AI Claims Processing: A Practical Guide for Insurers

2026-10-21 · Tommaso Maria Ricci

In July 2025 Allianz switched on a system in Australia that settles small food spoilage claims in about a day instead of several. Seven AI agents check coverage, weather, fraud signals and payout, and then a human claims professional makes the call. That is what AI claims processing looks like when it works: not a robot adjuster, but a pipeline that removes the waiting time between steps and leaves the judgment where it belongs.

Most insurers and claims operations are nowhere near that. They have pilots, a document extraction tool, maybe a fraud score that nobody fully trusts, and a claims backlog that grows every time a storm hits. The gap is not technology. It is sequencing, data and governance.

This guide is for the people who own that gap: heads of claims, COOs at carriers and MGAs, TPA leaders and the CFOs who sign off the budget. It covers where AI actually helps in the claims lifecycle, where it creates legal exposure, how to measure it, and a 90 day plan to move from pilot to production without becoming the next lawsuit headline.

What AI claims processing actually means in 2026

The phrase gets used for five very different things, and mixing them up is the first reason projects stall. When I review a claims AI plan, I sort every proposed use case into one of these layers before anything else.

1. Intake and triage. Reading first notice of loss, emails, photos, PDFs and call transcripts, extracting the facts, and routing the claim to the right queue.
2. Coverage and liability support. Checking the policy, the dates, the exclusions and the likely liability split, then presenting a recommendation to an adjuster.
3. Fraud and leakage detection. Scoring claims for anomalies, duplicate billing, staged losses or inflated repair estimates.
4. Settlement and payment. Calculating payouts on simple, rules-bound claims and preparing the payment for approval.
5. Communication. Keeping the claimant informed, answering status questions and drafting letters.

Each layer has a different risk profile. Intake errors cost time. Coverage and denial errors cost lawsuits. Treating the whole thing as one "AI claims project" is how companies end up with a fast intake bot sitting in front of the same slow adjudication process, which moves the bottleneck instead of removing it.

Automation versus augmentation

The second distinction that matters: is the AI deciding or recommending? For low-value, high-frequency claims with clear rules, straight-through processing is realistic. For anything involving judgment, bodily injury, large losses or ambiguity in the policy wording, the system should prepare the file and the human should decide.

Allianz's design is explicit on this point. In its own description of Project Nemo, the company states that payout decisions are never automated by design, and the audit agent passes its summary to a person who confirms the outcome. You can read the Allianz account of the system for the full architecture. That is not timidity. It is a design choice that keeps accountability clear, and Allianz still put the system live in under 100 days.

Why claims is where insurance AI pays back first

Claims is the moment the insurance promise is tested, and it is also the largest controllable cost line for most carriers. That combination makes it the natural first target.

The adoption data from US regulators shows how far the industry has already moved. The NAIC's 2025 health insurance survey found that 86 of 93 responding companies, or 92%, currently use, plan to use or plan to explore AI or machine learning. In the breakdown of models in production by operational area, fraud detection appeared at 70% and claims adjudication at 62%. Earlier NAIC surveys put the equivalent figure at 88% for private passenger auto insurers, 70% for home and 58% for life. The NAIC health survey memo has the detail, including the caveat that the health survey counts GLMs and GAMs as AI, which the auto and home surveys did not.

So "should we use AI in claims" is no longer the question. The question is whether your use is producing measurable results or just producing pilots.

The results that are on the record

There are few published, verifiable numbers in this space, which tells you something about how many projects are still unproven. The ones that exist are instructive.

• Aviva (UK). According to a McKinsey case study on the program, Aviva built more than 80 AI models across the claims journey, cut the average time to assess liability on complex motor claims by 23 days, improved routing accuracy by 30% and reduced customer complaints by 65%. UK trade press reported in March 2026 that AI in motor claims saved Aviva more than £60 million in 2024.
• Allianz (Australia). Project Nemo reports an 80% reduction in processing and settlement time on food spoilage claims under AUD 500, with steps before human review completing in minutes.

