Are AI Agents Ready to Serve?

Meet Friendly John.

If you need a homeowners or auto insurance policy in the United States, he’s your bot. Gainfully employed at Boca Raton, Florida-based e-brokerage Nsure, Friendly John will ask questions, compare more than 50 carriers’ policies, and suggest the best one for the client’s needs. He’s almost, but not quite, full service: a human agent must bind the policy.

“When you go on to our website and you want to run a quote, the process that’s taking place in the background is through the Friendly John Rater; those AI agents are…essentially running all of your quotes,” says John Haisch, Nsure vice president of automation and AI. “They’re collecting based on your data, going to carriers, processing the quotes, and then bringing them back into the system. You’ll have all of your insurance quotes, bindable, real quotes in anywhere from 10 to 15 minutes from a dozen or more carriers.”

Friendly John is at least adjacent to being an AI agent: software systems that autonomously perform tasks by interacting with users, systems, and data, says Luis Blando, chief product and technology officer at low-code development platform OutSystems. AI agents apply large language models (LLMs), machine learning (ML), natural language processing, and automation to conduct various actions, ranging from answering questions to offering recommendations, Blando says. Through a process of testing and analysis, the systems can take on increasingly complex challenges and operate more autonomously, to benefit corporate efficiency, decision-making, and user satisfaction, he adds.

Many IT providers tout agentic AI as the next iteration of development and functionality for generative artificial intelligence. Sources for this article say 2025 will likely prove decisive for AI agents in the insurance industry and beyond. This could be the year that these systems reach levels of reasoning, flexibility, comprehensiveness, and logic that will make them indispensable partners in addressing many basic insurance tasks. A key question will be to what degree and how fast AI agents can move up the value chain to the more complex functions that are required for commercial lines.

“Where we see the industry right now is in an experimentation phase with AI agents and absolutely coming to production and scale in 2025,” says John Duigenan, distinguished engineer, U.S. financial services industry, at IBM, one of the software giants designing LLM platforms required for agentic artificial intelligence. “It will scale massively this year and be a key theme.”

Some go further and say AI agents in 2025 may begin to take on the actions and reasoning needed to carry out simpler job functions and product lines— potentially making them competition for junior agents and brokers.

“I think the best way to think about AI agents is like an intern,” says Rotem Iram, co-founder and CEO of San Francisco-based, AI-enabled cyber insurance carrier At-Bay. “You can bring them in, give them simple tasks, provide a little bit of supervision, and they can get a lot of the dirty work out of the way. But like any intern, they will get better in time, and in a year they can become a junior producer; in two years they can be a senior producer; and probably in four to five years they can run the whole thing.”

The insurance industry is only beginning to consider how agentic AI might be applied, cautions Lindsey Klarkowski, policy vice president— data science, artificial intelligence, and cybersecurity at the National Association of Mutual Insurance Companies. “In terms of use of AI agents in the insurance ecosystem, given that agentic AI potential and viability is only starting to be explored in the big tech spheres, there isn’t necessarily a trend to point to currently in the insurance industry context.”

Agentic AI’s progression could also be slowed by regulatory limits and trust issues by consumers, others say.

It’s an Agentic AI World

Agentic AI appears to be the next evolution in adoption of generative AI, with use cases to match, according to Klarkowski.

Major tech players seem to concur, with Microsoft, OpenAI, Google, and others announcing updates on their use of the technology in late 2024.

The enthusiasm was evident at the Insuretech Connect Conference last October in Las Vegas (ITC Vegas), discussed in workshops and on the exhibition hall floor, notes Kacie Conroy, vice president of information technology at M3 Insurance, a national brokerage that is exploring deployment of agentic AI tools.

“When we see a theme, we pay attention to it,” Conroy says. Agentic AI is likely to follow a path set by the Internet, which initially seemed of ambiguous value to insurance businesses but ultimately became essential, she adds.

Industry sources offered differing opinions over whether AI agents are a discrete category of artificial intelligence, the next stage in a gradual evolution of AI capabilities, or, more cynically, a marketing term to make AI seem new and hot again.

“Agentic AI appears to be the next step or phase of generative AI, in that it combines classic automation with generative AI—LLMs, specifically,” Klarkowski says.

IBM’s Duigenan believes agentic AI represents the third, and most advanced, of three levels of artificial intelligence. Chatbots able to perform simple data retrieval were the first level, he says, followed by today’s state of the art—AI assistants that can perform chatbot tasks along with associated assignments, such as routing customers to another automated IT function to pay a bill or perform other transactions, though often with human assistance.

