End-to-end insurance claims handling automation

Claims handling is the largest cost block of an insurer and the moment customers judge it. Agentic AI now chains the previously separate steps (intake, coverage check, document reading, damage estimate, fraud screening, settlement proposal, customer messaging) into one orchestrated process. Simple, high-frequency claims can run nearly touchless; complex, large or suspicious ones are prepared for a human adjuster, who keeps the final decision on payout and rejection.

Problem & value

• Today: a claim passes through several systems and teams. Adjusters re-key data, request missing documents by email, look up policy clauses and write letters. Leakage (overpayment), slow settlement and weak fraud detection follow.
• AI value: shorter cycle time for simple claims, consistent coverage application, better triage (straight-through vs. specialist vs. special investigation), and adjusters who spend time on judgement and empathy rather than data entry.
• Evidence: Allianz reports that its agentic "Project Nemo" for food-spoilage claims in Australia cut processing times by about 80%, using seven task-specific agents with a human taking the final payout decision [1]. Shift Technology reports early-adopter figures for its agentic claims product (vendor-reported, not independently audited) [2]. Treat vendor numbers as self-reported.

Variants

• Motor (own damage and glass): photo-based damage estimate, repair-shop network, total-loss logic (see VIS-023).
• Property and weather events: mass-claim surges after storms; aerial imagery and weather data for triage.
• Travel, pet and simple health reimbursements: low value, high volume, best candidates for touchless handling.
• Commercial and liability lines: agents only prepare files, summarise documents and draft reserves; no autonomous settlement.
• Health and life claims: special-category data; the strictest human oversight and AI Act scrutiny.
• Broker, TPA and MGA operations: same pipeline run for several carriers with per-tenant rules.

How it works

1. FNOL (app / chat / email / phone)
2. intake agent (structured claim, missing-info questions)
3. coverage agent (policy, exclusions, deductible, waiting periods; cites clause)
4. document agent (OCR/VLM on invoices, reports, photos)
5. assessment agent (damage/cost estimate, tools: repair-cost APIs, VIS-023)
6. fraud & complexity triage (rules + ML + graph signals, see ANOM-003)
7. settlement agent (proposal, reserve, letter draft)
8. [human adjuster: approve / amend / reject; mandatory above thresholds or on any doubt]
9. payment + customer notification + audit log

• Orchestration: a supervisor agent routes to specialised agents with narrow tools, exposed via APIs or MCP servers on top of the core claims system (Guidewire, Duck Creek, Sapiens, Majesco etc.).
• Deterministic core: coverage rules, limits and payment calculations should run as code or a rules engine that the LLM calls, not be improvised by the model.
• Human-in-the-loop: adjusters approve every payout above a threshold, every rejection or partial denial, every fraud referral and every complaint. Customers can always request a human review.

Data preparation

Data needed: claim forms, policies and coverage terms, claim histories, photos, expert reports, invoices and correspondence. In insurers these live in the claims or policy core system, the DMS, mail inboxes and on paper scans, often with inconsistent metadata. Typical gaps: scans without OCR, photos without claim link, free-text coverage rules, duplicate customer records, missing outcome labels, and health data with broad internal access rights.

Preparation steps: (1) inventory sources and map them to the process steps (PREP-002); (2) OCR and classify incoming documents (PREP-006, DOC-003, DOC-014); (3) turn policy conditions and handling guidelines into an authoritative knowledge base (PREP-013); (4) harvest labels from closed claims - coverage decision, settlement, fraud flags (PREP-020) - and build an evaluation set with experienced adjusters (PREP-018); (5) pseudonymise data for development and testing (PREP-022); (6) fix permissions on health and claims data (PREP-023, PREP-024) and check purpose and legal basis (PREP-025).

Effort: high, about 40-50 % of the project, because history is spread over legacy systems and labels must be validated by adjusters. A readiness check per GUIDE-003 prevents surprises.

Implementation paths

Which path for whom

• Enterprise: Path B - claims handling is the insurer's core process and holds special-category data; building on core-system integration with EU-hosted model APIs keeps control, while a vendor suite (Path A) suits insurers without engineering capacity.
• Developer: Path B - cloud APIs in EU regions plus an agent framework give the fastest route to a testable claims workflow; Path C for sovereignty-driven projects.

