Case Study
Insly AI: Dynamic Pricing Engine
Ando Kivilaid built Insly's dynamic pricing engine — a sub-100ms microservice layer that combines machine-learning loss predictions with classical actuarial rating grids. The system increased auto-binding rates from 22% to 68% while keeping loss ratios bounded, and let insurers deploy new risk-adjusted algorithms in days instead of months.
Auto-binding 22% → 68% · sub-100ms pricing
Published May 19, 2026 · Last updated June 1, 2026
Static rating grids — maintained in spreadsheets or legacy policy admin systems — cannot react when loss trends, weather risk, or macro conditions shift between renewal cycles. Insly's insurer and MGA clients needed a pricing layer that preserved actuarial governance while allowing rapid model iteration and instant quote decisions at the point of sale.
I architected and shipped the dynamic pricing engine as a core Insly AI capability, integrating with existing product configuration, underwriting rules, and broker quoting flows. The design prioritized latency (sub-100ms p95), explainability for actuarial review, and safe rollback when new models underperformed in shadow mode.
This document covers the business problem, technical architecture, and six-month production outcomes including auto-binding lift and loss-ratio guardrails. Metrics cite measurement windows and sample definitions for teams evaluating ML-augmented rating engines.
See also the underwriting automation case study for the document-intake layer that feeds clean risk attributes into this pricing API, and the MCP server project for how external data can reach AI tooling in production.
Problem
Traditional rating engines encode premiums in versioned Excel grids or mainframe tables. Actuaries need weeks to publish grid changes; brokers lose deals when competitors quote faster with updated risk factors. Insly clients reported auto-binding rates stuck near 22% because manual referral rules triggered on any model uncertainty.
Machine-learning prototypes existed in data science notebooks but could not meet production latency, audit, or integration requirements. Underwriters distrusted black-box scores without mapping to familiar rating factors. Regulators and internal actuarial teams required traceable formulas linking ML outputs to base premiums.
The target state: a single API returning bindable premiums in under 100 milliseconds, combining ML risk signals with approved rating grids, with automatic referral only when governance rules — not model noise — require human review.
Architecture
The engine is a stateless microservice cluster behind Insly's quoting API. Each request carries product code, risk attributes, and optional external signals (weather indices, macro series). A feature store assembles normalized inputs; parallel workers fetch historical loss runs and cached third-party enrichments.
An ML inference tier produces loss-cost adjustments and confidence intervals from gradient-boosted models trained on anonymized portfolio data. A rating orchestration layer applies actuarial formulas — base rates, territory factors, experience mods — then scales premiums by the ML adjustment within actuary-approved bounds. Every step emits an explanation object for underwriter and actuarial dashboards.
Shadow deployment compares candidate models against production without affecting bind decisions. Promotion requires actuarial sign-off and automated back-tests on holdout portfolios. Circuit breakers revert to grid-only pricing if latency exceeds SLO or model outputs breach loss-ratio guardrails.
Clients configure products through Insly's admin UI: which ML features are active, bound multipliers, referral thresholds, and grid versions per territory. Version pinning ensures reproducible quotes — the same inputs and model hash always return the same premium for audit disputes.
Results
Across six months of production (H2 2024 – H1 2025) on three live commercial lines, auto-binding increased from 22% to 68% of quote requests that previously referred to underwriters. Median API latency held at 74ms p95 under peak load. Loss ratios on auto-bound business stayed within ±2% of actuarial targets.
Insurers deployed four new risk-adjusted algorithms in an average of 11 days from model approval to production — down from 8–12 weeks with grid-only releases. Broker NPS on quote speed improved measurably on pilot territories. Revenue per underwriter hour increased because referrals concentrated on genuinely non-standard risks.
The engine now underpins pricing for multiple Insly platform tenants. Key lesson: ML adjustments work in insurance only when actuaries own the bounds and every quote remains explainable to a regulator. Speed without governance is a liability; speed with governance is a competitive moat.
Product teams reused the same shadow-mode and circuit-breaker patterns when rolling out form automation and document extraction — reducing cross-team risk when multiple AI services share broker-facing quote flows. That shared platform discipline accelerated every subsequent Insly AI release on the roadmap.
Key metrics
| Metric | Value | How measured |
|---|---|---|
| Auto-binding rate | 22% → 68% | Share of eligible quotes bound without underwriter referral, three commercial lines, 6-month window vs. prior-year baseline. |
| Pricing latency | 74ms p95 | End-to-end API response time at 95th percentile under production peak load, measured continuously via APM. |
| Deployment time | 11 days avg | Calendar days from actuarial model approval to first production quote, four algorithm releases in 2024–2025. |
| Loss ratio variance | Within ±2% | Auto-bound portfolio loss ratio vs. actuarial plan, rolling 6-month windows on pilot lines. |