Knowledge Graph vs RAG for Insurance Claims Adjudication

---

For claims adjudication, use both, split by question. In Han et al. (2026), on temporal questions in one benchmark built from English news articles, GraphRAG scored 49.06 against RAG’s 43.74 with retrieval tokens matched. Atlan’s MCP server applies “the same persona, role, and domain policies that govern the UI” to every tool call and logs each call.

Vector RAG reads what a clause says; a knowledge graph traces who links to whom and which endorsement changes which clause. Picking the form in force on the loss date is an ordinary date filter that SQL runs as well as a graph.

---

A claim file asks three kinds of question. Wording: what does the water-damage exclusion say? Version: which form applied on the date of loss? Relationship: does this contractor appear on other claims? Each takes its own path: vector RAG, a filter or a knowledge graph. In Han et al. (2026), token-matched scores were 71.01 for RAG and 71.17 for GraphRAG overall; without matching, RAG scored 25.73 on temporal questions, so much of the gap was context budget. These are news and general question sets, so read them as direction. Point both patterns at an enterprise context layer for definitions and policy.

• A date filter picks the governing version first; RAG reads the clause inside it.
• Overlapping amendments look alike to embeddings, so unfiltered similarity can rank a later edition first (TimelyRAG, 2026).
• Access is checked before retrieval and at every hop, and an adverse answer returns its provision and version.

Dimension	Vector RAG	Knowledge graph	Use both when
Claims question it answers best	What an exclusion, condition or definition says	Who links to whom; which endorsement modifies which clause	The answer needs the right version and its wording
What it returns	Quoted passage with form and section	Nodes and edges with record IDs	An evidence bundle carries both
Policy versions	Needs an effective-date pre-filter; unfiltered, a newer edition can win (TimelyRAG, 2026)	Holds effective intervals plus modifies and supersedes edges	The date filter works in SQL too; a graph adds precedence chains
Benchmark signal (Han et al., 2026, MultiHop-RAG, built from news articles)	Better on single-hop and detail questions; 71.01 overall with tokens matched	GraphRAG on temporal questions: 49.06 vs RAG’s 43.74 with tokens matched (25.73 unmatched); 71.17 overall	Overall scores nearly tie
Unanswerable questions (Han et al., 2026, appendix)	91.36 on null questions	13.95 for GraphRAG global community summaries	Abstain and route to a person
Typical claims failure	Clause split from its exception (StrucChunk, 2026)	Missing edge in a text-extracted graph read as a missing fact	Text stays the authority
Where access is checked	Before retrieval (Authorization-First Retrieval, 2026)	Again at every hop (Retrieval Pivot Attacks, 2026)	At both points

---

Which claims questions need RAG and which need a knowledge graph?

The type of claims question picks the starting pattern, so tag every question first (checklist item CRP-01, below). The general trade-off between knowledge graphs and RAG covers latency and cost; this table keys on the claim file.

Three things go by “graph” here. A date lookup is a filter that SQL runs as well as a graph. A graph loaded from policy and claims systems earns its cost on party links and endorsement precedence. GraphRAG extracts a graph from text and summarizes it, for sensemaking across many claims.

Claims question (example)	Question type	Starting pattern	Must return	Evidence
“What does the water-damage exclusion say?”	Single-hop wording	Vector or hybrid RAG over whole-provision chunks	Quoted provision, form ID, edition, section	RAG vs. GraphRAG (2026); StrucChunk (2026)
“Which form and endorsements were in force on the date of loss?”	Version	Effective-date filter in SQL, index metadata or the graph, then RAG inside that version	Policy ID, term, form editions, endorsement IDs with intervals	TimelyRAG (2026)
“Does endorsement E change exclusion X?”	Multi-passage composition	Graph edge (modifies or supersedes), then retrieve both texts	Both provisions quoted, plus the precedence edge and its source	StrucChunk (2026); RAG vs. GraphRAG (2026)
“Is this repair shop linked to prior claims on other policies?”	Relationship, multi-hop	Graph traversal with typed party roles	Path from claim to party to prior claims, with record IDs	RAG vs. GraphRAG (2026)
“How many open claims share this phone number and bank account?”	Count or intersection	Deterministic graph or SQL query from a tested template	The count plus the query used	Tested query template; no benchmark cited
“What patterns run across this quarter’s denied water claims?”	Global sensemaking	Community-summary GraphRAG	A summary labeled non-evidentiary	From Local to Global (2024); GraphRAG docs
“Is there enough to deny?”	Adverse decision	Version filter, provision retrieval, graph facts, abstention gate, human review	The evidence bundle	Null-question scores, RAG vs. GraphRAG (2026); NAIC Model 902 Section 7A