The third level, tested in late 2024 but expected to be deployed at scale by 2025, is agents, which can orchestrate different AI functions to autonomously perform complex multistep processes, Duigenan says.

Here, an AI agent might use tools to conduct an entire planned course of action. Human involvement could be limited to approving that approach.

A key measure of whether an AI system has reached agentic capabilities is whether it can essentially reason for itself upon encountering a problem or something beyond the data on which it has been trained. That can include knowing that it needs to run another AI application, search databases, or consult with a human to provide it additional information or guidance, to fully understand all the dimensions of a human request, solve problems that arise, and progress further toward completing a larger task, Duigenan says.

“An AI agent is an intelligent piece of software that can do things independently and can basically self-determine what it’s going to do,” adds Pete Miller, president and CEO of The Institutes. “It can make decisions about how to branch and do different things with an outcome-based approach. We think these are going to progress pretty rapidly.”

Agent Applications

AI agents may be used in various ways within the insurance ecosystem, according to sources for this article. IBM, for example, is examining four use cases for AI agents for carriers, Duigenan says: sales and distribution, underwriting, servicing, and claims.

“Within those four domains, we’ve looked for all of the places where an overlay of AI assistants can assist either customer or agent experience,” he says. “For example, it could assist an underwriter by aggregating information in documents that are prepared and generating outcomes. Similarly, for policy clauses in servicing, it could respond to customer questions such as, ‘Can I make a payment,’ or ‘Have I made a payment?’ ‘Is my policy up to date?’ and so on.”

But Duigenan says that as of late 2024 it was too early to say exactly how artificial intelligence agents that reason and learn could safely be applied to integrate these AI assistant processes with larger tasks without better understanding of risk factors.

Some say their products or services already employ AI agents. Applied Systems, a leading software vendor to the insurance industry, is scheduled in spring 2025 to release Applied Book Builder, which will use agentic AI technology to help customer service representatives (CSRs) review commercial policies to be renewed, update client risks and exposures, and suggest coverages that are not included in current policies and placements, says Chief AI Officer Elad Tsur.

“That recommendation is the outcome of multiple AI agents that communicated: an agent that understands the current coverage, including the legal language of policy documents, reading documents with hundreds of pages of legal wordings, understanding what coverages are included and what coverages aren’t; multiple additional agents that all collaborate to take public web data and understand what coverages are needed for a business; and another AI agent that makes the recommendation decision,” Tsur says. “Of course, these AI agents don’t replace CSR judgment, but instead enhance their productivity and accuracy.”

Gallagher Bassett has several AI systems that it plans to further refine and integrate with the help of an AI agent, says Chief Digital Officer Joe Powell. They include a system used in North America that can quickly and briefly summarize hundreds of pages of documentation from a claim file and a customer service application for workers compensation lines that was rolled out in July in Australia.

“We’re building skills that currently serve a single purpose and that can be built together to create powerful multiagent AI,” he says. Eventually, a master agent will orchestrate all of those individual AI assistants, according to Powell.

Underwriting And Risk Mitigation

AI agents will be invaluable to expediting completion of less complex underwriting tasks and increasing the granularity at which various risks can be measured and thereby more accurately underwritten or mitigated, some interviewed said.

“The use cases I see for AI agents [at At-Bay] are, first, basically automating the back office jobs of both underwriting and claims,” Iram says.

That can mean eliminating many routine, mundane aspects of underwriting by automating underwriting that already relies on formulas and tables.

“I would say that the average underwriter probably spends the majority of their time doing tasks that do not require a lot of judgment,” Iram says. “So what we are doing is we are leveraging gen AI to take away the 70% worst part of an underwriter’s job, which allows our underwriters to spend all of their time making judgment calls, working with brokers and with customers, winning deals, building relationships, crafting coverage requests, spending more of their time on the more complex accounts, and spending less time on simple accounts.”

The second big job [for AI agents] is around risk evaluation, quickly finding hard-to-locate signals in massive data sets. AI agents are particularly useful for cyber risk analysis, Iram notes, given that the risks and possible data points are constantly expanding and changing as hackers change their infiltration strategies.

More capable AI agents supporting underwriting and risk analysis could expand the number of services insurers provide, freeing up humans to assist clients to reduce their risk profiles and thereby their premiums, says Iram.

