AI Agents, Workflows, and Knowledge Graphs in Insurance (2026 Guide)

TL;DR: AI agents can read a submission, pull the loss runs, check the account history, and hand an underwriter a triaged file in minutes instead of hours. But they only do that reliably when they sit on top of a connected model of your business, not a pile of disconnected systems. In insurance, the same insured, broker, and claim show up differently in your policy admin system, your claims system, and an inbox full of PDFs. A knowledge graph resolves those into one entity and records how every answer was derived, which is what makes agent output trustworthy and audit-ready. Agents do the work, agentic workflows sequence it, and the graph connects it. Skip the graph and you get a fast demo that stalls in production.

What do AI agents, workflows, and knowledge graphs mean for insurance?

Insurance runs on context that lives in too many places. A single commercial account touches a policy admin system like Guidewire or Duck Creek, a separate claims platform, an agency system such as Applied Epic, an actuarial or pricing model, and a broker email thread with ACORD forms and loss runs attached as PDFs. The work of underwriting, claims, and renewal is largely the work of assembling that scattered context into a decision. That is why AI agents in insurance are compelling: the highest-cost activity in the shop is not the decision itself but the hours spent gathering the information the decision depends on.

Three technologies are usually discussed together, and they are not interchangeable. An AI agent is software that can take a goal, plan steps, call tools and systems, and produce an output, such as a triaged submission or a claims summary. An agentic workflow is the orchestration layer that sequences those agents across a real business process with checkpoints and handoffs to people. A knowledge graph is the connected data model underneath, where insureds, brokers, policies, claims, and documents exist as resolved entities with explicit relationships between them. The graph is what an OutcomeCatalyst-style AI context layer for insurance provides: a company brain that agents can reason over instead of guessing across silos.

Key takeaways

• Context, not model quality, is the bottleneck. Frontier models are already good enough to read an ACORD form or a loss run. What they lack in most carriers is a reliable, connected view of the account they are reasoning about.

• Insurance and banking lead enterprise AI-agent adoption. Financial services sits at the front of the deployment curve, which means your competitors are moving and the playbook is being written now.

• A knowledge graph is the foundation, not a feature. Entity resolution, explicit relationships, and data lineage are what turn a plausible-sounding agent into one an underwriting or claims leader can defend in an audit.

• Explainability is a regulatory requirement, not a nicety. A graph records where each answer came from, so you can produce lineage when a regulator or reinsurer asks how a decision was reached.

• You do not need a two-year platform project. The fastest path is to connect a few systems around one high-friction workflow, prove it, then expand, rather than boiling the ocean with a data warehouse rebuild.

Why most insurance AI projects stall

The common failure is not the model. It is that the model has no dependable picture of the business. Celent has noted that roughly 80% of policy-level data is trapped in unstructured documents such as submissions, binders, quotes, and loss runs, arriving as PDFs and fragmented Excel worksheets rather than clean records. When a pilot works on a curated sample and then breaks on live traffic, the reason is almost always that the real data was messier and less connected than the demo assumed. The agent could read the document; it just could not tell that the "ABC Logistics" on the submission was the same insured already sitting in the claims system under a slightly different legal name.

The appetite is real and the leaders are pulling ahead. McKinsey estimates generative AI could add $50 billion to $70 billion in value across the insurance industry, and points to concrete results such as Aviva deploying more than 80 AI models in claims, cutting liability assessment time on complex cases by 23 days and reducing customer complaints by 65%. On the adoption side, Deloitte reports that 57% of insurers were using AI in 2025 and predicts that half of enterprises using generative AI will deploy agentic AI by 2027, up from a quarter in 2025. The gap between the firms getting value and the firms stuck in pilots is rarely the algorithm. It is whether the AI has a trustworthy, connected context layer to stand on.

The three technologies, and how they fit together

It helps to be precise about which layer does what, because vendors blur the terms and buyers end up paying for one when they needed all three.

1. AI agents: doing the work

An agent is the unit that performs a task end to end. Give it a goal such as "triage this submission" and it plans, retrieves what it needs, calls the systems it is allowed to touch, and returns a result. In insurance the useful agents are specific: a submission-reading agent that parses an ACORD 125 or 140 and the attached loss runs, a data-extraction agent that normalizes a broker's bordereaux, a claims-summary agent that assembles adjuster notes and prior claims into a briefing, a fraud-signal agent that flags patterns across linked claims. An agent on its own is powerful but narrow. It is only as good as the context it can reach and the boundaries it is given.

