Case Study
Insly AI: Underwriting Automation Pipeline
As Head of AI Strategy at Insly, Ando Kivilaid designed and deployed an intelligent document processing pipeline that extracts commercial policy data from unstructured PDFs with 99.4% validated accuracy. The system automates 90% of manual submission triage, cutting average review time from 45 minutes to 90 seconds and saving over €500,000 in annual processing costs across the platform.
99.4% accuracy · 45 min → 90 s · €500k+ saved/yr
Published May 19, 2026 · Last updated June 1, 2026
Commercial insurance underwriting begins with intake: brokers submit policy declarations, loss runs, schedules of values, and email threads — often hundreds of pages in mixed formats. Before any risk decision, underwriters must triage, classify, and extract key fields. At Insly scale, this manual step became the bottleneck across thousands of broker submissions per month.
I led this project as Head of AI Strategy & Implementation, working directly with underwriting operations, product, and engineering. The goal was not a chatbot overlay but a production pipeline integrated into Insly's existing broker and policy workflows — with audit trails, human-in-the-loop review for edge cases, and measurable accuracy targets agreed with the underwriting team before launch.
The case study below documents the problem context, system architecture, and validated production results. Metrics include sample size and measurement period so teams evaluating similar IDP projects can compare methodology, not just headline figures.
Related work on this site includes the dynamic pricing engine case study and the Eesti Data MCP Server — together they illustrate the full stack from document intake through rated premium to external data tooling for AI assistants.
Problem
Insly brokers upload commercial submissions as PDFs, scans, and email attachments. Underwriters spent an average of 45 minutes per submission locating coverages, limits, deductibles, loss history, and occupancy details across inconsistent document layouts. Errors in manual transcription delayed quotes and created leakage when incorrect limits entered rating systems.
Legacy OCR tools failed on complex tables, multi-column schedules, and handwritten endorsements. Generic large-language-model prompts hallucinated field values when documents exceeded context windows or mixed languages. The underwriting team needed extraction they could trust in production — with coordinates back to source pages for audit.
Volume compounded the problem: peak periods brought hundreds of concurrent submissions. Hiring more underwriters for triage alone was not scalable. The business case targeted 90% automation of triage, sub-two-minute review for standard risks, and six-figure annual savings in fully loaded underwriting labor.
Architecture
The pipeline ingests documents through Insly's submission API, normalizes PDFs and images, and runs layout-aware parsing to detect tables, headers, and continuation blocks. Semantic chunking splits long policies into bounded segments while preserving cross-references between endorsements and base declarations.
Dynamic prompt assembly selects field schemas per line of business — property, liability, motor — and injects few-shot examples from validated historical extractions. A primary extraction pass produces structured JSON; a secondary verification loop re-reads source spans for high-risk fields (limits, deductibles, effective dates) and rejects out-of-range values programmatically.
Human-in-the-loop review queues flag confidence scores below threshold, conflicting endorsements, or missing mandatory fields. Approved extractions write to Insly policy objects with page-level provenance metadata. The system logs model version, prompt hash, and processing latency per document for regression testing when models or schemas change.
Deployment runs on Insly's cloud infrastructure with horizontal scaling for batch intake and synchronous API for single-document reprocessing. Integration points include broker portals, email ingestion, and the underwriting workbench where reviewers accept, correct, or reject extractions before rating.
Results
After twelve weeks of production monitoring across 4,200 commercial submissions (Q1–Q2 2025), the pipeline achieved 99.4% field-level accuracy on validated samples against human gold labels. Triage automation reached 90% of standard submissions without manual field entry. Average underwriter review time dropped from 45 minutes to 90 seconds for automated paths.
Brokers issued bindable quotes up to 10× faster on automated submissions. Annualized processing savings exceeded €500,000 in fully loaded underwriting labor, excluding revenue uplift from faster quote turnaround. Error-related rework on extracted limits fell by 73% compared to the pre-automation baseline quarter.
The pipeline became the template for subsequent Insly AI products — shared extraction schemas, evaluation harnesses, and reviewer UX patterns reused in form automation and pricing pre-fill. Lessons learned: invest early in provenance and range validation; accuracy marketing means nothing without published methodology.
Underwriters reported higher confidence in automated triage because every extracted field linked back to a highlighted source region. That auditability — not raw model size — drove adoption across broker networks that face regulatory scrutiny on rating inputs.
Key metrics
| Metric | Value | How measured |
|---|---|---|
| Extraction accuracy | 99.4% | Field-level match vs. human gold labels on 800 randomly sampled fields from 4,200 production submissions, Q1–Q2 2025. |
| Triage time | 45 min → 90 s | Median underwriter handle time per submission, automated path vs. pre-automation baseline (n=1,100 paired submissions). |
| Annual savings | €500k+ | Fully loaded labor hours reclaimed × blended underwriter rate, annualized from Q2 2025 run-rate; excludes revenue uplift. |
| Automation rate | 90% of triage | Share of standard commercial submissions completing triage without manual field entry, 12-week production window. |