Elsewhere, Pleasanton, California-based cyber insurance provider Cowbell is focused on using AI agents on the customer and broker-facing submission side, says co-founder and Chief Product Officer Rajeev Gupta. In January, Cowbell released an AI agent to largely automate submission handling.

“The end goal is to make it seamless for agents and brokers, utilizing AI, to ingest and extrapolate submission data received via email, review attachments, and assemble the necessary information to create a submission record and generate an automated quote,” Gupta says. “The agent or broker then receives an email in a matter of minutes requesting additional information or providing a fully bindable quote.”

Agency and Broker Uses

Wholesale brokerages can employ AI agents to reduce costs for servicing smaller accounts that have previously not been profitable, says Sivan Iram, co-founder and CEO of Mountain View, California-based, AI-driven brokerage Flow Specialty [Editor’s note: Sivan Iram is the brother of Rotem Iram].

“The challenge that [retail] agents have today with smaller accounts is that they are being ignored by the traditional wholesaler brokerages because they are not profitable,” he says. “A $2,000 policy is not profitable for a wholesaler.”

For a $10,000 premium account, a typical wholesaler would have a negative margin of 16% given the time to manage an account that would ultimately yield a $1,000 commission. Consider also that just one of four deals closes, Sivan Iram adds.

Flow, however, can provide wholesalers a positive 65% margin on those deals, the CEO says. Two aspects of agentic AI enable such efficiencies, he explains.

“First, we employ AI agents to help streamline our processes and do them faster: things like data extraction, taking unstructured data formats like emails and structuring them into downstream systems,” Sivan Iram says. “That’s the low-hanging fruit, but to me the second bucket is the more exciting long-term AI use case, which is digitizing human knowledge and being able to scale up knowledge and expertise. So, for example, every time that we get a quote from one of our carriers, we do a thorough analysis of the coverage and we provide a recommendation to the insured using our AI agent.”

Such capabilities will edge AI agents toward the type of judgment calls that humans make through years of insurance industry experience, such as “gut instinct” calls and more nuanced deductions that, say, a typical solution will not apply to a particular customer’s risk situation or business model, Sivan Iram says.

These capabilities will also become pervasive as AI agents’ price point comes down. “These AI agent capabilities were not available two or three years ago in part because of cost,” he notes. “The promise of technology is not only in the tip of the spear of technological capabilities; it’s also how quickly it can be made accessible and democratized.”

Other brokerages are moving in the same direction. M3 Insurance is considering implementing agentic AI to make interactions with clients better and faster, Conroy says.

“Data entry is a really big area that we and a lot of brokerages are looking at,” she says. “There’s a lot of information that comes to us from the client to the carrier and we’re the translator to the clients and translating things back to the carriers. So how can we use AI to help streamline those data entry and materials that you’re providing for carriers, materials you’re providing for your clients, in a more automated way? No one wants to be a keyboard warrior and do all the data entry. They want to receive the information and be a consultant working to provide value to our carriers and our customers.”

M3 is exploring a proof of concept on using AI agents to retrieve information from carrier documents, help draft employee benefit spreadsheet comparisons, and provide such information to the client manager.

“What co-pilots do is take away a lot of the time-consuming research,” Nsure’s Haisch says. “And it goes one step beyond that by disqualifying poor policy options [such as policies that cost more but do not deliver more value]. I think the trick is that it will need to explain why it is not doing this option versus this other option, right? But I think a lot of these decisions will eventually go to co-pilots and to AI agents.”

Implementation Challenges

IBM’s Duigenan says a key challenge to expanding AI agent usage is transparency regarding what exactly the system is doing and how, and verifiability that its deductions are correct.

Customers are wary of operations that involve little to no human intervention, IBM research shows. An October 2024 IBM report on AI found that only 29% of surveyed customers were comfortable with generative AI virtual agents providing customer service, and only 23% with them providing insurance advice. Just 26% trusted the reliability and accuracy of advice from generative AI. “In this environment, generative AI personalization may not improve customer trust; it can erode it. Enterprises are repeating past mistakes by not taking the time to understand their customers’ needs and concerns.” Despite those concerns, the report also noted that 77% of surveyed industry executives said they “need to adopt gen AI quickly to keep up with rivals.”

Duigenan says the trust gap could at least be partially addressed by progressing from AI assistants to more intelligent agents that can better reason through difficult judgment calls by better and more comprehensive learning from human professional experience data on which they are trained and by communicating with humans more effectively and more naturally.