Notice what both have in common. Neither started with a general-purpose chatbot. Both started with a specific, high-volume claim type, rebuilt the workflow around it, and kept humans on the decisions that carry liability.

The fraud number everyone quotes, and why to be careful

You will see the figure that insurance fraud costs the US economy $308.6 billion a year. It comes from a 2022 Coalition Against Insurance Fraud report. It is also contested: a 2024 analysis from Rutgers Law School argues the methodology relies on property and casualty fraud percentages that are decades old.

I mention this because claims AI business cases are often built on headline fraud numbers that the finance committee will eventually challenge. Build yours on your own leakage data. If you cannot measure your current fraud and leakage rate, that is your first project, not an AI one.

How to use AI in insurance claims processing, stage by stage

Here is how I map AI onto the claims lifecycle, with the realistic use, the prerequisite and the risk at each stage. This is the core of the work, so it is worth going slowly.

Stage 1: First notice of loss

What AI does well: reading unstructured intake from every channel, whether web form, email, phone transcript or photos, and turning it into a structured claim record. Language models are now good at extracting dates, locations, parties, vehicle details and damage descriptions from messy text.

Prerequisite: a target data model. If your claims system does not have clean fields for what you want to extract, the AI produces structured output with nowhere to go.

Risk: low to moderate. Errors here are usually caught downstream, but they add rework. Measure extraction accuracy on a sample every week during rollout.

Metric to watch: time from first contact to a complete, assignable claim file.

Stage 2: Triage and routing

What AI does well: predicting complexity, severity and the right handler. This is where Aviva reported its 30% routing accuracy gain, and it is often the highest-return step because a misrouted claim loses days before anyone works on it.

Prerequisite: historical claims with outcomes, so a model can learn which features predict a complex claim. Three years of closed claims is a reasonable starting point for a mid-size book.

Risk: moderate. Bad routing does not deny anyone, but it can create unequal service levels between customer groups if the model learns proxies for protected characteristics. Test for that.

Metric to watch: reassignment rate and cycle time by segment.

Stage 3: Coverage verification

What AI does well: matching the claim to the policy, endorsements and exclusions, and flagging the clauses that apply. This is document comparison work, and retrieval over your policy wordings handles much of it.

Prerequisite: a clean, versioned library of policy wordings tied to each policy number. Many carriers discover here that they cannot reliably say which wording applied on the date of loss.

Risk: high if automated, moderate if presented as a recommendation. A coverage denial driven by an AI misreading is exactly the kind of decision that generates bad faith claims.

Metric to watch: adjuster override rate on AI coverage recommendations. If it is near zero, check whether adjusters are actually reading them.

Stage 4: Damage assessment and reserving

What AI does well: estimating repair costs from photos for standard auto and property damage, comparing estimates against historical repair data and suggesting initial reserves.

Prerequisite: labeled image data and repair outcomes, or a vendor that has them. This is a stage where buying usually beats building.

Risk: moderate. Photo-based estimates are reliable on common damage and weak on edge cases. Set a confidence threshold below which the claim goes to a human appraiser automatically.

Metric to watch: variance between AI estimate and final paid amount, by damage type.

Stage 5: Fraud and leakage

What AI does well: network analysis across claims, providers and repair shops, anomaly detection on billing patterns, and duplicate detection. Machine learning has been doing this longer than generative AI has existed.

Prerequisite: a feedback loop. A fraud model that never learns which referrals were confirmed degrades quickly.

Risk: high for customers. A fraud flag that delays a legitimate claim is a customer experience failure, and a flag correlated with neighborhood or ethnicity is a regulatory one.

Metric to watch: referral hit rate, meaning the share of AI fraud referrals that investigators confirm.

Stage 6: Settlement

What AI does well: calculating and preparing payment for simple claims where coverage, liability and amount are clear and within limits.

Prerequisite: explicit rules for what qualifies as straight-through. Value cap, claim type, no injury, no fraud flag, customer tenure, whatever your risk appetite allows.

