1. The paperwork is solved
In one of the more advanced operations I have seen, the AI pipeline covers intake from any channel, document classification and summaries, data extraction, cover and duplicate checks, triage and routing, automated chasing of missing information, an initial reserve suggestion and liability advice. Tens of thousands of claims have gone through it in about eight months. That is a serious deployment, and it still covers less than 10% of the operator's annual claim volume.
The gains are concrete. One correspondence flow handles a four-figure number of emails a day and saves about three minutes of sorting on each, roughly 7 to 8 FTE of capacity. One invoice flow frees about 2 FTE. The internal language is hours refocused, not cost removed, and the headcount question has been deferred.
This is not unusual. A 2026 survey of 110 European insurance professionals found that 51% automate 10% or less of claims-processing decisions. Only 8% reported that more than half of decisions happen without human review. FNOL and fraud detection are ahead; reserve setting remains one of the least automated steps.
The industry is getting very good at autonomous claims processing.
It is much less willing to automate autonomous claims decisioning.
2. Autonomy ends where the consequential money starts
What is not broadly live says more.
Payment, recovery and closing do not generally run autonomously across a broad P&C book. There are now several public exceptions for bounded, high-volume and predictable claims.
Allianz reports that 49.7% of German pet insurance claims were fully automated in 2025, including payout, with uncertain cases routed to human experts. Direct Pojišťovna reports 60% full automation for windshield claims, with uncomplicated claims paid within minutes. Nationale-Nederlanden says certain high-volume, low-complexity household and windscreen claims can now be processed and paid in around six minutes with no manual handling. And DOMCURA says its KIM AI platform can autonomously process qualified claims and pay them in about ten minutes.
Outside P&C, Kooperativa and MAMA AI report fully automated processing of standard life claims from notification to payment decision in around 12 seconds, with exceptions escalated to humans.
These are genuine examples of autonomous AI-enabled claims processing. But the common feature is important: the decision space is narrow enough to define, constrain and monitor.
Map the claims journey stage by stage and the pattern becomes clearer.
Intake and triage can reach AI acts alone. Cover checks and damage assessment can also reach AI acts alone for bounded claim types; reserving increasingly reaches AI acts after approval. Liability and fraud detection generally sit at AI recommends. Payment can reach AI acts alone for bounded claims, while recovery and closing remain much closer to AI suggests when the claim is consequential.
The boundary is not literally the first euro. Public evidence now shows AI paying some claims without manual handling.
The boundary is the first consequential euro.
The learning loop is missing too. In many deployments, handler corrections are logged but are not yet turned into a closed learning loop that systematically improves the decisioning system. A human review process that never feeds back into the model, rules or operating guidelines will keep needing the same human judgement.
3. The headline KPI is the model marking its own homework
A number I have encountered repeatedly is 95%+ AI confidence.
That is the model rating its own certainty, checked against human validation. It is a useful tuning signal.
It is not accuracy.
And confidence does not tell you:
- how many files went through without a person;
- how many decisions were wrong;
- how much faster claims closed; or
- what happened to leakage.
Public case studies rarely publish all four numbers together.
That distinction matters because there is a growing gap between what the technology can do and what the operating model actually delegates to it. The 2026 European claims survey found significant gaps between the benefits insurers expect from automation and the benefits they report actually achieving.
Labels drift the same way.
One pipeline described as fully autonomous turned out to be human-validated end to end. In my experience, self-assessed maturity runs about one level above what the floor actually does. It is the AI label inflation I wrote about earlier this year, one layer further down.
Meanwhile the cost side has become precise. Inference is a real line item, token consumption is hard to forecast, and business cases now carry model consumption as a separate cost lever.
So the cost of claims AI has its own budget line, while the benefit is still often expressed in hours nobody has taken out of a budget.
Until operators publish straight-through rate, decision accuracy and leakage, nobody can properly price the benefit. Buyers will therefore cap vendor pricing at whatever benefit can be evidenced.
4. The bottleneck is process knowledge
Building an agent is the easy part.
Calibrating it and writing the guidelines it follows takes people who know the process in detail. Those are often the same experienced handlers the business wants to free up.
You can only automate what you can describe consistently.
Most claims processes can be codified consistently up to the point where judgement starts. That point is usually where liability, money or customer outcome is decided.
This is why agentic claims AI is less a technology problem than an operating-model problem.
The difficult question is not:
Can the agent perform the task?
It is:
Can the insurer define the conditions under which the agent is allowed to decide?
That is a much higher bar.
5. Regulation is not the main brake
The EU AI Act's high-risk list covers AI used for risk assessment and pricing of individuals in life and health insurance. Claims management falls outside that category. That does not leave claims AI unregulated. AI literacy obligations apply to providers and deployers, and the Act's transparency rules apply to certain AI systems, including systems that interact directly with people. Insurers also remain bound by GDPR and sectoral governance requirements, including Solvency II and EIOPA's 2025 opinion on AI governance.
GDPR Article 22 matters more. It applies when a decision with legal or similarly significant effects, such as refusing a claim, is based solely on automated processing. Article 22 permits such decisions only in limited exceptions, including where they are necessary for entering into or performing a contract, and safeguards still apply: human intervention, the opportunity to express a view and the right to contest the decision. Where special-category health data are involved, the bar is higher: explicit consent or a substantial public-interest basis in law. A handler who merely signs off whatever the model recommends is not meaningful human involvement.
The brake is therefore not simply regulation.
It is internal.
Nobody has measured outcomes well enough to let go of the decision.
What this means
The operators doing this work are ahead of most of the market, and the capacity they have released is real.
The risk sits in the vocabulary.
When faster intake, automated assessment and workflow orchestration are bought and budgeted as autonomous claims handling, the business case starts resting on decisions the operating model still reserves for people.
The next step up the ladder belongs to operators that measure outcomes on the decisions themselves.
Reserving and cover are the obvious candidates. AI already recommends there. Handler corrections are already being generated. The data exists.
Using it to move from recommendation to delegated decisioning is an operating decision insurers are only beginning to take — and, so far, mainly for bounded claim populations.
There is a simple test.
Ask any claims operator what share of files closed this year with zero human touch.
Then ask how much of that share came from AI rather than rules-based automation, and how much came from tightly bounded claim types.
That second number — and the complexity of the claims it covers — is where claims AI really is.