Hybrid retrieval that pairs BM25 with dense vectors helps the wording row, since form numbers are exact strings. Relationship rows resemble fraud-ring detection with a knowledge graph. A shared contractor is a referral signal for a person, never denial evidence on its own.

When is RAG plus SQL enough? If your agent answers only wording and version questions, an index with a date pre-filter from the policy system covers both, and a graph is one more pipeline to keep current. Add one when party links or precedence chains enter the workload. Counts need a tested query template either way.

For AI agents for insurance, routing is a context engineering decision made per question.

---

What must a claims retrieval answer return before it can support a denial?

Claims-handling rules fix the output a claim file needs, and that output shapes retrieval more than benchmark scores do.

NAIC Model 902, the property and casualty claims model regulation, bars denying a claim on a policy provision unless the denial references it (Section 7A). The NAIC’s Unfair Claims Settlement Practices Act treats failing to give “a reasonable and accurate explanation of the basis” for a denial, and knowingly misrepresenting policy provisions, as unfair claims practices when committed flagrantly or often enough to indicate a general business practice (Sections 4L, 4A and 3). A wrong-edition explanation puts accuracy at risk. Model 902 also requires claim data that is “accessible and retrievable for examination” (Section 4A) and claim files that “permit reconstruction” (Section 4B).

The NAIC AI Model Bulletin, adopted December 4, 2023, says an insurer’s actions must not violate those laws “regardless of the methods the Insurer used” and names claim administration among the AI uses in scope. Model 902 covers property and casualty only, and states adopt their own versions, so check your state’s text.

For AI risk management, the NIST AI RMF 1.0 (2023) defines explainability as “a representation of the mechanisms underlying AI systems’ operation.”

In the EU, Annex III point 5© of the EU AI Act lists AI systems “intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance” as high-risk (Regulation (EU) 2024/1689). The point does not name claims handling.

Source	Provision	What it asks of retrieval output
NAIC Model 902, property and casualty claims model regulation	Section 7A: no denial on a provision unless the denial references it	The exact provision ID and text from the governing version
NAIC Model 902	Sections 4A and 4B: claim data accessible and retrievable for examination; claim files that permit reconstruction	Keep the evidence bundle in the claim file
NAIC Model 900, Unfair Claims Settlement Practices Act	Section 4L: reasonable and accurate explanation of a denial; Section 4A: knowingly misrepresenting policy provisions; Section 3: flagrant, or frequent enough to indicate a general business practice	A wrong-version explanation is not an accurate one
NAIC AI Model Bulletin, adopted 2023-12-04; 29 jurisdictions listed as adopted on the NAIC map, status 2026-10-08	1.6 claim administration and fraud detection; 3.2 data currency and lineage; Section 3 explainability to the consumer	Data currency and lineage are named program items
Colorado, listed on the same NAIC map under insurance-specific regulation	3 CCR 702-10, effective 2023-11-13; amendments effective 2025-10-15	Read the regulation itself for the lines it covers
NIST AI RMF 1.0 (January 2023)	3.4 Accountable and Transparent; 3.5 Explainable and Interpretable	Vocabulary for model-risk review
EU AI Act, Annex III point 5©	Risk assessment and pricing for natural persons in life and health insurance	High-risk scope as written; claims handling is not named
W3C PROV-O (Recommendation, 30 April 2013)	prov:used, prov:wasDerivedFrom, prov:wasAttributedTo	A standard way to record what an answer used

---

How does RAG handle policy wording, and where does it break?