Risk: varies with the cap. Start low. Allianz started with claims under AUD 500.

Metric to watch: share of claims settled straight through, and the error rate found in post-payment audit.

Stage 7: Customer communication

What AI does well: proactive status updates, answering "where is my claim" questions and drafting letters for adjuster review.

Evidence from adjacent work is encouraging. An NBER study of customer support agents found a 14% average productivity gain from an AI assistant, with the largest gains among less experienced staff. Claims handlers answering status questions are doing similar work.

Risk: moderate. A chatbot that promises a payment or misstates coverage can bind the company. The Air Canada chatbot case in 2024 established that a company is responsible for what its bot tells customers.

Metric to watch: inbound status calls per open claim.

Where AI claims processing goes wrong: the lawsuits to learn from

The two cases every claims leader should know are both from US health insurance, and both turn on the same issue: the allegation that a machine effectively made decisions that a qualified human was supposed to make.

Cigna and PXDX. In March 2023 ProPublica reported that Cigna medical directors denied more than 300,000 claims over two months in 2022 using a system called PXDX, spending an average of 1.2 seconds per claim. Cigna disputed the characterization, said the system was used to pay claims with correct codes automatically and stated it does not use AI. A class action followed regardless.

UnitedHealth and nH Predict. Plaintiffs in the Lokken case allege that UnitedHealth used an algorithm to cut off post-acute care for Medicare Advantage patients. In February 2025 a federal court allowed the breach of contract and good faith claims to proceed, as GovInfoSecurity reported. UnitedHealth's position is that the tool supports care decisions and does not make coverage determinations. The case is ongoing.

You do not need to decide who is right in either case to take the lesson. The legal exposure comes from three things:

1. The gap between what you say and what you do. If your policy or your regulator filing says a qualified professional reviews each claim, the review has to be real, documented and long enough to be meaningful.
2. Denials, not approvals. No one sues over a claim paid fast. Every automation decision that can result in a denial, a reduction or a delay deserves a separate risk review.
3. Discovery. In the Lokken case the court ordered broad discovery into the AI-driven claims process. Assume every model, prompt, override log and internal email about your claims AI could be read by opposing counsel.

My rule of thumb on denials

Automate the yes, recommend the no. If the system is confident a claim should be paid and it is within your straight-through limits, let it pay. If it thinks a claim should be denied or reduced, it prepares the reasoning and a human makes the decision with enough time and information to disagree.

The regulatory map for claims AI

Rules differ by jurisdiction, and claims AI sits under general insurance law more than under AI-specific law. A summary of the frameworks that matter most for a carrier operating in the US or Europe:

The NAIC bulletin

The NAIC adopted its model bulletin on insurers' use of AI systems in December 2023. By March 2025, 24 jurisdictions had adopted it with little or no material change, according to Quarles & Brady's tracker. The bulletin expects a written AI systems program, an accountability structure that includes claims, compliance and legal, and due diligence on third-party AI vendors including audit rights. If you run claims in an adopting state, your claims AI should be inside that program, with a named owner.

The EU AI Act

For European operations, the AI Act lists as high-risk the AI systems used for risk assessment and pricing of natural persons in life and health insurance (Annex III, point 5). Claims handling is not listed by name, but a claims model that feeds into pricing or eligibility can drift into scope, so map it carefully.

The timeline changed in 2026. After Regulation (EU) 2026/1744, the high-risk obligations for Annex III systems apply from 2 December 2027, and those for AI embedded in regulated products from 2 August 2028. Transparency obligations under Article 50, such as telling people they are talking to an AI system, apply from 2 August 2026. Article 4 asks companies to take measures that support the AI literacy of staff working with these systems; it does not set a specific level.

For a claims chatbot in the EU, the practical consequence today is simple: disclose that the customer is talking to an AI.

