A specialist MGA operating in the Lloyd’s market
Submission triage from 4.2 days
to 18 minutes
400+ submissions a week, twelve co-operating agents, and not one underwriter replaced.
- Sector
- Specialty insurance · MGA
- Engagement
- Production deployment
- Timeline
- 14 weeks
- Stack
- Azure · LangChain · GPT-4o
The challenge
What the team was living with
A specialist MGA writing across the Lloyd’s market was receiving upwards of four hundred new submissions a week. Each one needed manual extraction, risk categorisation, sanctions screening, capacity routing and preliminary underwriting notes before a senior underwriter could form any view at all.
The bottleneck sat at the very start of the deal, which is the most expensive place for one to sit. Broker relationships are decided in the first days of a submission, and the team was spending four of them on administration.
There was also no audit trail on routing decisions — an FCA exposure entirely separate from the commercial cost, and one that would have surfaced at the next review.
- 400+ weekly submissions processed entirely by hand
- 4.2 days average from receipt to underwriter desk
- 28% of submissions declined without review, for want of capacity
- Inconsistent triage quality across the team, with no defined standard
- No audit trail on routing decisions — an FCA compliance risk
- Senior underwriters spending around 40% of their time on administration
The approach
Twelve single-purpose agents that read, score and route
before any human opens the file
Design the pipeline before writing any of it
Task boundaries, tool permissions, escalation logic, output schema and failure modes were specified and signed off for all twelve agents first. Nothing was built until the hand-off contract between each pair of agents was agreed.
Build ingestion and extraction
Multi-format ingestion across email, PDF and ACORD SLIP, then entity extraction into a structured schema. Extraction quality sets the ceiling for everything downstream, so it was built and validated first.
Add screening and capacity as hard gates
Sanctions screening and capacity checks were implemented as gates rather than signals. A submission does not progress past a failed gate regardless of how attractive the risk looks.
Score risk, appetite and priority
Risk profiling across six underwriting criteria, then appetite scoring against the MGA’s own written criteria and historical decisions, then a priority score. Each step XAI-traced.
Route with a context pack, and log everything
The routing agent selects the underwriter and the context pack agent assembles extracted entities, screening results, capacity position and comparable historical risks. Every decision in the chain is logged, explained and attributable.
Validate against a senior underwriter benchmark
The pipeline was measured against senior underwriter routing decisions on a hold-out set before go-live, then monitored for three months in production against the same benchmark.
The architecture
Twelve agents, one auditable chain
Each agent is purpose-built for a single task. No monolithic prompts, no hallucination chains. Every step produces a structured, auditable output that the next step consumes.
- 01 Ingestion & format parse
- 02 Document classification
- 03 Entity extraction
- 04 Sanctions screen
- 05 Capacity check
- 06 Risk profiling
- 07 Appetite scoring
- 08 Historical comparison
- 09 Priority scoring
- 10 XAI summariser
- 11 Routing decision
- 12 Context pack builder
- Underwriter review
- Override capture
- Audit log
Outcomes
What changed, measured
The change underwriters describe is not the automation. It is that submissions now arrive on their desk with the screening done, the capacity position known and comparable risks attached — four days earlier than before, and with every step of how it got there available to read.
Take it with you
The full engagement, in twelve pages
Architecture decisions, what we got wrong on the way, and the measurement methodology behind every figure on this page.