Making Dollars and Sense of AI

Early insurance brokerage exuberance regarding the potential of AI is giving way to more careful, nuanced assessments based on actual implementation of the technology. In many cases firms must prioritize among different use cases competing for dollars and employee time, say brokerage and IT solutions executives.

AI’s return on investment can be greatest for back- and middle-office solutions that sift through data to support humans and the business more comprehensively, quickly, and thoughtfully. In a 2025 study of business use of artificial intelligence, MIT researchers found that “while 50% of organization generative AI budgets go to sales and marketing, back-office automation often yields better ROI [return on investment].”

“I think that MIT study is right,” says James Thom, chief product officer at insurance technology solutions company Vertafore. “People don’t appreciate how much work there is in the back office.”

In the property and casualty sector, middle- and back-office work is fragmented, unstructured, and document-heavy, says Chris Watkins, chief information officer for insurance services provider ReSource Pro. Brokerages and brokerage operations-focused vendors are eyeing AI to help streamline many of these processes, conducting case studies of tools for processing and issuing certificates of insurance, policy checking, submissions, and billing reconciliation.

Companies are experimenting with the flood of AI-based systems, says Anupam Gupta, chief product officer for insurance technology company Applied Systems. But “nowhere close” to half of Applied customers are using artificial intelligence for their key workflows, he adds.

Broker Mindset

For AI investments to succeed, brokers say they must be integrated into an overarching strategy.

“Early signs are pointing to organizations struggling to attribute material value from AI because they’ve failed to embed it into core processes alongside rigorous frameworks that demand impact,” says Benjamin Funk, chief technology and chief AI officer at fintech and brokerage Acrisure. “We’re witnessing history repeating itself as the latest AI adoption efforts mirror the digital transformation missteps of the past, where initiatives launched without clear accountability or measurability, inevitably producing a lackluster outcome.”

Funk adds that successful AI deployments are carefully tailored to a company’s needs and fully integrated into core business operations. “Winners will be decided by those that fully activate a forward-deployed engineering model: [inserting] technical talent into business teams, understanding fundamental day-to-day challenges, and taking accountability for designing and delivering AI solutions in-the-field,” says Funk. “This assures that AI becomes part of the operational fabric of the institution tied to core business outcomes versus just an underutilized technical tinker-kit.”

For its part, Acrisure is incorporating AI education and training into all aspects of its organization, with hands-on enablement, role-specific tools, and a culture of experimentation that empowers every colleague to become an AI-native contributor.

The company’s priorities for the technology include increasing daily efficiency via uses such as communication acceleration, calendar scheduling, and code development, as well as helping to meet critical growth and profitability targets by identifying client needs, optimizing placement strategies, and accelerating go-to-market execution through intelligent automation.

Global broker Marsh takes a similar approach, organizing AI deployments around five themes, says Puneet Satyawadi, global chief operating officer for Marsh Risk. Among these: increasing everyday productivity; making data access more intuitive and robust across the organization; and having AI tools working quietly in the background, providing employees with the information needed to be effective in their roles.

West Des Moines, Iowa-based brokerage Holmes Murphy has been exploring AI tools for several years, says Dave Ashton, the company’s chief information officer. While an initial thrust was deploying general AI tools across the agency, in 2025, Holmes Murphy completed a comprehensive exercise to identify opportunities for applying artificial intelligence across the insurance product life cycle. The brokerage found that many of the most promising opportunities align with middle- and back-office functions, where processes tend to be data-intensive, according to Ashton.

“The next step will be to put more formality around the business case, the business benefits, the savings, and then that will help us justify the value of that AI relative to the cost of the AI,” he says, adding the firm hopes to complete that analysis in first-quarter 2026. Preliminary analysis suggests that AI solutions relating to functions such as policy checking and data submission are likely to be deployed later this year.

While Holmes Murphy is evaluating AI solutions from many insurtechs, advances in coding tools used by the company’s AI software engineers make the possibility of developing back-office solutions internally more attractive.

“We are finding that we will have more of an ability to do this than what we had believed when generative AI first became really well known out there,” Ashton says. “And that would allow us to ensure that our solution fits into our overall technology strategy and the experience we want for our employees; whereas when you use a [third-party] product, especially from a startup, they don’t necessarily have the same integration capabilities and don’t have the same experience as what we’ve been providing to our employees.”

