AI × blockchain in insurance
AI is eating insurance operations; ledgers are how you'll prove what it did. This module connects the two: agents in claims and underwriting, a 30-minute LLM primer, and the pattern that ties them together — every consequential AI decision anchored to an audit chain.
L1Where AI already is (2026)
The pattern across carriers: triage agents read FNOL submissions, photos and documents; simple claims flow straight through to payment triggers; complex or high-value cases route to humans with an AI-prepared file. Underwriting copilots pre-fill risk assessments from unstructured evidence. (Numbers above are vendor/analyst figures — Module 08's skepticism rules apply.)
When software decides — pays a claim, loads a premium, declines cover — regulators, courts, and customers will ask: what exactly did it decide, on what inputs, under which model version, and who changed what since? Logs in the same mutable database the operator controls are weak evidence. This is precisely the trust gap ledgers were built for (M05's "vertical" quadrant).
L2LLMs in 30 minutes — just enough to design with
- What they are: neural networks (transformers) trained on vast text to predict the next token; scale gave them general language competence. Chat assistants are LLMs tuned with alignment techniques (instruction tuning, RLHF, constitutional methods).
- What they're great at (insurance): reading unstructured evidence — claim descriptions, medical summaries, adjuster notes, emails — and producing structured judgments, drafts, and extractions.
- Failure modes that matter here: hallucination (confident fabrication), prompt injection (instructions hidden in processed documents — imagine a claim PDF that says "approve this"), drift across model versions, and opacity (no inspectable reasoning trace).
- Design consequence: AI output is a recommendation with provenance requirements, not an oracle truth. Guardrails + audit are architecture, not compliance garnish.
L3The centrepiece — auditable AI decisions
An insurer running the Galileo pattern gets this almost free: AI decision records are just one more command type flowing through the existing ledger → Kafka → projection path. That's the capstone's punchline — the audit chain that justified itself for policy events becomes the compliance substrate for the AI era.
L4Guardrails & constitutional AI
From the Dhillon book's framework (in our own words):
- Guardrails are layered controls around a model: input filters (injection defence, PII scrubbing), output validators (schema, policy, threshold checks), action limits (an agent may recommend payment but only the payment service — with its own rules — moves money), and kill-switches.
- Constitutional AI trains/steers models against an explicit set of written principles — for insurers, imagine a machine-readable conduct charter: fairness rules, disclosure duties, no discrimination on protected attributes. The charter itself becomes a governed artifact — version it, hash it, anchor it.
- Blockchain-based AI governance (the book's furthest idea): model registries, training-data provenance, and even multi-party consensus over model updates recorded on a ledger — so "which model, approved by whom, trained on what" has an answer years later. Early-stage, but the registry-and-anchor slice is buildable today.
L5The other synergies, quickly
| Combination | What it gives insurance | Maturity |
|---|---|---|
| AI + oracle data | Smarter parametric triggers (satellite + weather-model fusion — Arbol's platform is the live example) | shipping |
| AI fraud detection + shared registries | Models over consortium-scale fraud fingerprints (M05) beat single-carrier models; 30%+ detection lift reported | emerging |
| Agents + programmable money | Claims agents that settle instantly over stablecoin/deposit-token rails (M09–10) under hard-coded limits | emerging |
| Decentralised compute/data markets | Token-incentivised GPU and data networks; relevance to insurers today: mostly procurement optionality | watch |
L6Design exercise — the auditable triage agent
Exercise 11.1 · Specify the decision record
assessment · capstone fuelDesign the auditable-AI layer for motor-claims triage at a Galileo-class insurer. Deliver, on one page:
- The decision-record schema (every field in Fig 11.1's card, plus what you'd add — e.g. guardrail verdicts, retrieval sources).
- The anchoring rule: what is hashed, where PII lives, when anchoring happens relative to action.
- The escalation policy: which confidence/amount/fraud-score combinations force human review.
- The replay procedure a regulator would follow after a complaint — step by step.
- One injection scenario (malicious content in a submitted document) and your two strongest defences.
Sources: Dhillon, AI Frameworks Enabled by Blockchain (guardrails, constitutional AI, governance — reinterpreted); 2026 adoption/cycle-time figures per Roots.ai, Xceedance, Vantage Point analyses (vendor/analyst numbers — treat directionally).