.svg)
of European insurers already use generative AI
EIOPA, 2026
Use case · Insurance claims
Agentic claims processing for insurers
AI agents move a claim from first notice of loss to settlement. People approve the decisions that matter.
People approve key decisionsEvery action loggedDesigned for European rules
- IntakeReads the first notice of lossAgent
- TriageScores complexity, urgency and fraud signalsHuman check
- CoverageChecks the policy, wording and limitsHuman check
- DocumentsAsks for missing documents and chases repliesAgent
- AssignmentSends the claim to fast track or the right handlerAgent
- SettlementDrafts the amount, calculation and customer letterHuman check
Every action logged: decision, reason, source
AI Agentscoordinated by an orchestrator
3 human checksfraud and exceptions, partial or decline coverage, settlement review and approval
The idea
What an agentic workflow does differently
In short
Agentic claims processing uses a team of AI agents, each doing one job in the claim, to move an insurance claim from first notice of loss to settlement through the insurer’s own systems. An orchestrator coordinates the agents, tracks the claim and records every step. People approve the decisions that matter: fraud and exceptions, partial or decline coverage, and settlement review and approval. Moxie designs and builds these workflows with insurers, starting with one claim type in one business line.
| Capability | Rules automation | Chatbot | Agentic workflow |
|---|---|---|---|
| Reads unstructured documents | No | Partly | Yes |
| Plans the next step | No, fixed script | No | Yes, within set limits |
| Acts in core systems | Yes, fixed paths | Rarely | Yes, approved tools only |
| Hands decisions to people | Only by rule | No | At defined checkpoints |
9% of generative AI use cases reported to EIOPA are agentic. In 2026, a quarter of newly disclosed insurer use cases showed agentic AI. Sources: EIOPA GenAI Market Survey, February 2026; Evident AI Index for Insurance, June 2026.
Why now
Most insurers use AI. Few have scaled it.
The gap sits between a promising pilot and daily claims work at scale.
Sources: EIOPA GenAI Market Survey, February 2026; Capgemini World P&C Insurance Report 2026.
of their generative AI use cases are in production. 64% are still proofs of concept.
EIOPA, 2026of property and casualty insurers are successfully scaling AI
Capgemini, 2026The workflow
One claim, several agents, three human checks
The orchestrator keeps the claim’s status, decides which agent works next and logs every step. Fraud scoring is dynamic and reruns whenever new data arrives.
1
Intake
Reads the first notice of loss from any channel.
2
Triage
Scores complexity, urgency and fraud signals.
Human check: fraud flags and exceptions3
Coverage
Checks the policy, wording and limits.
Human check: partial or decline coverage4
Documents
Asks for missing documents and chases replies.
5
Assignment
Sends the claim to fast track or the right handler.
6
Settlement
Drafts the amount, calculation and customer letter.
Human check: settlement review and approvalIllustrative design. Agent roles and checkpoints are set by each insurer.
Tools and guardrails
Agents act only through approved tools
Every action is logged with the decision, the reason and the data source.
Policy administration
Can: read policies, wordings, limits
Guardrail: read only
Claims core system
Can: create and update claim records
Guardrail: every change logged
Document store
Can: read, sort and file documents
Guardrail: personal data masked when not needed
Customer channels
Can: send requests and status updates
Guardrail: approved templates only
Fraud scoring service
Can: request a score from a bought tool
Guardrail: the score informs, people decide
Payments
Can: prepare the payment
Guardrail: released only after approval
Illustrative design. Permissions and guardrails are set by each insurer.
Human review
People stay in control where it matters
| Checkpoint | The agent prepares | The person decides | Why it is here |
|---|---|---|---|
| Fraud and exceptions | Fraud signals and the reasons behind them | Investigate, or release to fast track | False flags hurt honest customers |
| Partial or decline coverage | The clauses that apply, quoted from the policy | Confirm, change or overrule | Refusing cover directly affects the customer |
| Settlement review and approval | Amount, calculation and draft letter | Approve above their authority limit | Payment authority stays with people |
| Customer disagrees | The full decision log for the claim | Review the claim manually | Customers keep the right to a human review |
Across lines of business
How the workflow adapts to each business line
Motor
- Intake from phone, app and photos, often at the roadside
- Triage separates glass-only claims from injury and liability
- Photo damage estimates come from a bought tool the agents call
Home
- Storms create surges of similar claims within days
- Cover depends on the cause of damage, so wordings matter
- Documents include photos, contractor quotes and invoices
Pet
- High volumes of small vet invoices, often paid by the owner first
- Diagnoses checked against exclusions and pre-existing conditions
- Clinical history requested directly from the vet
Health
- Medical invoices and referrals arrive in many formats
- Health data is special category data under GDPR
- Cover depends on plan, provider network and prior approval
Next step
Start with one claim type in one business line
Choose a high-volume, low-complexity claim type
Measure today’s time, cost and customer outcome
Book a discovery with Moxie
FAQ
Agentic claims, answered
A team of AI agents, each doing one job in the claim, works through the insurer’s own systems from first notice of loss to settlement. An orchestrator coordinates the agents, tracks the claim and records every step, and people approve the decisions that matter.
Rules automation follows a fixed script and a chatbot rarely acts in core systems. An agentic workflow reads unstructured documents, plans the next step within set limits, acts through approved tools only and hands decisions to people at defined checkpoints.
At defined checkpoints: fraud and exceptions, partial or decline coverage, and settlement review and approval above a person’s authority limit. Customers keep the right to a human review, and further checks can happen when new data is uploaded.
Claims handling is not high-risk under the EU AI Act by default, but life and health pricing and risk assessment are, so those decisions stay outside the claims agents. Supervisors expect governance, human oversight and documentation (EIOPA AI Opinion, August 2025), and health and injury data are special category data under GDPR. This is general information, not legal advice.
With one claim type in one business line. Choose a high-volume, low-complexity claim type, measure today’s time, cost and customer outcome, then book a discovery with Moxie.
Let’s talk
Book a discovery with Moxie
Tell us which claim type you would start with. We will take it from there.
Talk to an expert
Prefer email? Write to hello@moxie.pt.