Vector RAG is the right tool for reading clauses; its claims failures come from how documents are cut and versioned.

According to Han et al. (2026), RAG performs better on single-hop and fine-grained detail questions.

Fixed-size chunking cuts a clause from its exception. In a legal-text preprint, structure-aware chunking reached 100% boundary precision against 7.7% for recursive splitting on an Indian procedural code, and kept 97.6% of cross-references against 0% (StrucChunk, 2026). Use chunking strategies for RAG that follow clause boundaries (CRP-04).

Amendments override earlier clauses but keep most of the text, so editions overlap. In TimelyRAG (2026), a study aimed at amended laws, policies and regulations, temporal ranking improved nDCG@10 by up to 28.6%. The plainer fix is metadata pre-filtering on effective dates before ranking (CRP-02, CRP-03).

In a 2024 Quebec auto-insurance study, 5% to 13% of RAG answers included a false statement that could mislead a customer (NLLP 2024). A 2026 Quebec insurance benchmark found RAG lifted weak models by over 35 percentage points but caused “context distraction” in others (arXiv 2603.07825). Both are familiar reasons RAG answers go wrong; test each model and retriever pair end to end (CRP-12).

---

How does a knowledge graph handle policy versions, parties and prior claims?

A claims knowledge graph models policies, forms, endorsements, parties and claims as typed nodes with dated edges; it beats a table on relationships, not on a date lookup.

The OMG P&C Data Model 1.0, published October 2014, “addresses the data management needs” of the property and casualty insurance community. Typing those entities is ontology design; the ontology and knowledge graph split sets where rules live.

On 13 OMG P&C tables, Claim among them, a large language model answered 16% of questions correctly over direct SQL and 54% over a knowledge graph of the same database (Sequeda et al., 2023). That compares a graph with SQL, not with RAG.

Effective intervals map to the hasBeginning and hasEnd properties of OWL-Time, a W3C Candidate Recommendation Draft (15 November 2022). Precedence is where the graph earns its place: a modifies or supersedes edge is a stored fact, never a similarity guess (CRP-05). Run entity resolution on claimants and contractors before you traverse.

Graphs extracted from text miss things: in Han et al., only about 65.8% of answer entities appeared in the graph built for HotpotQA, a multi-hop question set. A graph loaded from systems of record fails through stale loads and bad entity matches instead. Validate knowledge graph construction, track temporal validity in a context graph, and watch for static graphs that drift out of date.

---

Worked example: one homeowners claim through both patterns

One hypothetical burst-pipe claim runs the order end to end: date filter, RAG, graph, evidence bundle. Claim CLM-0001, loss date 2026-02-14, burst pipe. Policy HO-123 renewed on 2026-01-01 on a new form edition; a water-backup endorsement was added on 2026-03-01, after the loss. Both queries are hypothetical openCypher, grammar-checked only.

1. Find the version in force on the loss date. The query runs it in the graph; the same predicate works as a SQL WHERE clause or an index pre-filter. It drops the 2026-03-01 endorsement a similarity search could return.

MATCH (c:Claim {id:'CLM-0001'})-[:UNDER]->(p:Policy)-[:HAS_FORM|HAS_ENDORSEMENT]->(d:PolicyDocument)
WHERE d.effective_from <= c.loss_date AND (d.effective_to IS NULL OR d.effective_to > c.loss_date)
RETURN d.id, d.edition, d.effective_from, d.effective_to

2. Retrieve whole-provision chunks whose document_id is in the step 1 result, and return the quoted text and section.
3. Find other claims that list the same contractor as a party, each with its record ID. A match is a referral signal for a person; re-check access at every hop.

MATCH (c:Claim {id:'CLM-0001'})-[:HAS_PARTY {role:'contractor'}]->(v)<-[:HAS_PARTY]-(other:Claim)
RETURN other.id, v.id

Neither query shows access checks; enforce them in the layer that runs each query.

