Insurance is a paperwork business hiding behind a technology suit. A typical claim may require a phone call, three emails, a police report, a repair estimate, and a dozen boxes of data entered by hand into a 60-year-old database. Working out a commercial risk can mean reading broker submissions of all shapes and sizes. Even a minor address change can hop between two teams.
That's the promise of agentic AI, software that not only responds to questions but constructs a plan of action and executes it across systems. And that's what insurers are interested in - and what exists as a disconnect. Evident's tracker has found agentic AI at 21% of public insurance AI deployments, with 56% of those being in claims. And a survey of 347 European insurers by EIOPA in February 2026 found that autonomous agentic applications remained predominantly in proof-of-concept mode.
Most carriers are in the middle of the curious-cautious spectrum. This post addresses three situations where agentic AI makes sense, what a reasonable agentic AI should be like, and things to consider before commissioning Agentic AI Development Services for any of these situations.
**
What "agentic" means in an insurance setting
**
A copilot proposes. An agent executes. Provide a claim file to a copilot and it generates a summary. Send the same file to an agent and, within parameters you specify, it reads the notice, verifies the policy, detects a lack of repair estimate, emails the insured, records the request and alerts a human when there are questions.
That difference is significant because insurance work is largely a series of small steps through a bunch of different applications: policy admin, claims, document repositories, email, and payments. Agents are designed for chains like that. They also behave differently from copilots in error, because a derelict agent causes an actual error, not just a poor draft. Remember that when reading the rest.
**
Claims intake
**
First notice of loss (FNOL) is when the customer is on edge and the insurance company is disorganized. Phone, app, email, broker; the insurance company receives a communication that isn't usually complete.
The intake agent can extract the essential detail from a call, locate the policy and verify its activeness, determine which items are not yet in the claim file and start the claim with the fields pre-populated. It can provide the claimant with a clean list of remaining evidence with reminders if needed. Simple, low-value claims go into a fast lane. Complex claims arrive at the handler as a neat file, not a mountain of separate emails.
The human boundaries must be clear. Coverage decisions, denials, claim and liability disputes, and anything else that indicates potential fraud must remain human. An agent who gathers the file saves time. An agent who denies a claim that needs no investigation is a ticking regulatory and reputational time bomb.
**
Underwriting support
**
Commercial underwriters devote a disproportionate portion of their day to filling out forms. A broker provides a submission: application forms, loss runs, schedules of locations or vehicles, perhaps a PDF scan of something hand-printed. Prior to the underwriter pricing a risk, someone has to read all of this and transcribe it into the system.
An underwriting support agent will do that preparation. It takes the data, runs it against appetite and guidelines, pulls in external data (property or business data), highlights the gaps, writes a short summary, and includes questions that should be put to the broker. The underwriter opens the file ready for decision, rather than one in the process of being assembled.
The industry is at the beginning of this. In WTW's 2026 survey of 59 property and casualty insurers in the US and Canada, only 16% of firms said they were currently using AI for human underwriting. Sixty percent said they would be prioritizing this technology by 2028. It's all about framing as much as the data: Machines aren't replacing underwriters for tough-to-assess risks.
**
Policy servicing
**
Servicing ranks last and is usually the easiest to start with. Common, rule-based, repetitive requests such as a driver addition, address change, certificate of insurance, billing query or mid-term change are all standard.
The servicing agent can check who the request is from and if they are enabled, run the appropriate change rules, update the policy system, issue the documents and come back to the customer. If the request does not meet the rules it is escalated with the full context so the customer is not asked to explain everything again.
Since the rules are simple and there's a large volume, the outcomes are simple to quantify: how long did it take, how many mistakes and how much of it was carried by a machine instead of a human?
**
The risks worth taking seriously
**
Insurance, as it should be. There are a few things that need to be planned from day one.
Regulation. In the US, increasingly states have followed the NAIC's model bulletin on the use of AI by insurers, which calls for a documented governance programme. In the EU, the AI Act classifies AI used for risk assessment and pricing of life and health insurance as high-risk and introduces additional standards on governance and reporting. You'll want to have the design in front of your compliance team before you code.
Bias and fairness. Agents screening claims or scoring risks may learn patterns from historical data that can be discriminatory. Measure for this before release and periodically afterward.
Auditability. If a regulator or a policyholder questions how a decision was made, you need to be able to reproduce what the agent saw and did. Logging every step is non-negotiable.
Explicit, limited authority. Predefine what the agent is authorized to do alone; what requires approval and what it is never permitted to do. Coverage reporting, limits of payment, and denial authority are clear opportunities for human approval.
**
Where to start
**
Don't begin with the entire claims cycle. Select one specific and narrow flow that has a baseline, such as certificates of insurance, or intake of one personal line, and then quantify it fairly.
It's important because the results have been diverse. For instance, EXL's 2026 insurance survey found that 45% of the agentic AI projects that insurers have launched have been considered successful. And then there's Celent's survey, which reports that 22% of insurers have an agentic AI project scheduled to deploy before the end of 2026. Carriers are in the experimentation stage, which is arguably a reason to go slow.
When you're considering Agentic AI Development Service Providers, ask how they work with your legacy systems, because those are typically the most difficult to interface with. Ask how they test agent behavior pre-launch and how they will keep an eye on things post-launch, and who is accountable if failures occur. And ask them if they're comfortable recommending a pilot project smaller than the project you initially asked for - vendors that are "yes men" might not be the best choice.
**
Choosing a partner
**
No matter if you end up with an in-house team or outsourcing to a company like DianApps, the specs will be similar. You want an insurance workflow expert and technologist who includes governance in the build process and can demonstrate how success is achieved. Watch out for anyone who offers a "one and done" automation of complex decisions on the first release.
**
The bottom line
**
Agentic AI in insurance is new, which works in your favor if you're prepared to tread lightly. Claims entry, underwriting prep and policy service are all recurrent, document-centric, rules-based processes-precisely those things the agent does best.
Maintain human involvement in decisions that have legal or financial implications, document everything, and start with one narrowly focused process. Those insurers who view Agentic AI Development Services as a means of addressing point-of-friction challenges instead of replacing their entire business should see results after just one year.
Top comments (0)