Governance frameworks

The NIST AI Risk Management Framework is voluntary but useful, and it maps cleanly onto what the NAIC bulletin asks for. I use its four functions (govern, map, measure, manage) as the structure for a claims AI register. For the broader policy layer, see my guide to AI governance for business and the companion piece on AI data governance, which covers the data lineage questions regulators now ask.

Build, buy or partner: choosing claims AI

Claims AI vendors fall into three groups: core claims platforms adding AI features, point solutions (photo estimating, document extraction, fraud analytics), and general AI platforms you configure yourself. The right mix depends less on technology than on how differentiated your claims process is.

Buy when the capability is commoditized and data-hungry: photo damage estimation, document OCR and extraction, standard fraud scoring. Vendors have training data you will never accumulate.

Build or configure when the workflow is specific to your book: triage rules for your product mix, coverage logic for your wordings, the orchestration between steps. This is where Allianz and Aviva put their own teams.

Partner when you need speed and lack internal capability, but insist on owning the data, the prompts and the decision logs.

I go deeper into the economics in build vs buy for AI software. For claims specifically, add one test: can the vendor explain, for a single claim, exactly why the system recommended what it did? If the answer is a demo and not a log, keep looking.

Vendor due diligence checklist

• Where does the training data come from, and does it include data from your competitors?
• Can you export every decision, input and model version for a given claim?
• What is the documented accuracy on claims like yours, not on the vendor's best customer?
• How is bias tested, how often, and will they share results?
• Does the contract give you audit rights and regulator cooperation, as the NAIC bulletin expects?
• What happens to your data if you leave?

The economics: building the business case

A claims AI business case rests on four value levers. Be explicit about which one each use case pulls, because they show up in different lines of the P&L.

1. Loss adjustment expense. Fewer handler hours per claim. The easiest to measure and the one finance trusts most.
2. Indemnity leakage. Paying the right amount, through better estimates, fraud detection and subrogation. Larger in money, slower to prove.
3. Cycle time. Faster settlement reduces rental, storage and secondary damage costs, and it reduces complaints.
4. Retention. Customers who have a good claims experience renew. Hard to attribute, real over two or three years.

A simple model

Take a carrier handling 10,000 claims a month. If AI-assisted intake and triage saves 25 minutes of handler time per claim, that is around 4,200 hours a month. At a loaded cost of 60 dollars an hour, that is roughly 250,000 dollars a month in capacity, before any effect on leakage.

These are illustrative inputs, not benchmarks. Replace them with your own handling time study. The point is the structure: time saved per claim, times volume, times cost, minus the run cost of the system, with leakage and cycle time as upside rather than the base case. My article on AI ROI for business walks through the same method in more depth.

Where the money is not

Two traps. First, capacity saved is not cost saved unless you reduce hiring, overtime or outsourcing. Decide upfront what you will do with freed capacity. Second, vendors price per claim or per seat, and costs scale with volume. Model the cost at your peak catastrophe volume, not your average month.

What I have seen outside insurance

I have not run a carrier's claims department, and I will not pretend otherwise. What I have done is lead AI projects in operations with the same shape: high-volume requests, a bottleneck of skilled people, and customers waiting. At a medical center, an AI program added about 20% capacity. At a hotel, an AI program helped move revenue from 9 million to 10 million.

The lesson that transfers to claims is the same in both: the gain did not come from the model. It came from redesigning the queue the model sits in, and from deciding in advance what the freed capacity would be used for.

Data readiness: the part nobody budgets for

Every claims AI project I have reviewed underestimated data work. The usual problems:

• Claim notes as free text with no consistent structure, written in abbreviations only the team understands.
• Outcome data missing. You know a claim closed, but not whether it was reopened, litigated or found fraudulent later.
• Policy wording versions not linked to policies.
• Photos and documents stored in a separate system with no reliable link to the claim record.
• Vendor data (repair shops, medical providers) with inconsistent identifiers.

Before you buy anything, pull 500 closed claims at random and try to answer, for each one: what happened, what was paid, how long it took, and why. If you cannot do that from the systems, neither can an AI. The AI readiness assessment I use with clients starts exactly there.

