10decoders
The Pilot Worked. Then Someone Asked Who Approves the Denial
Picture a mid-sized property and casualty carrier that pilots an agent on auto glass and minor property claims. The agent reads the first notice of loss, checks the policy, extracts the repair estimate, and drafts a settlement. Cycle time drops, the adjusters like it, and the steering group approves a wider rollout. Then a compliance lead asks a plain question: if the agent denies a claim and the policyholder complains, who made that decision, and on what evidence? Nobody has a clean answer. This is an illustrative scenario, but it is the point where many claims pilots stall.
The stall is rarely about model quality. Deloitte's survey ranked lack of business-line support as the top reason gen AI implementations fail, ahead of funding, with poor data foundations, legacy IT, and weak collaboration between business and technology teams close behind. Respondents also rated themselves least prepared on talent and workforce skills. Those are ownership problems. A claims agent touches coverage, payment, and customer communication at once, and each of those belongs to a different leader.
McKinsey offers a useful budgeting rule from its work with more than 200 insurers: for every dollar spent on AI development, plan at least another dollar for adoption and scaling. Change management, it says, is about half the total effort. In claims, much of that second dollar buys decision rights: who can approve what, which actions need a licensed adjuster, and how a bad outcome gets reversed. A carrier that has not paid that bill has a demo, not a production system.
"A claims agent is not production-ready when it is accurate. It is production-ready when every decision it touches has an owner."
Six Claims Steps Where Agents Fail Quietly
These are the points in the claim lifecycle we examine first. The severity column reflects how much harm a wrong output does before anyone notices.
| Claims step | What breaks | Severity |
|---|---|---|
| First notice of loss intake | The agent captures facts the claimant never stated, such as a guessed loss date, and the guess becomes the record | Moderate |
| Coverage verification | The agent reads the wrong policy version or misses an endorsement, so a covered loss is flagged as excluded or the reverse | Critical |
| Document extraction | Photos, estimates, and medical bills arrive in mixed formats, and a misread amount flows into reserves without a flag | High |
| Fraud flagging and SIU referral | Scores treat unusual as suspicious, so honest claimants are slowed and the special investigations queue fills with noise | High |
| Settlement recommendation | The model anchors on past payouts, so earlier inconsistencies are copied into every new file at speed | Critical |
| Denial and claimant communication | A fluent, confident letter states a wrong reason, and the claimant has no obvious route to a human | Critical |
Not sure where your claims AI decision gaps are?
10decoders builds and tests AI agents for regulated workflows, including document-heavy processes like claims. We start by mapping each agent action to an owner, an evidence trail, and a fallback, so you can see what stands between your pilot and production.
Book a Free AI Assessment →Why Good Averages Hide Bad Claims
Aviva's reported results show what is possible. McKinsey cites more than 80 AI models in its claims operation, a 23-day reduction in liability assessment time for complex cases, a 30% gain in routing accuracy, and a 65% drop in complaints. Notice what that list describes: many specific models, each pointed at a narrow task such as routing or assessment. It does not describe one general agent given authority over the whole claim.
Averages are also where claims risk hides. An agent that is right on 95% of glass claims can still be wrong in a pattern, for example on older policy forms, on a particular repair network, or on claimants who write in a second language. Because the agent is fast and fluent, wrong outputs look the same as right ones. Without segment-level monitoring, the first sign of a pattern is a complaint, a regulator inquiry, or a lawsuit, and by then hundreds of files carry the same flaw.
The practical answer is to treat each claims action as its own product with its own limits. Reading a document, proposing a reserve, paying a small claim, and refusing a benefit carry different consequences and need different levels of oversight. Writing those levels down is dull work, and it is the work that lets a carrier widen autonomy over time with evidence instead of hope.
The Agent Summarizes and the Adjuster Decides
Agents extract documents and draft notes. Every decision stays with a person, and nobody measures whether the drafts are accurate or only fast.
The Agent Routes and Recommends Under Review
Agents sort claims, suggest coverage and settlement, and auto-handle a narrow low-value band. Review is manual, so it becomes the bottleneck and the weak point.
Autonomy Levels, Evidence Trails, and Live Monitoring
Each action has an owner and a cap, every recommendation is traceable, adverse actions get human sign-off, and monitoring widens or narrows autonomy using data.
Checklist: What a Production-Ready Claims Agent Can Show
Pick one claim type and ask whether the carrier can demonstrate each item from configuration and logs, not from assurances.
"Widen the agent's authority one claim type at a time, and let the monitoring data, not the roadmap, decide the pace."
What to do this week
01 Inventory every action the pilot agent can take
List each action as a verb: read, classify, extract, recommend coverage, propose reserve, pay, deny, send letter. For each one, write who owns the outcome today and whether the agent suggests, acts with review, or acts alone. Expect to find at least one action running at a higher autonomy level than anyone intended. Bring the list to the claims, compliance, and IT leads in the same room.
02 Set dollar and line-of-business caps in configuration
Choose a payment ceiling for the narrowest, best-understood claim type and enforce it in the workflow engine, not in the prompt. Route bodily injury, litigated files, and anything touching a regulated line to a licensed adjuster by rule. Document the cap, its approver, and the review date so the next expansion starts from a record instead of a conversation.
03 Require evidence links on every recommendation
For ten recent claims, ask the adjuster to trace a recommendation back to the policy clause, the document, and the data field that drove it. Count how many take more than two minutes or cannot be traced at all. Those are the gaps a regulator will find first. Fix the highest-volume gap before the agent sees more claims.
04 Review adverse actions and sample outcomes by segment
Put a human sign-off on every denial and partial payment letter, and sample a fixed number of agent-handled files each week for a senior adjuster to re-decide blind. Compare results by policy form, region, and claimant language. Record disagreements with reasons. A pattern in the disagreements is your earliest drift signal, and it costs far less than a regulatory finding.
Let 10decoders Stress-Test Your Claims AI Before It Scales
We map each agent action to an owner, an autonomy level, an evidence trail, and a fallback, replay a sample of closed claims to find where recommendations diverge from adjuster decisions, and draft the monitoring thresholds with your claims and compliance leads. You leave with a ranked gap list and a thirty-day plan.