Others say that retaining trust is in part a function of establishing a red line of automation that they may never cross. “At Zurich, we don’t ever commit ourselves to just automating responses with AI,” says Barry Perkins, chief operations officer for Zurich North America. Even highly sophisticated AI technologies can’t ensure the output won’t be hallucinated—providing a false or misleading response to a human prompt, he says, adding that allowing AI agents to operate freely without human checks could also put companies in regulators’ crosshairs.

The challenge of addressing those trust issues increases as the complexity of the insurance use cases and insurance lines compounds, Perkins notes.

“I think the underwriters, the carriers, all want a human being reviewing it at some point,” The Institutes’ Miller says.

Some deployed AI models that fall short of AI agents already have embraced a great deal of autonomy, introducing large amounts of risk for their human users in some cases, says Michael Berger, head of the Insure AI team for Munich Re and HSB. As an example, Berger says that a credit card fraud detection model, which Munich Re insured beginning in 2018, automatically decided, by itself, to accept or decline certain transactions.

“That placed the human user at the mercy of the AI model to really perform well and, if it didn’t, then a lot of fraud loss would have occurred for the merchant,” he says.

Berger says that error rates can vary significantly between different artificial intelligence use cases, including for AI agents, particularly for model systems that address many different kinds of use cases. Some of those have just a 1% error rate while others might have a 95% error rate. This makes it important for insurers to properly test those systems for their intended use on the right data to determine actual error rates and not just rely on the model’s design, Berger says.

Small, focused, flexible, and highly customized agentic AI models may address some of these challenges, Duigenan says.

“What I find with the trust factor is that it’s a very individualistic position,” M3’s Conroy says. “Some people are like, ‘I love this. I wanted it 10 years ago. This is very exciting.’ And there are others that because they’re so familiar with their current workflow, they trust that more because this is very new, uncharted territory.”

M3 takes the trust factor seriously, she says. While AI is a resource to support administrative operations, a human is still analyzing the output. Some AI systems that M3 is exploring will also measure the accuracy of their own output.

It will also take time to integrate AI agents into existing processes, particularly if they are central to work processes, rather than ancillary enhancements, Perkins says.

“You only get value from these tools if they’re integrated into your workflow,” according to the Zurich executive. “Building them on the side, and then trying to bolt them on is a much more difficult proposition than if I’ve got a piece of software in which I do the rating and I produce the documents and if I’ve got AI built into that—then you see the expansion of usage because companies that maybe couldn’t afford to have a team of difficult-to-find data engineers and LLM specialists for products built off to the side now don’t need them because they are already built in without further investment.”

Increasing Accuracy for More Complex Uses

Even for some functions that appear similar to fairly typical robotic process automation, Rotem Iram notes that agentic AI may be increasingly called on to determine how to manage more difficult-to-process data.

“When you’re looking at a capability to capture data, the first 50% is pretty easy, but then it gets incrementally harder,” he says.

At-Bay’s technology and engineering teams want to know what percentage of cases the agent can manage autonomously and its accuracy level, according to Rotem Iram. In the preceding 12 months, the percentage rose from 20% to 90% of cases, thanks to “the exponential improvement in the quality of the AI LLM models.” Meanwhile, accuracy of AI decision-making has risen from 70% to the upper 90% range— better than humans, who generally hit just above 70%.

“The challenge is to go from an AI agent that works well on some cases to an agent that works well on virtually all cases—but also knows when it doesn’t know,” Rotem Iram says.

At Applied Systems, the internal standard is to allow AI to conduct a task only when it performs better than 99% of humans asked to do the task, Tsur says.

Over the last year, AI applications’ capacity to learn and draw distinctions has greatly increased in a variety of areas, Rotem Iram says. A key step was AI agents reaching human-level comprehension of images—both identifying items and understanding their context.

“This has translated into an incredibly useful tool for automating insurance back office processes,” Rotem Iram says.

Implications for the Industry

There are several lessons for brokerages and human agents based on the emergence of AI agents, those interviewed said.

First, as AI assistants progress to AI agents and worthy co-pilots, brokers and agents who do not know how to exploit their capabilities will be at a decided disadvantage. They will be making recommendations and taking actions based on a smaller knowledge base and set of possibilities, will not be checked for things they may have forgotten to do (much as the AI in Spellcheck currently finds spelling flaws in communications), and some functions without AI assistance will take longer.

