Blog

Agentic AI in Insurance: What Executes, What Only Suggests

Key Takeaways What Is Agentic AI in Insurance? Agentic AI in insurance is software that can take a defined goal and work out what needs to happen next, using the tools and systems available to move the task forward without someone directing every step. A larger goal first has to be broken into smaller tasks, …

| 12 MIN READ

Key Takeaways

  • Agentic AI in insurance falls into two distinct categories: agents that execute a task and write directly to a system of record, and agents that only produce a recommendation for a person to approve.
  • Submission intake, FNOL and claim setup, midterm adjustments, and legacy policy logic can run unattended. Bordereaux ingestion and premium reconciliation are two high-volume workflows that get less attention outside the London market.
  • Human judgement is still required for decisions around complex risk, disputed claims, pricing outside pre-set bands, and fraud referrals, even when an agent has already prepared the case.
  • Authority, audit trail, input quality, and accountability determine how far an agent can go for a given task.
  • Executing agents change the operating model itself, shifting the focus from individual task speed to end-to-end cycle time, capacity, and combined ratio.

What Is Agentic AI in Insurance?

Agentic AI in insurance is software that can take a defined goal and work out what needs to happen next, using the tools and systems available to move the task forward without someone directing every step.

A larger goal first has to be broken into smaller tasks, a process known as goal decomposition. From there, an agent might query a database, call an API, or pull information from a document. It can also take action across systems, such as updating a record, routing a file, or triggering the next step in a workflow.

However, not every action should happen automatically. An AI agent might complete one step on its own and reach a point later in the workflow where a person needs to review or approve what happens next. A system that recommends an action and one that has the authority to execute it are different products, and they need different controls.

AI agents in insurance have plenty of room to grow. In The State of Insurance Operations, an August 2026 survey of 150 insurance leaders that mea commissioned from ISG, 78% said they use production AI for incremental efficiency gains or limited automation. Only around 13% described an advanced AI posture today, meaning workflows redesigned around AI or AI-native operations, while 52% intend to get there within two years.

The Difference Between an Agent That Executes and One That Recommends

The model itself doesn’t tell you which mode you’re in. Two agents can run on the same architecture, reason over the same data, and produce outputs that look identical on screen. The difference comes afterwards. One may be allowed to write that output directly to a system of record, while another has to send it to a person for approval.

That comes down to authority. An agent might be perfectly capable of drafting a coverage decline, for example, without having permission to issue it. The same applies to actions like binding a policy or releasing a payment. The model can figure out only part of what determines how far it can act. Its permissions determine what it can actually change without a person in between.

You can see the difference in common insurance workflows:

WorkflowRuns unattendedStill needs a humanWhat sets the line
Submission intakeExtract, structure, write to systemAppetite fit call, outside of predefined rulesInput quality
Clearance and sanctionsMatch, flag, clearAdverse match reviewTolerance for false negatives
Bordereaux ingestionParse, normalize, reconcileException handlingCover holder data consistency
FNOL and claim setupIntake, coverage check, reserve suggestionReserve approvalReserving authority
EndorsementsParse request, recalculate, draftBind the changeBinding authority
Premium reconciliationMatch, allocate, postDisputed balancesLedger write permissions
Risk selection and pricingAssemble the evidence packThe decision itselfAccountability, not capability

The rightmost column is what matters most. Input quality, permissions, risk tolerance, and accountability all affect how much authority an agent can be given. So even when the technology can carry a task further, an insurer may choose not to give it that authority. Where execution stops and human judgement begins is as much a governance decision as a technical one.

Where Agentic AI Already Executes in Insurance Operations

Four workflows come up most often:

  1. Submission intake: extracting and structuring broker submission data and writing it into the underwriting system.
  2. FNOL and claim setup: handling intake, checking coverage, and suggesting a reserve before a handler opens the file.
  3. Midterm adjustments: parsing a policyholder or broker request, recalculating the premium impact, and drafting the endorsement.
  4. Legacy policy logic: reverse-engineering undocumented rating and eligibility rules from old batch systems.

Two other high-volume workflows get less attention, especially outside the London market:

  1. Bordereaux ingestion: parsing inconsistent MGA and cover holder files, normalizing them to a standard schema, and reconciling them against the books.
  2. Premium reconciliation: matching, allocating, and posting entries against bank statements and broker remittances, with unmatched balances flagged for review.

ISG’s survey asked leaders how far AI should go across 20 insurance activities over the next 24 months. Repeatable work drew the strongest appetite for end-to-end AI: 28% for standard-risk quote generation, 26% for claims data extraction and validation, and 19% for submission intake, triage, and classification. For bordereaux processing and FNOL intake and triage, the figure was 15%. These are stated preferences rather than deployment figures.

mea Operations is built to execute operational work across underwriting, claims, finance, and broking, working across the systems teams already use and bringing people in when an exception requires their attention.

Where Agentic AI Still Only Suggests

Even after an agent extracts the data, checks coverage, and assembles everything a person would need, some decisions still need to be made by a human. These include binding or declining complex risk, coverage calls on borderline claims, pricing moves outside pre-set bands, and fraud referrals to SIU.

How far the agent goes depends on its authority and who remains accountable for the outcome.

ISG’s survey shows the same split. 83% of insurance leaders support AI for repeatable operational work, while 86% believe people should retain ownership of consequential decisions. For complex or contentious claims adjudication, 28% want no AI involvement at all and only 5% want it end to end.

An agent’s authority determines what it has permission to change in a system of record. Input quality can narrow the agent’s room to act, particularly when a workflow relies on unstructured documents that are incomplete or difficult to interpret. mea’s technology handles submission ingestion, validation, and triage as part of that workflow.

