Automated Claims Moderation: How AI Automates Claims Review, QA and Audit
Before Notch was an AI company, we ran an MGA. We handled claims, and we knew the routine for checking them. An adjuster made the call on each file. A manager pulled a handful of closed claims at the end of the month to see how the team was doing. Once a year an auditor arrived with a checklist and a sample, and we spent a week answering questions about files we had closed months earlier.
Each of those checks did useful work. They ran in separate teams, on separate timelines, against different readings of the same guidelines, and between them they looked at a small share of the claims we paid.
Claims moderation joins them up. You run review, QA and audit as one process, on every claim, while the claim is still open.
What is claims moderation in insurance?
Insurance claims moderation is the continuous checking of claim files for accuracy, compliance and handling quality across your whole book. It covers three activities most carriers run as separate functions: claims review, claims QA and claims audit. In a moderated operation, AI agents read every file and apply your rules, and your claims, QA and compliance teams act on the exceptions the agents surface.
Think of moderation as the umbrella, with review, QA and audit as the stages underneath it. All three ask whether a claim was handled the right way. They ask at different moments and for different audiences, which is how they drifted apart in the first place.
The three stages of insurance claims moderation
Claims review: checking decisions while the claim is live
Review happens during handling. Someone confirms coverage, checks that the documents support the loss, tests the payment against limits and deductibles, and decides whether the file needs a senior adjuster or the SIU. In most operations the handling adjuster does this alone, with a supervisor stepping in above an authority limit. Review is the one stage that can change the outcome for the policyholder, so every error you catch here is one you never have to recover later.
Claims QA: scoring how well each claim was handled
QA looks at the quality of the work. A reviewer reads a handled file and scores it against your best practice standard. Did the adjuster make contact on time, did reserves track the exposure, do the notes explain the decision, and was the letter to the policyholder clear? The scores feed coaching and training. QA tells you whether your team did good work, a separate question from whether the work met the law.
Claims audit: testing compliance with policy, contract and regulation
Audit tests the claim against rules that people outside the claims team care about. Internal audit, finance, regulators, reinsurers and carriers overseeing a TPA all want proof that payments were accurate, statutory deadlines held and authority limits were respected. Audits tend to run on closed files, on a sample, and on a calendar set by the audit plan instead of by the claim.
How does automated claims moderation work?
Automated claims moderation operates as an intelligence layer within existing claims systems. It uses OCR and language models to extract file data, validates claims against deterministic policy rules, and flags billing anomalies or fraud. Compliant claims pass through straight-through processing, while flagged exceptions escalate directly to human adjusters. When a global insurance and pensions broker ran its back office this way, 65% fewer cases needed a person.
Reading and verifying every document on the claim file
A property claim can arrive as twenty three pages of photos, police reports, estimates and invoices, and your adjuster reads all of it to find six facts. An automated agent classifies each document as it lands, extracts the fields that matter, totals the amounts across attachments and names any document still missing. OCR handles scans and handwritten notes. Language models read the narrative material, such as correspondence and adjuster notes. Each extracted fact links back to the page it came from, and that link is what lets your QA team trust everything that follows.
Checking each claim against policy rules, limits and authority levels
With the facts in place, the agent tests the claim against your rules: policy in force on the date of loss, cover for the cause, sublimits, deductibles, reserve thresholds, state deadlines and the handling adjuster’s authority. You configure those rules as deterministic logic, so the file from your newest adjuster gets the same test as the file from your most senior one. The agent records each check and its result against the claim.
Flagging fraud and leakage risk before payment
Leakage hides in line items. The agent compares invoices with the estimate and the loss description, catches duplicate bills, flags labour rates outside your agreed schedule and spots charges with no link to the reported damage. Risk models score the patterns your SIU watches for, such as a large loss a few weeks after inception.
Escalating exceptions to adjusters, QA reviewers and SIU teams
Most claims pass every check. If you allow it, the ones inside all your thresholds can move through straight through processing with no person touching them. Anything outside the rules goes to a person along with the full history, the extracted facts and the specific rule that triggered the handoff. Few carriers do this yet: WTW’s 2026 Advanced Analytics and AI Survey found just 14% of North American P&C insurers use straight through processing in claims, with another 36% planning to introduce it. Coverage disputes, denials and contested liability stay with licensed adjusters by design. You set those boundaries, and you write them as rules the agent cannot reason its way around.
Working inside your existing claims system, such as Guidewire or Duck Creek
Nobody wants to replace a core claims platform to add a QA layer. Notch connects to the claims, policy and document systems your team already uses and writes each finding back into the file. We start with one workflow, run the agent against historical claims, compare its findings with your QA team’s and expand once both sides trust the output.
What triggers a claims audit at an insurance carrier?
A claims audit is usually triggered by a calendar, a complaint or a pattern. The calendar covers the annual audit plan, a reinsurer’s review or a regulator’s market conduct exam. Complaints and litigation pull specific files into scope. Patterns are the ones you can spot yourself: a large loss soon after inception, payments sitting just under an authority limit, reserves that never moved, or a TPA whose closure times look too good to be true. Checking every file means you find these before anyone writes to you.