Brokers and agents would do well to begin experimenting with AI, which is being embedded in Internet browsers, word processors, and other applications they’re already using, The Institutes’ Miller notes. By experimenting with IT uses with a data set with which a human agent is thoroughly familiar and the AI agent has been trained on, the technology’s capabilities and limitations can more readily be gauged. Dial up the complexity or obscurity of the question posed and see what happens.

Organizations should also bring their stakeholders together to understand which processes may be ripe for agentic AI and test some use cases, says Joe Schueller, data analytics director at Waukee, Iowa-based brokerage Holmes Murphy. The company’s first hackathon, in November 2024, was a “smashing success” that demonstrated solutions that AI and agentic AI could provide to sticky business problems, he says.

Schueller says that while initial AI use cases are likely to involve existing process automation, they could eventually evolve into more sophisticated agent responses, such as whether a certain loss event is covered by insurance: “For use cases where accuracy is paramount, you have to be really buttoned down and I am not sure we are there yet with this technology. But I believe the technology will develop to that point in due time, perhaps sooner than any of us think.”

Agencies and brokerages should refine, digitize, and train secure LLMs on their proprietary data as soon as possible, so that AI agents can increase efficiencies in their various work processes and become reliable agentic assistants. Training on such a database could allow brokers and agents to learn from, and use, the best practices and experiences of the entire firm if the information from those experiences is accurately and consistently digitized.

Brokerages that manage this process successfully will generate smarter and more accurate AI agents keyed to the most relevant data, that of agents’ and brokers’ own firms, according to Miller. But this may require insurance carriers and brokers to improve their IT game. “The important thing that generative AI has done is it’s really exposed data governance problems in organizations,” he says. “Because you get much better results, many fewer hallucinations, if your data is well-organized and accurate when you go in to train the system. There’s a lot of bad data governance.”

AI agent costs must also be tracked. While LLM costs have declined, they remain expensive to rely on at scale given the vast number of AI functions and decisions they will need to perform iteratively to approach human capabilities.

“Even a single AI automated workflow is going to hit the LLM API 20 to 30 times,” Cowbell’s Gupta says. “And if you multiply that by the number of use cases you’re looking at, the cost can quickly increase exponentially.”

Using less-expensive small language models (SLMs) is sufficient for some steps in the workflow, Gupta notes.

Sources interviewed for this article offered varying opinions regarding how AI agents might impact the careers of human agents and brokers. At least initially, Gallagher Bassett’s Powell says, agentic AI might serve as a form of quality control for agents.

“To be honest, AI shouldn’t impact what the right answer is,” Powell says. “Ideally, it’ll improve your ability to communicate a little bit better and drive a more consistent outcome. Maybe [use by agents and brokers of AI co-pilots] will drag your bottom performers up to your better performers.”

To what degree AI agents displace human agents may depend on a given insurance organization’s operating philosophy and desire to balance cost efficiencies with a human touch, Powell notes.

“There’s a good number of companies who are going to say, ‘Let’s get as much cost reduction as possible out of this,’” he says. “And, ‘Let’s reduce headcount by taking our tasks off people’s desk and then taking the people out of the equation.’”

For high complexity lines, the amount of risk and the number of decisions necessary mean humans should remain involved for the foreseeable future, according to Powell. But that aspect could start to disappear from simpler processes that are effectively “just a set of steps” to follow without needing human judgment, he adds.

At Nsure, Friendly John has already impacted agent employment. Several years ago, while the Friendly John AI application was emerging, the company employed 100 agents and brokers. Now, there are only 30, says co-founder and Chief Operating Officer Wojtek Gudaszewski. Remaining team members manage more complex user requests, serve clients who refuse to use Friendly John, and handle clients with confusing or complicated data that cannot be interpreted by the AI.

“There is a big question of what happens to junior talent in insurance as AI expands,” Rotem Iram says. “Junior people have two things working against them right now. One is that they are the first ones that the AI can probably match up with. The second one is that they work from home and don’t go to the office as often as they used to, and so nobody knows their name and face and they don’t build relationships. It’s a precarious position to be in.”

But over the longer term, more senior brokers and agents could feel the heat of agentic AI, some say. Rotem Iram says the rise in AI intelligence and capabilities and the related cost efficiencies it presents through agentic AI will eventually disrupt the world as we know it, including the insurance world.