, industry classification, geographic risk data). (3) Feature store for underwriting AI: underwriting features (pre-computed features for each risk submission — loss history analysis, peer comparison, aggregate exposure impact, pricing benchmarks), claims features (claims development patterns, frequency/severity trends, large loss indicators), and portfolio features (aggregate exposure by peril, geography, industry — updated in near-real-time as risks are bound). (4) Real-time exposure pipeline: streaming exposure updates (every bound risk immediately updated the aggregate exposure view), catastrophe model integration (exposure data automatically formatted for catastrophe model ingestion — enabling daily model runs rather than monthly), and alert generation (automated alerts when aggregate exposure approached limits — "Florida windstorm aggregate at 92% of limit"). (5) Regulatory and Lloyd's reporting pipeline: automated data submission preparation (Lloyd's performance management data, regulatory returns — formatted from the pipeline's unified data rather than manually compiled).

Underwriter productivity: data retrieval time reduced from 35% of underwriter time to under 10% — freeing approximately 25 percentage points of capacity for actual risk analysis. Revenue impact: increased underwriting throughput enabled the managing agent to evaluate GBP 120M in additional premium opportunities in the first year. Data quality: actuarial model input data quality improved from 82% accuracy (manual preparation) to 97% accuracy (automated pipeline with validation) — reducing reserve volatility by 18% and improving pricing model accuracy. Historical data: 15 years of underwriting and claims data made queryable — enabling ML model training that improved pricing accuracy by 8% (measured by combined ratio improvement). Real-time exposure: aggregate exposure monitoring moved from monthly to near-real-time — preventing 6 instances of aggregate limit breach in the first year that would have occurred under the old monthly monitoring. Lloyd's reporting: quarterly submission preparation reduced from 3 weeks to 2 days. Annual value: GBP 12.8M (underwriting throughput improvement), GBP 8.5M (pricing accuracy — 8% combined ratio improvement on GBP 3.5B portfolio), GBP 4.2M (exposure management — avoided limit breaches), GBP 2.1M (data quality — reduced reserve volatility), GBP 0.9M (reporting efficiency) = GBP 28.5M. Development cost: GBP 380,000. Payback: under 5 days.