4. Write the evidence bundle to the claim file as decision traces, using W3C PROV-O terms prov:used, prov:wasDerivedFrom and prov:wasAttributedTo. It holds document IDs with editions, quoted provisions, graph paths, retriever and model versions, the access decision and the reviewer.

The bundle carries the provision reference Model 902 Section 7A asks for and the explanation Model 900 Section 4L requires, and stays retrievable under Sections 4A and 4B. It supports a reviewer; it does not make a denial correct.

To build a knowledge graph for AI agents, pick a graph database that supports your query language and access controls, and store each bundle as versioned context.

---

Where do permissions leak in vector and graph retrieval?

Studies of general RAG systems find leaks at two points, after retrieval and at graph expansion, and in claims those are the points where claimant and medical records would leak.

In one multi-agent RAG test setup, retrieve-then-filter pipelines exposed unauthorized context in 86.1% of base queries, while authorizing first removed structural leaks by construction (Authorization-First Retrieval, TrustNLP 2026) (CRP-09).

Hybrid systems add a second door: an allowed chunk links through a shared contractor or clinic into other claimants’ records. In a 2026 study of retrieval pivot attacks, tested on enterprise document sets rather than claims data, enforcing authorization at the graph-expansion boundary cut measured leakage to zero (CRP-10).

Health claims raise the stakes; read the HIPAA obligations for AI agents before medical records enter the graph. Define agent access control once, through role-based access control in a context layer, and enforce it on every retrieval path: your vector index and claims graph need their own enforcement points.

---

What goes wrong when claims teams use one pattern for everything?

Nine failure modes recur when one pattern, or one unguarded hybrid, answers every claims question. Several are pitfalls of enterprise knowledge graphs; the last is about definitions, which is how a context layer reduces hallucinations.

#	Failure	What happens in a claim	Fix	Source
F1	Fixed-size chunking splits a clause from its definition or exception	The agent quotes an exclusion without the carve-back	Structure-aware chunks that keep clause boundaries and resolve cross-references	StrucChunk (2026)
F2	Version drift	The agent retrieves a newer edition or a post-loss endorsement	Effective-date filter before ranking; “latest” and “in force on loss date” are different queries	TimelyRAG (2026)
F3	Extracted-graph gaps	A missing edge in a graph built from text is read as a missing fact	Keep source text authoritative; never infer “no prior claims” from a graph miss	RAG vs. GraphRAG (2026)
F4	Community summaries used as evidence	The denial rests on a summary instead of a provision	Summaries for triage only; evidence is quoted text	RAG vs. GraphRAG (2026); From Local to Global (2024); GraphRAG docs
F5	Retrieve-then-filter permissions	The model sees restricted claimant or medical context before the filter runs	Authorize before retrieval	Authorization-First Retrieval (2026)
F6	Vector-to-graph pivot	An allowed chunk links through a shared entity into other claimants’ records	Re-check authorization at graph expansion and each hop	Retrieval Pivot Attacks (2026)
F7	Retrieval makes some models worse	Adding a corpus degrades answers for one model and prompt	Test the exact model, prompt and retriever end to end	Quebec insurance benchmark (2026)
F8	Fluent but false	Answers include statements that could mislead a customer	Quotation-only evidence, an abstention gate, human review on adverse outcomes	Quebec automobile insurance QA (2024)
F9	Uncertified definitions	“Open claim” or “reported date” resolves from a non-authoritative table	Certified glossary terms and lineage from the context layer	Atlan docs: glossary, lineage, certificates

---

Is GraphRAG the right choice for claims adjudication?

GraphRAG’s community summaries suit pattern-spotting across many claims, not the evidence for one determination.

GraphRAG builds community summaries for “global sensemaking questions over datasets in the 1 million token range” (Edge et al., 2024). The open-source GraphRAG project is MIT-licensed, and its query docs call global search “a resource-intensive method, but often gives good responses,” while local search combines graph data with text chunks.

In one evaluation, the summaries seldom abstained. On null questions, vector RAG scored 91.36 against 13.95 for global community GraphRAG in Han et al. (2026), so an abstention gate (CRP-11) is mandatory on this pattern.