2. Agentic workflows: sequencing the work

Real insurance processes are multi-step and involve people, so a single agent is rarely the whole answer. An agentic workflow chains agents together and inserts control points: read the submission, resolve the insured against the book, pull the loss history, score against appetite, then route to a human underwriter with a recommendation and the evidence behind it. The workflow decides what runs, in what order, what requires human sign-off, and what happens when confidence is low. This is where agentic AI workflows in insurance move from a clever tool to an operating change: the process itself becomes faster and more consistent while a person still owns the decision that carries regulatory and financial weight.

3. Knowledge graphs: connecting the work

The graph is the connected model the agents reason over. Instead of an agent querying five systems and hoping the records line up, the knowledge graph for insurance holds each insured, broker, policy, location, and claim as a single resolved entity, with the relationships between them made explicit: this broker placed this policy for this insured, which has these three open claims and this loss history. When the workflow asks "what is our total exposure to this insured across all lines," the answer comes from one connected model, not a fragile join across systems that use different keys. This is the layer that is usually missing, and its absence is why agents that dazzle in a demo become unreliable in production.

AI agents across the insurance workflow

The clearest way to see the value is workflow by workflow, using the decisions your teams actually make.

Underwriting submission intake and triage

Submissions arrive as email with PDF ACORD forms, supplemental applications, and loss runs, often for accounts with hundreds or thousands of locations spread across fragmented spreadsheets. Underwriters spend hours reading, keying, and assembling context before they can even decide whether a risk is worth quoting. An AI underwriting workflow reads the submission, extracts and structures the exposure data, resolves the insured and broker against your existing book, pulls prior loss history, checks the risk against your appetite and referral rules, and delivers a triaged file with a clear-to-quote recommendation or a specific reason to decline. McKinsey has documented cases where this kind of acceleration cut commercial quoting from days to hours. The underwriter still decides; they just start from an assembled file instead of an empty desk. This is the core of our underwriting triage work.

Claims adjudication support

AI claims processing agents do not replace the adjuster; they compress the assembly work around the decision. A claims-support agent gathers the first notice of loss, adjuster notes, policy terms, prior claims for the same insured, and relevant documents into a single briefing, checks coverage against the actual policy language, surfaces inconsistencies and possible fraud signals across linked claims, and drafts the correspondence. Because the agent works from a connected model, it can see that a claimant appears on two other claims under a different spelling, a link a siloed system would miss. The adjuster reviews a complete, evidence-backed picture and adjudicates faster. See our approach to claims adjudication support.

Audit and regulatory readiness

Insurance decisions have to be explainable, and regulators, reinsurers, and internal audit increasingly ask not just what was decided but how. When agents operate over a knowledge graph, every answer carries its lineage: which document, which policy version, which prior claim fed the recommendation. That turns audit preparation from a fire drill into a query. Instead of reconstructing a decision months later from email and memory, you can show the chain of evidence directly. This is a decisive advantage over black-box AI, and it is central to our audit and regulatory readiness work.

Renewals and portfolio review

Renewal season is a context-assembly problem at scale. For every account, someone needs the loss experience, exposure changes, prior underwriting notes, and open claims, and someone needs to see the portfolio view to catch concentration and adverse selection. An agentic workflow can pre-assemble every renewal file, flag accounts where loss experience has deteriorated or exposure has grown, and give underwriting leaders a portfolio-level read across the book. Because the graph already resolves entities across systems, aggregate exposure to a given insured or industry is a question you can answer on demand rather than a quarterly spreadsheet exercise.

Why a knowledge graph is the foundation, not a nice-to-have

Three properties of a knowledge graph are what separate insurance AI that survives production from AI that does not.

• Entity resolution. The same insured, broker, or claimant appears with different names, IDs, and spellings across your PAS, claims system, agency management system, and email. The graph resolves these into one entity so an agent reasons about the real account, not a fragment of it. Without this, every downstream answer is quietly wrong in ways that are hard to catch.

• Relationships. Insurance risk lives in the connections: which broker placed which policy, which claims belong to which insured, which locations sit under which program, how a group of claims relates to a single event. A graph stores those relationships as first-class facts, so questions about exposure, concentration, and fraud become traversals of the model rather than brittle multi-system joins.

