Fewer than 100 Fellows nationally. PHP 300K–600K monthly market rate. Most mid-market HMOs and life insurers operate without in-house actuarial capability. The Actuarial AI Agent is a capability enabler — IC-compliant reserving, fraud detection, and lapse modeling without a Fellow on staff.
Manual chain-ladder spreadsheets become same-day automated computations. Static lapse assumptions become Cox regression on policy-level data. Fraud review becomes Day-1 anomaly scoring.
Quarterly IBNR cycles take 3–5 business days per product line. Mortality assumption updates run annually at best. Lapse modeling uses static industry averages. IC valuations are outsourced at PHP 150K–250K per engagement. Fraud review happens post-settlement.
Upload claims data. The agent validates, computes mortality A/E ratios, IBNR development factors, lapse survival curves, and fraud scores automatically. Query results in plain English. Every output traceable to a specific computation. Audit-ready by design.
Computation lives in Layer 2 — proven Python actuarial libraries. AI reasoning lives in Layer 3 — but only ever reasons over Layer 2's structured outputs. No hallucinated numbers.
Mid-market insurer with 1 Fellow + 1 Associate, or HMO with no in-house actuary at all. The platform is a productivity tool for one and a capability enabler for the other.
Five phases. Each independently testable. Lead actuary co-designs the NL query UX in Phase 3 — adoption isn't an afterthought.
FastAPI scaffold. PostgreSQL schema. Schema validation. Sample data integration.
IBNR chain-ladder. Mortality qx. Morbidity GLM. Lapse Kaplan-Meier.
Phoenix integration. Context assembler. Grounded NL queries with actuary co-design.
Isolation Forest scorer. XGBoost risk. Threshold tuning. Validation set AUC > 0.82.
React dashboard. IC report formats. Audit log UI. Lead actuary UAT sign-off.
No reserve figure, assumption update, or regulatory submission is generated without actuary review. The platform produces every supporting schedule the Appointed Actuary needs — the actuary signs off.
"This platform doesn't replace actuaries. It removes the computational burden that occupies 60–70% of their working hours — data preparation, triangle construction, assumption lookups, report formatting — and returns that time to the work that requires professional judgment."
Two-week scope kickoff. Twelve-week build. Sixteen-week UAT-ready. Then continuous operation with quarterly model recalibration.