Insights topic
Insurance AI operations
Insurance work is decisions, not tasks. These pieces cover what changes when agents take part in claims, underwriting, and servicing.
Claims automation fails when it treats a decision like a task
A claim is a decision, not a document to be routed. Claims automation projects that focus on extracting fields, triaging intake, and speeding up the paperwork usually leave the actual judgment — is this covered, is this valuation reasonable, is this pattern fraud — exactly where it was, made by an adjuster with less time because the easy parts got automated and the hard ones got concentrated. The honest framing is to ask which decisions in the claims lifecycle can tolerate a probabilistic recommendation with human sign-off, and which need to stay fully human, rather than automating the workflow end to end and hoping the judgment comes along for free.
Nobody priced the judgment that underwriting actually depends on
Underwriting automation tends to get built by modeling the inputs an underwriter reviews, not the judgment they apply when the inputs conflict — the submission that looks clean but doesn't sit right, the risk that's technically in appetite but isn't one you want. Nobody prices that judgment into the automation's scope, so the system performs well on the straightforward 80% of submissions and silently degrades underwriting discipline on the 20% that actually determines loss ratio. Good AI underwriting programs make that judgment explicit: what triggers referral to a human, what the model is and isn't allowed to bind, and how referral thresholds get revisited as the book changes.
Half a job automated is a new failure mode, not a win
It's common to automate half a job — the data entry, the status update, the first-pass summary — and count it as a claims automation success, while the part that actually required expertise stays manual and now has less context because the automated half stripped out the nuance a human would have noticed. Brittleness was often a feature of the old process, forcing exceptions to surface to a person; smoothing it away with automation can hide exceptions that should have been escalated. Insurance AI operations that hold up are designed around where accountability sits, not just where the labor is cheapest to remove, because a wrong claims decision or a mispriced policy costs far more than the labor it displaced.
A claim is a decision, not a task
Claims is rightly called the highest-leverage workflow in insurance to automate — straight-through processing, lower costs, faster settlement. But the framing treats a claim as an administrative hurdle, and a claim is not a task. It’s a decision about money, made under uncertainty, with two opposite ways to be wrong — and the scoreboard that justifies the automation can only see one of them.
Read
Insurance assumes a person
Putting agents into underwriting, claims and eligibility is where the money is, and the firms moving on it are reading their own economics correctly. But every instrument that sits underneath those functions was built around an assumption nobody wrote down — that a person decided. Remove the person and the machinery doesn't break. It just stops having a subject.
Read
You automated half a job
The roles being automated genuinely are automatable and the savings are genuinely real — we help clients capture them. But those desks were doing two jobs, and only one of them was ever written down. The second was a continuous, unmeasured correction of everything that arrived slightly wrong, and when the role goes, that work doesn't transfer anywhere. It stops.
Read
Brittle was a feature
The migration off brittle rule-based RPA onto AI automation that learns and self-heals is the right move, and we don’t dispute the upside. But in the migrations we watch, the thing that bites teams isn’t the part that failed to work. It’s the part that kept working when it shouldn’t have.
Read