AI Claims Processing: How P&C Carriers Cut Adjudication Costs

Carriers lose $0.10–$0.15 of every claim dollar to adjudication. Here's how production AI agents read FNOL, photos, and reports to draft settlements — and where humans stay in the loop. This is the practical case for AI insurance claims processing: not a chatbot bolted onto a portal, but an agent that does the file work an adjuster spends their day on.

The hidden cost in every claim dollar

Loss-adjustment expense — the cost of investigating, evaluating, and settling claims — runs roughly 10 to 15 cents on every claim dollar a P&C or auto carrier pays out. It is one of the largest controllable line items in the business, and it has barely moved in two decades. Premium leakage gets the headlines. LAE is the quiet drag that compounds across millions of files.

The reason it stays high is structural. Each claim arrives as a pile of unstructured documents from different sources, in different formats, with no shared schema. A first notice of loss from a call-center transcript. Photos from a policyholder's phone. A repair estimate as a PDF from one shop and a faxed version from another. A police report written in narrative prose. Medical bills. Recorded statements. Nothing lines up, and someone has to read all of it before a dollar moves.

That someone is an adjuster, and the bulk of an adjuster's day is not judgment — it is retrieval and transcription. Industry time studies consistently show that a large share of claim-handling hours goes to reading documents, re-keying figures into the claims system, cross-checking the policy, and chasing missing paperwork. The actual decisions — is this covered, what is it worth, is anything off — take minutes. The work that surrounds those decisions takes hours.

The expensive part of a claim was never the decision. It was everything an adjuster had to read before they were allowed to make it.

This is why headcount and offshoring have stopped bearing fruit. You can hire more adjusters or move the data entry overseas, but you are still paying a human to read a police report. The cost floor is set by the document work, not by wages. The only way through the floor is to automate the reading itself — which is exactly what modern AI claims processing automation now does reliably.

From FNOL to settlement: what AI automates

Start at intake. An AI agent ingests the first notice of loss the moment it lands — whether it arrives as a web form, a call transcript, an email, or an EDI feed. FNOL automation means the agent extracts the structured facts (loss date, location, parties, vehicle or property, cause of loss) and flags what is missing before the file ever reaches a queue. The policyholder who forgot to attach photos gets the request automatically, not three days later when an adjuster finally opens the file.

Then the agent reads the evidence. It interprets damage photos, parses the line items on a repair estimate, pulls the relevant facts out of a narrative police report, and reconciles a medical bill against treatment codes. Where a human reads each document once and holds the rest in their head, the agent reads everything and keeps every figure cross-referenced. The output is not a summary for a person to re-verify — it is structured data the downstream steps can act on.

Coverage verification is where the agent earns its keep. It maps the loss facts against the actual policy in force on the loss date: limits, deductibles, endorsements, exclusions, sublimits. It checks whether the named driver was covered, whether the peril is included, whether a prior endorsement changes the math. This is the step most prone to human error and inconsistency, because it requires holding the entire policy document in mind while reading the claim. The agent does it the same way every time.

Finally, damage assessment and settlement drafting. The agent compares the estimate against benchmark repair costs, applies the deductible, accounts for depreciation or betterment where the policy calls for it, and produces a recommended settlement with the reserve figure. Critically, it produces a draft — a settlement an adjuster reviews and approves, with every number traceable to the document it came from. The carrier's AI claims processing agent does the file work; the adjuster does the deciding.

• FNOL intake: extract structured facts from any channel, request missing items automatically.
• Evidence reading: photos, estimates, police reports, medical bills — parsed and cross-referenced.
• Coverage check: loss facts mapped against limits, deductibles, endorsements, and exclusions.
• Settlement draft: a recommended payout with reserves, every figure cited to its source document.

Straight-through processing vs. the exception queue

Not every claim should be automated end to end, and pretending otherwise is how automation programs lose the room. The right design splits the book into two streams: claims that can run straight through, and claims that route to a human with the agent's work already done.