Use summaries for triage, label them non-evidentiary, and keep combining knowledge graphs with LLMs for answers that cite sources.

---

Checklist: is your claims retrieval ready for adverse decisions?

Fourteen checks, CRP-01 to CRP-14, decide whether a claims agent can be trusted with adverse outcomes.

ID	Check	Pass condition	Source
CRP-01	Questions classified	Each claims question is tagged wording, version, relationship, count, sensemaking or adverse decision	Question-type decision table; RAG vs. GraphRAG (2026)
CRP-02	Version filter precedes ranking	Effective-date and jurisdiction filter, in SQL, index metadata or the graph, applied before any similarity ranking	TimelyRAG (2026)
CRP-03	All versions retained	Superseded forms and endorsements kept with effective intervals	TimelyRAG (2026)
CRP-04	Structure-aware chunks	Chunks align to clause boundaries; cross-references resolved	StrucChunk (2026)
CRP-05	Endorsement precedence explicit	Modifies and supersedes relationships stored, never inferred from similarity	Design rule, shown in the worked example
CRP-06	Provision cited in output	Every adverse answer returns provision ID, edition and quoted text	NAIC Model 902 Section 7A; Model 900 Section 4L
CRP-07	Graph edges have provenance	Each node and edge points to a source record	RAG vs. GraphRAG (2026); W3C PROV-O
CRP-08	Graph miss is not evidence	Text fallback when a graph lookup is empty	RAG vs. GraphRAG (2026)
CRP-09	Authorize before retrieval	Permission predicates applied before ranking	Authorization-First Retrieval (2026)
CRP-10	Re-authorize per hop	Access re-checked at vector-to-graph expansion and at each hop	Retrieval Pivot Attacks (2026)
CRP-11	Abstention gate	An “insufficient evidence” path routes to a person	RAG vs. GraphRAG (2026); Quebec automobile insurance QA (2024)
CRP-12	End-to-end test per model	The retrieval and model combination is tested on realistic adverse-decision cases before release	Quebec insurance benchmark (2026)
CRP-13	Certified definitions	Business terms resolve to certified glossary entries with lineage	Atlan docs: glossary, lineage, certificates
CRP-14	Evidence bundle stored	Bundle written to the claim file and retrievable for examination	NAIC Model 900 Section 4L; Model 902 Sections 4A and 4B

Score each check PASS, FAIL or UNKNOWN. Any FAIL or UNKNOWN on CRP-02, CRP-06, CRP-09, CRP-10 or CRP-11 means not ready for adverse decisions; any other FAIL or UNKNOWN means needs work; ready means PASS on all fourteen. The checklist makes no legal determination, and CRP-12 runs best against a plan for evaluating RAG systems.

---

How Atlan supports claims retrieval from the context layer

Your claims system, vector index and policy graph stay put; agents read definitions, lineage and policy from Atlan alongside them. The glossary maps business terms “to the exact assets they describe,” and certificates mark an asset Verified, Draft or Deprecated. An agent resolving “open claim” or “reported date” can then prefer the certified definition (CRP-13). Atlan captures column-level lineage automatically.

Personas scope what a team sees and can do. Purposes apply policies by tag: “A deny on PII wins over every grant.”

The Atlan MCP server applies “the same persona, role, and domain policies that govern the UI” to every tool call and logs each call (Atlan MCP security). Its docs list 41 tools across nine categories, including lineage over MCP and the business glossary (Atlan MCP tools).

That enforcement covers the context an agent reads through Atlan. Your own vector index and claims graph still carry CRP-09 and CRP-10.

---

FAQs about knowledge graph vs RAG for insurance claims

1. Is a knowledge graph better than RAG for insurance claims?

Neither wins overall. In a 2026 evaluation on general question sets, RAG did better on detail questions and GraphRAG on multi-hop and temporal ones. With retrieval tokens matched, overall scores were 71.01 and 71.17. Route each claims question to the pattern that fits it.