Why moderate every claim instead of a sample?
Sampling made sense when reading a file cost an hour of an expert’s time. Agents change that cost, and with it the case for checking a fraction of the book.
Manual sampling covers a small fraction of claims
Many traditional audit programmes review a small percent of a carrier’s claims. A sample that size tells you whether a team is on track overall. It cannot show you the documentation step one office keeps skipping, or the guideline two regions read in opposite ways, because those patterns only appear across thousands of files. Carriers tend to find them through an escalated complaint or a regulator’s exam. Our co-founder Elool Jacoby covers this gap in The QA Gap in Claims and Underwriting That Sampling Will Never Fix.
Moderating open claims before payment, not only after closure
An audit finding on a closed file becomes a recovery attempt or a training note. The same finding on an open file becomes a fix. Moderation runs at two points. During the workflow, the agent flags a missing proof of loss or a reserve nobody moved while the adjuster can still act on it. After closure, it reviews the full population of claims and surfaces the recurring patterns that should change your guidelines.
Turning moderation findings into adjuster coaching
Each moderated file gets a score against your standard. Grouped by adjuster, team or line of business, those scores show your claims managers where coaching will pay off and which guidelines people misread. Your QA team stops pulling files at random and starts deciding which patterns to act on, a far better use of people who know your guidelines inside out.
Automated vs manual claims moderation: speed, errors, staffing and audit trails
Manual moderation depends on people reading documents, and people run out of hours. The table sets the two models side by side on the measures your claims and compliance leaders track.
Key benefits of automated claims moderation
The gains show up in four places your leadership will measure.
Faster, more consistent claim decisions
Your adjuster opens a file that has been read, extracted and checked, so they start at the decision point. Simple claims that meet your thresholds can close in minutes. On complex claims, two adjusters with the same facts get the same rule checks, which closes the gap between how your best handler and your newest handler work a file. That speed is rare today: nearly two thirds of North American carriers told Gallagher Bassett their claims are not settling any faster than a year ago.
Reduced claims leakage and fraud
You catch overpayments before the money leaves the account, where an audit would find them months later. Duplicate invoices, out of schedule labour rates and unrelated charges get flagged on day one, and your SIU team sees risk scores with the reasons attached.
Lower QA and audit costs
Your reviewers stop spending their week reading sampled files and spend it on the exceptions. One Notch customer, a global insurance and pensions broker, cut manual reconciliation by 75%, so claims volume can grow without QA headcount growing alongside it.
Audit ready records for regulators
Each extraction, check and escalation carries a source reference and a timestamp. A market conduct examiner who asks why you paid a claim can see the documents and rules behind the decision in one view.
Keeping humans in control: AI governance in claims moderation
The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023, and by October 2026 27 states plus Washington, D.C. and Puerto Rico had adopted it.. It expects a written AI programme, documented governance and oversight of the vendors you use. The newer NAIC work on AI risk evaluation pushes traceability and human oversight further, which Elool covers in our NAIC AI pilot update.
In practice, governance comes down to a few habits. Tie each finding to a policy clause, a source document and a timestamp. Keep licensed people on denials, disputes and final audit sign off. Keep the logic that checks claims separate from the logic that handles them, so one model’s blind spots cannot repeat on both sides unnoticed. Compare the agent’s findings with your human reviewers’ on a regular schedule. Notch holds SOC 2 Type II, ISO 27001 and ISO 42001 certification, and you can deploy it inside your own cloud tenant. Phoenix, one of Israel’s largest insurance groups, runs Notch inside its own cloud after clearing its full security, regulatory and legal review.
How to measure claims moderation performance
Set baselines from your current programme before you switch anything on, then review the numbers monthly with claims and compliance leadership. Start with audit coverage, the share of open and closed claims checked against your full criteria, since moving from a small sample toward full coverage is the clearest sign the programme is running. Track quality scores by adjuster and team, and watch the spread as well as the average. A narrowing spread tells you coaching is landing.
Split leakage into amounts stopped before payment and amounts recovered after closure, and expect the first number to grow as the second shrinks. Normalise compliance exceptions per hundred claims so you can compare regions and TPAs on equal terms. Cycle time and your straight through processing rate round out the picture, since moderation should speed simple claims up and keep complex ones moving.
Where does Notch fit in?
We built Notch after years of doing this work by hand as an MGA, so we designed moderation around how claims teams operate. Our agents work across the claims lifecycle: they take first notice of loss, read and check every file, apply your rules and leave a record of what they checked and why. Across production deployments, Notch agents complete 70% to 73% of eligible interactions without a person, with a six times faster median time to resolution than a fully human team. ADAM, our operating layer, coordinates those agents and studies escalations to find the next rule worth tightening.
The platform has handled more than 20 million interactions in production for insurers, banks and other regulated businesses. You start with one claims workflow, prove the findings against your own QA team and expand from there.
If you want to see claims moderation running on your own files, book a demo and we will map your review, QA and audit work to agents your team can supervise from week one.