Straight-through processing belongs to the routine, high-volume claims where the facts are clean and the math is unambiguous — a single-vehicle auto glass claim, a low-severity property loss with clear photos and an in-network estimate, a total that falls well inside limits with no liability dispute. For these, the agent can verify coverage, price the loss, and issue payment within guardrails the carrier sets, with sampling for quality control. This is where P&C claims automation pays for itself: the simple claims that should never have consumed an adjuster's hour stop doing so.

Everything else routes to the exception queue — and this is a feature, not a fallback. Complex liability, suspected fraud, coverage that is genuinely ambiguous, large losses, anything contested or litigated: these go to an adjuster. But they arrive pre-digested. The agent has already read the file, verified coverage, summarized the evidence, and flagged the specific reason it could not proceed. The adjuster opens a file that is ready to decide, not a pile to assemble.

The routing logic itself is where carriers should invest scrutiny. Auto claims AI should be conservative about what it auto-processes and generous about what it escalates. A claim near a policy limit, a new claimant with thin history, a damage pattern that does not match the reported cause, an estimate that diverges from benchmark — any of these should pull a file out of straight-through and into human review. The cost of a wrongly automated claim is far higher than the cost of an unnecessary escalation, and the thresholds should reflect that asymmetry.

Compliance, fairness, and auditability

Insurance is one of the most heavily regulated industries an AI system can touch. Unfair claims settlement practices acts, state fair-claims requirements, and emerging model-governance rules all assume that a claim decision can be explained and defended. An automation program that cannot produce that explanation is not a cost saving — it is a liability waiting for a market-conduct exam.

The non-negotiable design principle is that every decision is documented and cited. When the agent recommends a settlement, it should be able to show which policy clause it relied on, which document each figure came from, and what rule led to the outcome. Not a confidence score — an audit trail. A regulator, an internal auditor, or a policyholder's attorney should be able to trace any number back to its source. This is what separates production AI claims processing from a black box that happens to be fast.

Fairness requires the same discipline applied before deployment and continuously after. Carriers need an evaluation framework: a held-out set of historical claims with known-good outcomes, run against the agent on every model or prompt change, measuring accuracy, consistency, and disparate impact across protected classes. Drift is real, and a system that was fair at launch can degrade silently. The evaluation is not a one-time certification; it is a standing test suite the carrier owns.

Human-in-the-loop is the through-line. The agent drafts; a person with authority approves. The escalation paths are explicit. The override is logged. The point is not to keep a human in the way of every claim — it is to keep accountable judgment attached to the decisions that carry legal and reputational weight, while the agent absorbs the volume that does not.

What carriers should pilot first

The failure mode for claims automation is scope. A program that tries to automate the entire book on day one collides with every edge case in the portfolio and stalls. The programs that succeed pick one thing and prove it.

Pick one high-volume, low-complexity claim type. Auto glass. Towing and roadside. Low-severity property losses under a fixed threshold. Choose the segment where the documents are predictable, the coverage questions are simple, and the monthly volume is large enough that a percentage improvement is a real number. A narrow wedge with high volume gives you fast feedback and a clean ROI story; a broad, ambiguous segment gives you neither.

Then define the ROI metric before you build anything. Decide in advance what success means: cost per claim, cycle time from FNOL to settlement, straight-through rate, adjuster hours redeployed, leakage reduction. Baseline it on the current process so the comparison is honest. If you cannot state the number you are trying to move, you are not ready to build — you are ready to do the measurement that comes first.

From there the path is incremental: ship the narrow wedge, measure against the baseline, widen the straight-through criteria as confidence grows, and add claim types one at a time. Each expansion is a decision backed by the evaluation framework, not a leap of faith. If you want to see how this maps to a fixed-scope engagement, the engagement model we use starts with exactly this kind of bounded, measurable first build.