# AI × Insurance — original notes

Short, cite-ready essays on what actually ships in regulated insurance AI — not hype decks. Each note links to full case studies or services with metrics.

## AI Audit vs AI Discovery Sprint: What your business actually needs

Leaders often buy the wrong first AI engagement. An AI audit inventories risk, data readiness, and governance gaps; a discovery sprint designs and partially validates one production-shaped use case. Choosing between them is a decision about evidence: do you need a map of the landscape, or a first working path through it? In insurance and other regulated domains, the wrong answer wastes a quarter and erodes underwriter trust.

See how engagements are structured →
## Custom AI builder vs general agency: When production systems beat retainers

Companies comparing a custom AI builder to a general digital agency are usually asking who will own accuracy, latency, and governance after the kickoff workshop. Agencies excel at campaigns, content, and multi-channel delivery. Builder-consultants excel when the product is a system underwriters and actuaries must trust daily — extraction pipelines, pricing APIs, agent tool layers — with published metrics and methodology notes.

Read the underwriting case study →
## Why underwriting AI fails without provenance

Brokers submit messy PDFs; underwriters need field-level trust. Production IDP pipelines must return coordinates to source pages, human-in-the-loop for edge cases, and accuracy measured against gold labels — not demo F1 scores on clean samples. At Insly we held launch until 99.4% field match on 800 sampled fields across 4,200 submissions.

Read the underwriting case study →
## Sub-100ms pricing is a governance problem

Fast ML-augmented rating only works when actuaries own bounds: shadow mode, circuit breakers, and explainable mappings from model outputs to approved grids. Speed without governance increases referral noise; with governance, auto-binding can jump from 22% to 68% while loss ratios stay bounded.

Read the pricing case study →
## MCP as the integration layer for public data

AI assistants need live facts, not stale training data. A hosted MCP server with annotated tools lets Claude or Cursor fetch Estonian electricity prices, company filings, or parliament votes in one step — no API keys, no copy-paste. Open source the server; operate the endpoint for reliability.

Explore the MCP server →