Underwriting Judgement and Risk Appetite

Appetite matching is one of the clearer cases where an agent’s output stays a recommendation rather than a decision. It can compare a submission against the insurer’s stated appetite, including class of business, limits, geography, and loss history. Some risks won’t fit neatly within those guidelines, though, and pricing may still require an underwriter’s judgement.

Binding agreements and MGA arrangements already set out which decisions need to be referred for approval. An agent working within that delegated authority can’t go beyond what the person it’s supporting is authorized to decide.

With mea Underwriting, the appetite and pricing work happens before the case reaches the underwriter. The underwriter still makes the final risk decision.

Complex and Disputed Claims

Coverage disputes, reserve-setting on uncertain files, and litigation-exposed claims are areas where an agent’s role is limited to preparation. It can pull the policy wording, claim history, and relevant correspondence and flag where the facts and coverage terms may conflict.

The decision itself stays with an adjuster or counsel. A reserve might change as new information comes into the file, or the outcome could turn on how particular policy wording applies to the loss. The agent can make that review faster by surfacing the information that matters. It doesn’t decide what the claim is worth or how the dispute should proceed.

How Do Executing Agents Change the Operating Model?

Removing a step from the operating model means it no longer exists as a separate piece of work. When submission intake feeds directly into what happens next, there is no re-keying step left to speed up.

For executives, cycle time is more useful when measured end to end, such as submission to bind or FNOL to reserve, rather than around a single task. The question is how much more volume the insurer can handle without growing the operations team at the same rate. Combined ratio takes the view beyond processing speed, showing whether operational changes are translating into better underwriting and expense outcomes.

ISG’s research points to the same shift. 86% of insurers using BPO providers said they would renegotiate or completely rework terms if contracts were renewed today. Against that backdrop, ingestion and triage make more sense as tasks within an operations suite than as standalone products bolted onto the existing process.

mea’s BPO solutions pair mea’s agents with operations teams for insurers that want the work run as a managed service. mea also works with a top-five global business processing outsourcer.

Running (Re)Insurance Operations on Agents That Execute

mea runs execution across underwriting, claims, and broker operations. In underwriting, that includes extracting and structuring the submission data before it reaches the underwriter. In claims, FNOL intake and coverage checks can happen before a file reaches an adjuster. On the broker side, mea for brokers handles work from client document ingestion through validation, triage, and preparation for placement.

A global P&C insurer’s documented rollout shows what that looks like in practice. The insurer moved from unstructured submission documents to structured data across its US lines of business, with the rollout completed in less than three months. It now extracts data from submissions across its US, APAC and EMEA lines, processing thousands every day. In mea Underwriting deployments, 90% of transactions are processed automatically and quote to bind takes one to two hours.

Get started to see where agentic execution could take work off your operations team.

Frequently Asked Questions

What is the difference between agentic AI and generative AI in insurance?

Generative AI produces an output when it’s asked, such as a summary of a claim file, a draft email, or a first pass at policy wording. Agentic AI pursues a defined goal across multiple tools and steps, without a person prompting each one along the way.

A generative system produces something a person reads and decides what to do with. An agentic system can change a record directly, for instance by updating a claim file or writing structured data into the underwriting system, without a person necessarily reading it first.

What is the difference between an AI agent and robotic process automation?

RPA follows a fixed script against a predictable interface: click here, copy this field, paste it there, in the same order each time. An agent works differently. Given a goal, it can work out which steps to take and which tools to use based on what it finds along the way.

RPA works best when the layout, fields, and locations stay consistent, so an unexpected format in a broker submission can interrupt the process. An agent can interpret the document, identify the information it needs, and adjust its next steps based on what it finds.

What is delegated authority, and why does it limit what an AI agent can do?

Delegated authority is the contractual right to bind coverage on another party’s behalf. A cover-holder may hold it under a binding agreement, or an individual underwriter may have it within limits their employer has authorized.

An agent can’t inherit that authority because it belongs to the person or firm named in the contract, not the tool they use. The agent can assemble the file, apply the rating logic, and prepare everything needed to bind. The actual bind still has to come from someone who holds the authority to make it.

Which regulators have published guidance on insurers’ use of AI?

In the US, the NAIC adopted a Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. At the EU level, EIOPA published an Opinion on Artificial Intelligence Governance and Risk Management in 2025. NAIC market surveys published in 2022 and 2023 found that 88% of responding US auto insurers and 70% of responding home insurers use, plan to use, or plan to explore AI or machine learning models. In the UK, the FCA and the PRA have said they will oversee AI through their existing rules, and the FCA’s Consumer Duty applies to customer outcomes that AI supports.

Do agentic AI systems need an insurer’s historical data to work?

Not necessarily. A system pre-trained on insurance data and workflows can already recognize common insurance documents, terminology, and structures before it sees an insurer’s own files. A system that relies on the insurer’s own corpus needs more work upfront to learn the documents and workflows it will encounter.

A pre-trained insurance system starts with much of the groundwork already in place, reducing the amount of preparation needed before it can begin producing usable results.

How quickly can an agentic AI system reach production?

For a system already pre-trained on insurance documents and workflows, production can take anywhere from a few weeks to a couple of months. The timeline depends on the work involved, including how many existing systems it needs to connect to, the validation required, and the exception paths that need to be defined. Systems that need additional training or configuration around an insurer’s documents and processes will take longer, since that groundwork must happen before the system is ready for production.

Cut the drag. Just win.

Insurance AI that shows up on your bottom line.