2. Can RAG alone handle policy endorsements and versions?

Not with similarity alone. Amendments keep most of the earlier text, so an unfiltered search can rank a later edition above the one in force. RAG handles versions when an effective-date filter, run in SQL, index metadata or a graph, narrows retrieval before ranking, and when chunks follow clause boundaries.

3. What should a claims AI agent return when it recommends a denial?

The quoted provision, the form edition and effective interval it came from, the records and graph paths it used, and the access decision. NAIC Model 902, the model regulation for property and casualty claims, bars a denial on a policy provision unless the denial references it; check your state’s adopted text.

4. Is GraphRAG accurate enough for claims decisions?

Its global community summaries suit triage, not evidence. In a 2026 evaluation, they scored 13.95 on unanswerable questions against 91.36 for vector RAG. Local search combines graph data with text chunks of the raw documents. Base adverse decisions on quoted provisions from the governing policy version.

5. How do you stop a claims agent from seeing records it should not?

Authorize before retrieval, then again at every graph hop. In one 2026 multi-agent test setup, retrieve-then-filter pipelines exposed unauthorized context in 86.1% of queries. Another study found that authorizing at the graph-expansion boundary cut measured leakage from vector seeds to zero.

6. Does the EU AI Act treat claims AI as high-risk?

Annex III point 5© lists AI for risk assessment and pricing of natural persons in life and health insurance as high-risk. The point does not name claims handling. Take legal advice on the full Annex III for your use case.

---

Sources

1. Unfair Property/Casualty Claims Settlement Practices Model Regulation, NAIC, July 1997
2. Unfair Claims Settlement Practices Act, NAIC, January 1997
3. NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, adopted December 4, 2023, NAIC
4. Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers, NAIC, status as of October 8, 2026
5. AI Risks and Trustworthiness, NIST AI Resource Center, excerpt from the Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023
6. Regulation - EU - 2024/1689 - EN - EUR-Lex, EUR-Lex, Regulation (EU) 2024/1689 of 13 June 2024, published in the Official Journal on 12 July 2024
7. About the P & C Data Model For Property And Casualty Insurance Specification Version 1.0, OMG, October 2014
8. PROV-O: The PROV Ontology, W3C, W3C Recommendation, 30 April 2013
9. Time Ontology in OWL, W3C, W3C Candidate Recommendation Draft, 15 November 2022
10. Overview - GraphRAG, GraphRAG project documentation
11. GitHub - microsoft/graphrag: A modular graph-based Retrieval-Augmented Generation (RAG) system, GitHub
12. RAG vs. GraphRAG: A Systematic Evaluation and Key Insights, arXiv, v3 4 March 2026
13. From Local to Global: A Graph RAG Approach to Query-Focused Summarization, arXiv, 24 April 2024
14. A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model’s Accuracy for Question Answering on Enterprise SQL Databases, arXiv, 13 November 2023
15. StrucChunk: Structure-Aware Retrieval for Legal Document RAG, arXiv, 19 October 2026
16. TimelyRAG: Semantic-Temporal Hybrid Retrieval for Time-Critical Question Answering in Overlapping-Evolving Documents, ACL Anthology, October 2026
17. Authorization-First Retrieval: Enforcing Least Privilege in Multi-Agent RAG Systems, ACL Anthology, July 2026
18. Retrieval Pivot Attacks in Hybrid RAG: Measuring and Mitigating Amplified Leakage from Vector Seeds to Graph Expansion, arXiv, 9 February 2026
19. Quebec Automobile Insurance Question-Answering With Retrieval-Augmented Generation, ACL Anthology, November 2024
20. Benchmarking Large Language Models for Quebec Insurance: From Closed-Book to Retrieval-Augmented Generation, arXiv, 8 March 2026
21. A Business Graph, Atlan Docs
22. Column Level Lineage, Atlan Docs
23. Add certificates, Atlan Docs
24. Personas, Atlan Docs
25. Purposes, Atlan Docs
26. Atlan MCP security, Atlan Docs
27. Atlan MCP tools, Atlan Docs