The Underwriter's Decision Stack

Eight layers. One closed loop. The intellectual foundation for everything UnderwriteAI builds.

Version 1 · 2025 · Proprietary

An Underwriting Decision Is Not A Moment

It is a stack of dependent judgments that compound from raw data at the bottom to a bound policy at the top — and then loop back through portfolio learning. Data flows upward through seven decision layers. Learning flows downward from the eighth back to the first. The system is a closed loop, not a pipeline.

Most AI in insurance today attacks one layer of that stack in isolation. A submission triage tool here, a pricing optimiser there, a CAT aggregation dashboard somewhere else. Each one locally intelligent inside a globally dumb system. The underwriter in the middle is still the integration layer — manually connecting outputs from disconnected tools under time pressure at the point of a decision with real financial consequences.

The Underwriter's Decision Stack reframes commercial underwriting as a vertically integrated reasoning pipeline where AI inserts at every layer — and where coverage structuring, pricing adequacy, and reinsurance protection are not adjacent functions but integral layers within the underwriting decision itself.


01
Layer 1: Data

Data

"What is true about this risk?"

Submission documents, SOV, COPE data, loss history, and external signals (satellite imagery, geocoded hazard, financials, news). The layer's job is evidence assembly.

AI Role: Extraction, normalization, gap detection, third-party data enrichment.
02
Layer 2: Signal

Signal

"What does the evidence mean?"

Occupancy risk grade, NatCat exposure score, COPE quality rating, moral hazard markers, and loss trend. Evidence becomes interpretation.

AI Role: Feature engineering, anomaly detection, signal scoring, explanation generation.
03
Layer 3: Coverage and T&C

Coverage and T&C

"What are we prepared to cover?"

Perils, limits, sub-limits, deductibles, warranties, and conditions. Coverage is defined before price.

AI Role: Term structuring, clause recommendation, coverage gap identification, T&C benchmarking.
04
Layer 4: Pricing

Pricing

"What is the adequate premium for these terms?"

Burning cost, exposure rating, GLM/GBM technical premium, CAT load, and capital loading. Pricing follows coverage.

AI Role: Model orchestration, scenario pricing, sensitivity analysis, deviation justification.
05
Layer 5: Judgment

Judgment

"What does experience say the model is missing?"

Market context, broker relationship, moral hazard read, and soft information no dataset captures.

AI Role: Structured judgment capture, peer benchmarking, bias detection, decision audit trails.
06
Layer 6: Reinsurance

Reinsurance

"How does this fit net retention and treaty economics?"

Fac need, cession logic, event limit consumption, net vs gross profitability, and RI cost per risk.

AI Role: Treaty consumption tracking, fac intelligence, net pricing optimisation.
07
Layer 7: Decision

Decision

"Bind, decline, refer, or counter?"

All upstream layers converge. Final terms confirmed.

AI Role: Referral routing, counter-offer generation, term optimisation, audit trail.
08
Layer 8: Learning

Learning

"What did this decision teach the portfolio?"

Portfolio metrics update, models recalibrate, and underwriter performance tracked. Closed loop.

AI Role: Closed-loop learning, drift detection, portfolio-aware re-underwriting signals.

Why This Structure Is Defensible

1. Mirrors real operations

Anyone who has underwritten a large commercial property schedule will recognise the sequence. Anyone who has not will not be able to fake it.

2. Coverage precedes pricing

Layer 3 (Coverage and T&C) sits above Signal and below Pricing. You define what you are covering before you price it. No credible actuarial or underwriting framework disagrees with this.

3. Pricing and Reinsurance are layers not silos

Every org chart treats actuarial and reinsurance as separate departments. This framework asserts they are integral layers — and the AI implications follow directly from that assertion.

4. Every layer has a clear AI insertion pattern

That means every layer can become a project, a repository, a whitepaper, an engineering engagement. The framework generates the roadmap.

5. It is diagnosable

A CUO can look at it and answer immediately: we are strong on Layers 1-2, adequate on 4, blind on 6. That is a consulting conversation that starts itself.


From Framework to Build

Every project addresses one or more layers. Every layer will eventually have a working AI system.

ProjectPrimary LayersDomainStatus
P1 SubmissionIQLayers 1, 2, 3, 4Commercial PropertyIn Build
P2 PriceDeskLayers 4, 2, 6, 8Technical PricingScoped
P3 LossLensLayers 7, 1, 2, 8Claims IntelligenceScoped
P4 RiskReaderLayers 2, 1, 3Risk EngineeringScoped
P5 TreatyDeskLayers 6, 5, 4Reinsurance AnalyticsScoped
P6 PortfolioMindLayers 5, 3, 6, 8Portfolio IntelligenceScoped
P7 AuditStackLayers 8, 7, 5Decision GovernanceScoped
Specialty Line Extensions
P8 ParametricDeskLayers 3, 4, 2Parametric InsuranceScoped
P9 RenewableUWDeskLayers 1, 2, 3, 4Renewable EnergyScoped
P10 DataCentreUWDeskLayers 1, 2, 3, 4Data CentreScoped
Explore All Projects →