Path A — Buy (off-the-shelf)

• Core-system vendors with agent layers: Guidewire (agentic framework and agentic FNOL), Duck Creek (Agentic AI Platform, announced April 2026) [3], Sapiens (Israel/US), Majesco, and EIS. Best when the agents must act inside the claims core.
• Specialists: Shift Technology (France, EU; claims automation and fraud) [2], Tractable (UK; photo damage assessment), CCC and Solera/Qapter (motor) Also Hypatos and Klippa (EU) for document capture, and Cognizant, Accenture and Capgemini as integrators.
• Choose A when: you run a mainstream claims suite, have no large ML team, and need vendor support for regulatory documentation.

Path B — Build on cloud APIs (recommended for enterprise insurers)

• Models: a frontier LLM with tool use (Claude, GPT, Gemini) or Mistral (EU provider), hosted in an EU region (Azure OpenAI EU, Google Vertex EU, AWS Bedrock Frankfurt); a VLM for documents and photos.
• Frameworks: LangGraph, Microsoft Agent Framework, or the vendor SDKs; MCP for tool access; a rules engine or BPMN engine (Camunda) for the deterministic process skeleton.
• Integration: claims-core APIs, DMS, payment system, a vector index over policy wording and guidelines.
• Choose B when: you want differentiated logic, have a platform team and can document the system yourself.

Path C — Self-host (open source)

• Models: open-weight LLMs such as Mistral, Qwen, Llama or gpt-oss-class models; Qwen-VL for documents and photos; Docling for parsing; served via vLLM.
• Stack: Camunda or Temporal, LangGraph, Postgres with pgvector, Langfuse for tracing.
• Hardware: HW3 for a 30-120B model serving an adjuster team; HW4 for surge capacity or fine-tuning.
• Choose C when: health or life data, national data-sovereignty rules or group policy forbid external APIs.

Path D — Custom training

Worthwhile for the fraud/triage model and for a fine-tuned damage-estimation model, trained on the insurer's own historic claims (thousands to millions of labelled outcomes). Fine-tuning the language model is rarely needed; retrieval over clauses and guidelines usually suffices.

State of the art (2026)

• Production agentic claims is real but narrow: leading carriers automate low-value, high-frequency lines and keep humans for the rest [1]. Core-system vendors now ship native agent frameworks [3].
• Pattern that works: many narrow agents with deterministic checks beat one general agent; every agent output carries a citation to the policy clause or document it relied on.
• Evaluation: replay historic closed claims through the agent and compare coverage verdicts and amounts with adjuster outcomes before going live (shadow mode).
• Limits and failure modes: wrong coverage interpretation from ambiguous wording, hallucinated amounts, prompt injection through uploaded documents or customer emails, manipulated or AI-generated photos and documents (see ANOM-017), unfair treatment of vulnerable customers, and automation bias of adjusters who simply click approve.

Shared resources & combinations

• Document stack: DOC-003 (claim forms), DOC-024 (submissions), shared VLM server.
• Damage and fraud: VIS-023 (photo damage assessment), ANOM-003 (claims fraud), ANOM-017 (forged documents).
• Prediction: PRED-037 (severity, litigation, reserving), PRED-010 (pricing feedback from claims data).
• Front door and platform: CONV-013 (customer assistant), AGENT-001 (mail triage), AGENT-013 (multi-agent patterns), AGENT-025 (agent governance and evaluation), AGENT-026 (analogous public-sector case processing).

Effort, cost & KPIs

• Path A: 3-9 months incl. core-system integration and change management; licence plus per-claim fees.
• Path B: 6-12 months for the first line of business; costs are integration, evaluation and compliance documentation rather than tokens.
• Path C/D: 9-18 months; GPU operations and MLOps team.
• KPIs: straight-through-processing rate, cycle time FNOL to payment, loss-adjustment expense per claim, indemnity leakage, share of overturned agent proposals, complaint and reopening rate, fraud hit rate, customer satisfaction (NPS), and fairness metrics across customer groups.