• Lineage and explainability. Every fact in the graph can carry its source and its derivation, so an agent's recommendation comes with a traceable chain back to the document, policy version, or claim it relied on. Tie this directly to regulatory audit: when someone asks how a decision was reached, you produce the lineage instead of reverse-engineering it. This is the difference between AI you can deploy in a regulated line and AI you cannot.

Knowledge graph vs. data warehouse vs. RAG for insurance

These three often get pitched as alternatives. They solve different problems, and understanding the difference saves you from buying the wrong thing.

• Data warehouse. A warehouse aggregates structured data for reporting and analytics. It is excellent for "how did the book perform last quarter." It is not built to resolve entities across messy operational systems in real time, and it does not model the relationships an agent needs to reason about a single live account. It answers questions about the past in aggregate, not questions about this insured right now.

• RAG (retrieval-augmented generation). RAG retrieves relevant text chunks and feeds them to a model. It is genuinely useful for document questions, such as finding a clause in a policy. But RAG retrieves passages, not resolved facts, so it cannot reliably tell you that two documents refer to the same insured or compute total exposure across systems. Gartner has noted that structured knowledge layers materially reduce AI error rates compared with standard RAG. RAG on its own answers "what does this document say," not "what is true about this account."

• Knowledge graph. The graph is the connected, resolved model of your business. It gives agents entities, relationships, and lineage, and it complements the other two: it can sit alongside a warehouse and can dramatically sharpen RAG by grounding retrieval in resolved entities. For agentic insurance workflows, it is the layer that makes the output trustworthy.

How to build this without a two-year platform project

The instinct to rebuild the data estate first is what kills momentum. A better path is narrow, fast, and additive.

1. Pick one high-friction workflow. Start where the pain and the volume are obvious, usually submission triage or claims summarization. One workflow gives you a measurable before-and-after and a contained blast radius.

2. Connect only the systems that workflow touches. For triage that might be your submission inbox, your PAS, and your loss-run source. You do not need every system on day one; you need the ones this decision depends on. The context layer connects to them where they are rather than forcing a migration.

3. Build the graph as you go. Resolve the entities and relationships this workflow requires, insured, broker, policy, claim, and let the graph grow with each workflow you add rather than modeling the whole enterprise up front.

4. Keep a human in the loop and instrument everything. The agent recommends; a person decides on anything that binds risk or pays a claim. Log the lineage from the start so explainability is built in, not retrofitted.

5. Measure, then expand. Track time-to-triage, quote turnaround, or claims cycle time, prove the lift, and reuse the same connected model for the next workflow. Each new use case gets cheaper because the graph is already there.

What good looks like: a walkthrough

Consider a commercial property submission that lands in a shared underwriting inbox at 9 a.m. It is an email from a broker with an ACORD 140, a supplemental application, and a loss run covering the last five years, plus a spreadsheet listing 340 locations across three worksheets. Under the current process, an underwriting assistant will spend the better part of the morning keying exposure data, looking the insured up in the PAS, and requesting a loss history before an underwriter even sees it.

With an agentic workflow on a context layer, the sequence runs differently. The intake agent reads the email and every attachment, extracts the 340 locations and their construction, occupancy, protection, and exposure details, and structures the loss run. The graph resolves the insured against the existing book and recognizes that this account has two prior submissions and one open claim from a related entity that the broker did not mention. A scoring agent checks the risk against appetite and referral rules and notes that total combined exposure to this insured across lines now exceeds the concentration threshold for that region. By 9:20 a.m. the underwriter opens a triaged file: structured exposure data, resolved account history, the concentration flag, a clear-to-quote-with-referral recommendation, and a lineage trail showing exactly which documents and records produced each conclusion.

The underwriter still makes the call. But they start from evidence instead of a stack of PDFs, they catch a concentration risk that manual triage would likely have missed, and if that decision is ever questioned in an audit, the reasoning is already documented. That is the shape of the win: not a decision made by a machine, but a decision made faster and better because the context was assembled and connected before a person had to think about it.

Common mistakes that sink insurance AI initiatives

• Buying a model and skipping the context layer. The demo works on clean samples, then production traffic exposes that the AI has no reliable, connected view of the account. This is the single most common way insurance AI stalls.