Claims AI readiness scorecard

Score each line from 0 (not in place) to 2 (fully in place). Be honest. The total tells you where to start.

0 to 8: start with data and measurement. AI will amplify the confusion you have.
9 to 16: pick one high-volume, low-complexity claim type and run a production pilot.
17 to 24: you are ready to scale across lines and add orchestration between stages.

A 90 day plan from pilot to production

This is the plan I use when a claims organization wants results in a quarter, not a two-year transformation.

Days 1 to 30: choose and measure

• Pick one claim type with high volume, low average value and clear rules. Glass, food spoilage, travel delay, minor property damage.
• Measure the baseline: volume, handling time, cycle time, cost per claim, complaints, reopen rate.
• Map the current workflow step by step, including waiting time between steps. Waiting is usually more than half the cycle time.
• Name the owner, the compliance reviewer and two experienced adjusters who will shape the design.
• Write the decision rules: what the AI may do alone, what it recommends, what it never touches.

Days 31 to 60: build and shadow

• Build or configure intake extraction and triage for the chosen claim type.
• Run in shadow mode: the AI processes every claim in parallel, humans work as usual, and you compare.
• Track agreement rate, extraction accuracy and the cases where the AI was confidently wrong.
• Set up the decision log: inputs, model version, recommendation, human decision, time spent.
• Review with compliance against the NAIC bulletin or your local equivalent.

Days 61 to 90: go live with limits

• Switch on for the chosen claim type with a value cap and confidence threshold.
• Keep human approval on any denial or reduction.
• Audit a random sample of AI-assisted claims every week.
• Report results against the baseline to the executive committee in money and days, not model accuracy.
• Decide the next claim type based on what you learned, not on what the vendor wants to sell next.

If you want a second pair of eyes on the plan before you commit budget, that is the kind of work I do: a short engagement to pressure-test the use case, the data and the governance before the expensive part starts. You can request one through the consultation page on this site.

Agentic AI in claims: what is real and what is a slide

Agentic AI, meaning systems where several AI components plan and execute multi-step tasks, is the most hyped part of this market. Allianz's Nemo is a genuine example: seven agents, each with a narrow job, orchestrated by a planner and checked by an audit agent before a human decides.

What makes it work is not autonomy. It is narrowness. Each agent does one verifiable thing: is the policy active, was there a storm at that postcode, does the claim pattern look like fraud. The orchestration saves the waiting time between those checks.

What does not work yet is a general claims agent that handles any claim end to end. Claims vary too much, the cost of error is too asymmetric, and the regulatory expectation of human accountability is too clear. For a broader view of where agents fit in a business, see my guide to AI agents for business.

Questions to ask any agentic claims proposal

1. What does each agent do, and how is its output verified?
2. What happens when two agents disagree?
3. Where exactly is the human decision point, and how long does the human have?
4. Can we reconstruct the full chain of reasoning for one claim six months later?
5. What is the fallback when the system is down during a catastrophe event?

How AI changes the claims team

The fear among adjusters is real and should be addressed directly, not with slogans. In the projects with published results, AI absorbed the clerical part of the job: reading documents, retyping data, chasing missing information, answering status calls. The judgment part stayed with people, and in several cases became more of the job.

That changes hiring and training. Junior adjusters used to learn by processing simple claims. If simple claims go straight through, you need a different path to expertise: shadowing, structured case reviews, and reviewing AI recommendations with a senior.

Two practical recommendations:

• Involve experienced adjusters in design from day one. They know which claims are simple only on paper.
• Measure override reasons. When an adjuster disagrees with the AI, capture why. That data improves the model and protects you in a dispute.

For the existing insurance strategy context, my broader piece on AI for insurance companies covers underwriting, distribution and pricing, and the guide for insurance agents covers the distribution side.

