AI claims triage: what to automate, what to route, and what regulators now expect

At first notice of loss, a model reads the description, photos and policy, estimates complexity and severity and routes the claim: fast-track settlement for simple, covered losses; an experienced adjuster for injuries or large losses; the special investigations unit when fraud indicators stack up. The value is in getting each claim to the right handler on day one.

The Coalition Against Insurance Fraud estimates fraud costs $308.6 billion a year in the US ^[1]. Models can find patterns people miss, such as the same phone number, repair shop or photo appearing across unrelated claims. A flag should send a claim for investigation; it should never deny a claim on its own.

The NAIC's model bulletin on insurers' use of AI, adopted in December 2023 and taken up by many states, expects a written AI systems programme covering governance, risk management, testing for unfair discrimination and oversight of third-party models and data ^[2]. Colorado's SB21-169 goes further, prohibiting unfairly discriminatory use of external consumer data and algorithms and requiring insurers to test for it ^[3].

• Keep an inventory of every model used in claims, with its owner and purpose.
• Test outcomes for unfair discrimination before launch and on a schedule after.
• Record which model version and inputs led to each routing decision.
• Get the same documentation from vendors whose models you use.

Triage only helps if it writes back into the claims system and the adjuster's queue. Industry data standards such as ACORD's help map data between systems ^[4]. Plan for this from the start, not after the model is built.

Found this useful?

Share the key finding on X.