People & change

Involve management and the claims department lead first, then adjusters (end users) as co-designers; IT and IT security for core-system integration and logging; the DPO early because claims often contain health data (special categories) - a DPIA is very likely, plus AVV with each processor; works council before the pilot starts.

Works council: yes. Claims systems record adjuster-level decisions, overrides and processing times, which makes them objectively suitable for monitoring (§87(1) no. 6 BetrVG, BAG case law); §90(1) no. 3 requires timely information about AI-supported work processes, §80(3) allows an AI expert, and a works agreement on AI use and no individual performance scoring builds trust. Public insurers involve the staff council.

Take measures to support AI literacy (AI Act Art. 4): train adjusters on reading agent proposals, spotting hallucinated coverage statements and documenting overrides. Start with one simple line (e.g. glass or travel claims) and a pilot group of experienced adjusters; communicate that the agent prepares and the adjuster decides on contested cases and all denials. Typical concerns are deskilling and customer fairness - address them with human decision rights, sample audits and an appeal path. Guides: GUIDE-004, GUIDE-005, GUIDE-008.

Risks, regulation & compliance

• EU AI Act: classification depends on scope. Annex III point 5(c) lists risk assessment and pricing for life and health insurance as high-risk [4]; pure property/motor claims handling is generally not listed but can become high-risk where it decides access to essential benefits or health insurance services, so assess each line. The Digital Omnibus (Regulation (EU) 2026/1744) moved the Annex III high-risk application date to 2 December 2027 [5]. No Art. 5 prohibition applies, but emotion recognition on callers or video claimants should be avoided (Art. 5 bans it at work and school, and it is high-risk elsewhere). Art. 4 AI literacy and Art. 50 transparency (customers must know they talk to an AI) apply now.
• EIOPA: the 2025 opinion on AI governance and risk management applies insurance-sector law (Solvency II, IDD) to non-high-risk AI: proportionate governance, data quality, human oversight, explainability and record keeping [6].
• GDPR: claims contain personal and often health data (Art. 9). Needs a legal basis (contract, Art. 9(2)(f) legal claims), DPIA, processor agreements and EU hosting. Art. 22 limits solely automated decisions with legal or similarly significant effect: keep meaningful human review for rejections [7].
• Other: DORA for ICT third-party risk, national claims-handling conduct rules (e.g. German VVG, BaFin MaRisk VA/VAIT), Product Liability Directive for software.
• Mitigations: thresholds for autonomy, shadow mode, sampling audits by quality teams, bias testing, document-injection defences, complete decision logs, and a documented customer path to a human.

Getting started

1. Pick one line with simple, high-volume claims (e.g. glass, travel, household); document the baseline KPIs.
2. Map the process and write down the deterministic rules; decide which decisions an agent may take and which it only prepares.
3. Build a golden set of 300-500 closed claims with adjuster outcomes and run the agent in shadow mode.
4. Add citations, logging and a review screen for adjusters; define thresholds, then pilot with real customers for one product.
5. Complete the AI Act classification, DPIA and governance documentation before scaling to more lines.

Sources

1. Allianz — "When the storm clears, so should the claim queue" (Project Nemo) — https://www.allianz.com/en/mediacenter/news/articles/251103-when-the-storm-clears-so-should-the-claim-queue.html
2. Shift Technology — insurance claims automation — https://www.shift-technology.com/
3. Duck Creek Technologies — Agentic AI Platform — https://www.duckcreek.com/
4. Regulation (EU) 2024/1689 (AI Act), Annex III — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
5. Digital Omnibus on AI, Regulation (EU) 2026/1744 — summary: https://www.orrick.com/en/Insights/2026/07/EU-AI-Act-Update-Digital-Omnibus-Finalizes-8-Compliance-Changes
6. EIOPA — Opinion on AI governance and risk management (6 Aug 2025) — https://www.eiopa.europa.eu/eiopa-publishes-opinion-ai-governance-and-risk-management-2025-08-06_en
7. Regulation (EU) 2016/679 (GDPR), Art. 9 and 22 — https://eur-lex.europa.eu/eli/reg/2016/679/oj