Common mistakes I see in claims AI programs

1. Starting with the chatbot. It is visible, so it gets funded first. It rarely moves cycle time, because the delay is behind it, not in front of it.
2. Measuring model accuracy instead of business outcomes. The executive committee does not care that extraction is 94% accurate. It cares about days and dollars.
3. Automating denials. The fastest route to a regulator letter or a class action.
4. Ignoring catastrophe load. A system tuned on an average month fails on the week after a hurricane, exactly when you need it.
5. No decision log. Without one you cannot improve the system, defend it or explain it.
6. Buying a platform before fixing the data. The vendor's demo uses clean data. Yours is not.
7. No plan for freed capacity. If no one decides, capacity disappears into other work and the savings never reach the P&L.

If any of these sound familiar, the fix is usually cheaper than the project that caused it. A focused review of your claims AI roadmap can save months, and you can ask for one through the consultation request on this site.

FAQ

What is AI claims processing?

AI claims processing is the use of artificial intelligence to handle parts of the insurance claims lifecycle: reading first notice of loss, extracting data from documents and photos, routing claims, checking coverage, estimating damage, detecting fraud, preparing payments and updating customers. In practice it ranges from simple document extraction to multi-agent systems that settle low-value claims in hours. The most successful programs automate clerical steps and keep humans responsible for decisions that can result in a denial or reduction.

How do you use AI in insurance claims processing without legal risk?

Start with one high-volume, low-value claim type with clear rules, and keep human approval on every denial or reduction. Document each decision with inputs, model version and the human's action. Put claims AI inside a written AI program with a named owner, as the NAIC model bulletin expects in adopting states. Test models for bias and drift on a schedule, and make sure what you tell regulators and policyholders about human review matches what actually happens in the workflow.

How much faster can AI make claims settlement?

Published results vary by claim type. Allianz reports an 80% reduction in processing and settlement time on small food spoilage claims in Australia, moving from several days to about one day or even hours. McKinsey reports that Aviva cut liability assessment time on complex motor claims by 23 days. Your result depends mostly on how much of your current cycle time is waiting between steps, which is where AI orchestration saves the most.

Will AI replace claims adjusters?

Not in the programs with published results. AI has taken over clerical tasks such as reading documents, re-entering data, chasing missing information and answering status questions. Judgment on coverage, liability, complex damage and disputed claims stays with adjusters, and regulators in the US and EU expect human accountability for decisions with legal effect. What changes is the mix of work and how junior adjusters build expertise, since simple claims increasingly go straight through.

Is AI claims processing considered high-risk under the EU AI Act?

Claims handling is not named as high-risk in the AI Act. Annex III lists AI used for risk assessment and pricing of natural persons in life and health insurance, which applies from 2 December 2027 after Regulation (EU) 2026/1744. A claims model that feeds pricing or eligibility decisions could fall into scope, so it should be mapped carefully. Chatbots that talk to claimants must disclose they are AI under Article 50 from 2 August 2026, and GDPR Article 22 already limits solely automated decisions.

Should we build or buy AI for claims?

Buy capabilities that depend on large external datasets, such as photo damage estimation, document extraction and standard fraud scoring. Build or configure the parts specific to your business: triage rules for your product mix, coverage logic for your policy wordings, and the orchestration between steps. Whatever you buy, insist on exportable decision logs for each claim, audit rights, documented bias testing and clear data ownership if you leave the vendor.

What data do we need before starting a claims AI project?

At minimum, three years of closed claims with outcomes (paid amount, cycle time, reopened, litigated, fraud confirmed), policy wordings versioned and linked to each policy, and documents and photos reliably linked to claim records. A practical test is to pull 500 random closed claims and check whether you can explain what happened in each from your systems alone. If you cannot, fix the data before buying a model.

How do you measure the ROI of AI in claims?

Measure four levers against a pre-project baseline: loss adjustment expense (handler hours per claim), indemnity leakage (overpayment and fraud), cycle time (days from notice to settlement) and complaints or retention. Calculate time saved per claim multiplied by volume and loaded labor cost, subtract the system's run cost at peak volume, and treat leakage and retention as upside. Capacity savings only count if you decide upfront how freed capacity will be used.