• Trying to build the enterprise data warehouse first. The two-year platform rebuild burns budget and goodwill before delivering anything. Connect around one workflow and expand instead.

• Ignoring entity resolution. If the same insured or claimant is not resolved across systems, every aggregate and every cross-reference is quietly unreliable, and the errors are hard to detect.

• Treating explainability as a later phase. In a regulated line, lineage you cannot produce is a decision you cannot defend. Instrument it from the first workflow.

• Automating the decision instead of the assembly. The value and the safety both come from compressing context assembly while keeping a human on anything that binds risk or pays a claim. Full autonomy on regulated decisions is where governance gaps become losses.

• No governance for the agents. Deloitte notes only about one in five organizations has a mature model for governing autonomous agents. Define what agents may touch, what requires sign-off, and how confidence thresholds route work, before you scale.

Frequently asked questions

What are AI agents in insurance?

AI agents in insurance are software that takes a goal, such as triaging a submission or summarizing a claim, then plans the steps, reads the relevant documents and systems, and produces a result. Useful agents are specific to a task, such as parsing ACORD forms and loss runs or checking coverage against policy language, and they work best when they can reason over a connected model of your accounts rather than querying siloed systems.

What is a knowledge graph in insurance?

A knowledge graph is a connected data model where insureds, brokers, policies, locations, and claims exist as single resolved entities with explicit relationships between them. It lets you answer questions like total exposure to one insured across all lines as a traversal of one model, instead of a brittle join across your PAS, claims system, and agency management system.

How is AI used in underwriting?

AI underwriting workflows read incoming submissions and attachments, extract and structure exposure data, resolve the insured and broker against your book, pull prior loss history, and score the risk against appetite before routing a triaged file to an underwriter. McKinsey has documented cases where this cut commercial quoting from days to hours. The underwriter still makes the decision from an assembled, evidence-backed file.

How does AI help with claims processing?

AI claims processing agents assemble the first notice of loss, adjuster notes, policy terms, and prior claims into a single briefing, check coverage against policy language, and surface inconsistencies or fraud signals across linked claims. Because the agent works from a connected model, it can catch that a claimant appears on other claims under a different spelling, which speeds adjudication while the adjuster keeps the decision.

What is an AI context layer for insurance?

An AI context layer for insurance is the connected model that sits between your systems and your AI agents. It links your policy admin system, claims system, agency management system, and document sources into one knowledge graph so agents reason over resolved entities and relationships with full lineage. It is what makes agent output reliable and audit-ready rather than plausible but unverifiable.

Is a knowledge graph better than RAG for insurance AI?

They solve different problems and work best together. RAG retrieves text passages and is good for document questions such as finding a clause. A knowledge graph provides resolved facts, relationships, and lineage, which RAG alone cannot, and grounding retrieval in the graph reduces errors. Gartner has noted that structured knowledge layers materially lower AI error rates compared with standard RAG.

How long does it take to implement AI agents in insurance?

A focused deployment does not require a multi-year platform project. By starting with one high-friction workflow, connecting only the systems it touches, and building the graph incrementally, carriers can stand up a working, measurable use case in weeks to months, then reuse the same context layer for the next workflow at lower marginal cost.

The bottom line

The insurers pulling ahead are not the ones with the best model. They are the ones whose AI has a trustworthy, connected view of the business to reason over. Agents do the work, agentic workflows sequence it, and a knowledge graph connects it, and the graph is the piece most projects skip and most projects fail without. If you get the context layer right, the same foundation compounds across underwriting, claims, renewals, and audit rather than fragmenting into one-off pilots.

If you want to go deeper, start with what an AI context layer is and how knowledge graphs power enterprise AI. For how this varies by sector, see our overview of AI implementation by industry and the parallel playbook in AI agents and knowledge graphs in private equity. When you are ready to map this to your own systems, our insurance practice covers underwriting triage, claims adjudication support, and audit and regulatory readiness end to end.

Sources

• McKinsey, The future of AI in the insurance industry

• McKinsey, Reimagining insurance with a comprehensive approach to gen AI

• Deloitte, Agentic AI in banking and financial services

• Deloitte, AI in Financial Services

• Celent, Commercial Lines Underwriting: Data, AI, and Automation

• CIO, Knowledge graphs: the missing link in enterprise AI