Underwriting and claims decisions a regulator can challenge.
When AI prices a policy or denies a claim, someone eventually asks why. The answer has to already exist in the system — reconstructable, auditable, and evaluated before it shipped, not assembled after the fact.
What this looks like inside an insurance technology engineering org.
These are the predictable result of underwriting and claims AI shipping faster than the explainability and audit infrastructure around it.
- Underwriting or claims-triage model decisions aren't reconstructable after the fact.
- A pricing or risk model updates, and the change wasn't evaluated against a held-out regression set first.
- An AI claims-processing agent has broader data access than the specific claim it's working actually requires.
- State regulators or auditors ask for a model inventory and risk classification that doesn't exist yet.
- Adverse-action or denial explanations generated by AI aren't consistent enough to stand up to a complaint review.
- No approval workflow exists between a model change and it going live in claims or underwriting.
Where in the model's life this shows up.
The risk isn't constant — it concentrates at specific points between a prototype and a regulator's request.
01 · Prototype
Pricing or risk model in testing
Explainability gets deferred — easy to skip when there's no policyholder yet to challenge a decision.
02 · Shipped to production
Real policies, real claims
A model update ships without clearing a held-out regression set first, and the next denial can't be defended.
03 · Scaling
Claims-processing agents with account-wide access
Access scoped to the account, not the specific claim — broader than any single task actually requires.
04 · Under audit
A state regulator asks for the model inventory
Most insurtechs don't have one until asked — and more than 20 states have already adopted the NAIC AI Model Bulletin requiring it.
Where this becomes engineering work.
Four disciplines cover most of what shows up in an insurance technology AI stack. Not every account needs all four.
AI Governance & Control
The deliverable: a model inventory and risk classification an auditor can read, built before the request lands.
AI Reliability Engineering
The deliverable: a regression eval harness for underwriting or pricing model changes, run before release.
AI Security Engineering
The deliverable: claim-scoped access controls so a claims agent can't see more account data than its task requires.
AI Systems Engineering
The deliverable: underwriting and claims-triage decisions built to be reconstructable, not explained after the fact.
Not sure where the inventory gap is? Run the AI Technical Debt Calculator.
Companies typically bring Crescent in when:
Capacity
You need specialist AI engineering capacity without building another team — project-based capacity, not staff augmentation.
Expertise
The project has crossed into an area where your existing team lacks specialized depth.
Speed
A production deadline is approaching and the internal path is too slow.
Critical project
Your core engineering team can't afford to divert months of capacity.
Independent review
You need an external technical second opinion before making an expensive decision.
Common questions.
Make the model's decisions defensible before you're asked.
Tell us where the gap is — an unreconstructable decision, an agent with too much account access, no evaluation gate before a model ships. We'll tell you which discipline it maps to and whether it's a fit.
NDA available on request · Scoped engagements · No surprise fees
Not sure where the gap is? Run the AI Readiness Score — five minutes, no call required.