Deploy Widely

Holmes Murphy’s initial approach of deploying AI generally across the organization has been replicated by other firms. To increase the overall AI savvy of employees, brokerages are working to make it easy for staff to access and use tools that will increase efficiency, regardless of their role.

Brokerage IMA Financial Group has implemented AI solutions across several back- and middle-office workflows, particularly in document comparison, policy review, contract analysis, and certificates of insurance, says Megan Cullen-Meyer, IMA’s vice president of data and AI.

Policy and quote comparison is one of the company’s most-used AI-enabled workflows, Cullen-Meyer says. In such comparisons, AI using PowerBroker software along with other tools and AI agents reviews policies, proposals, and quotes to identify differences in exclusions, limits, endorsements, and coverages. AI identifies discrepancies and coverage gaps, significantly reducing manual side-by-side comparison time, Cullen-Meyer says, adding that use of AI in such reviews has started reducing their duration by one third from the time that they previously took.

“Our goal was to manage document-heavy manual workflows and improve associate and client experience,” Cullen-Meyer says.

She notes how AI is used and scaled throughout the brokerage is as important as the benefits of a particular use case. “We operate on a unified technology stack and common core systems, which allows AI to operate on centralized data rather than fragmented broker systems. We empower ‘superusers’ across business lines to build and deploy AI agents, ensuring strong organic adoption.”

At specialty insurance distributor Amwins, Chief Technology Officer Michael DeGusta says general, enterprise-wide key buckets of AI applications include two areas: embedding AI throughout brokerage IT tools to increase performance by adding landing spots for employees to access AI software tools; and applying AI to improve document flow and generate new documents. “For example,” DeGusta says, “every time we receive this type of email, we need to do x.” This could include simple transactions such as a clearance request or a change to the broker of record. “Whatever the request might be,” DeGusta continues, “we will be able to not only determine next steps, but perform them efficiently.”

DeGusta says a variety of generative AI tools are available to Amwins employees. “All Amwins colleagues have Microsoft Copilot licenses, and our application development teams are utilizing Anthropic’s Claude to aid in rapid prototyping, coding, etc.,” DeGusta says. “In addition to ChatGPT, which is a great general-purpose AI model, we have our own AI model called AmChat, which allows us to embed custom prompts for various tasks unique to our organization.”

Amwins isn’t alone in building its own generative AI model. Many of the firms interviewed have launched their own generative AI platform, typically with the goal of creating something unique to their firm’s operational needs.

Marsh created LenAI for all employees more than two years ago. “LenAI is a homegrown platform with comparable capabilities to external tools, developed at roughly one-tenth the cost of commercial alternatives,” Satyawadi says. Generative AI has improved productivity, operational efficiency, and client service throughout Marsh’s businesses, with 60 AI-powered production systems already running across internal and client-facing applications.

Satyawadi says more than 80% of his colleagues are using LenAI regularly and the system is processing more than a million requests each week. “This means that for the majority of our people, work has already changed,” Satyawadi says. “As the volume of work increases, AI is helping us keep up with demand without the need to dramatically increase headcount.”

Holmes Murphy’s Hey Holmes was deployed in mid-2024 to assist in crafting emails, agendas, job descriptions, and other documents, all with the requirement that employees review those documents for accuracy, Ashton says. These tools reduce the time employees spend preparing documents by reviewing provided content for clarity, tone, structure, and accuracy, then refining them to be concise, professional, and impactful while preserving the sender’s intent. By providing context through prompts and responses, users collaborate with generative AI to articulate their messages, Ashton says.

ReSource Pro also offers an AI copilot internally to its thousands of employees.

“Our tool, AIChat, is similar to ChatGPT, but it’s got all the compliance security network we would we have, and it’s trained with some of ReSource Pro’s knowledge, so an employee can ask questions, for example, about insurance acronyms,” Watkins says. “It makes it easier for employees to onboard and have a central place to be able to ask questions.” AIChat was initially offered in January 2024 and about 50% of the company uses it at least once per month, Watkins adds.

“Some of our clients that are doing interesting things with AI are building their own GPTs that operate as their internal Help Desk,” Vertafore’s Thom says. “For many clients today, especially larger agencies and brokerages, their information used to be scattered across multiple systems and teams, forcing employees to hunt around for the data they needed. Now, they’re creating internal AI tools that consolidate their standard operating procedures, training materials, and workflow documentation into one chat-based interface. We’ve seen some agencies reach the point where 40%–60% of their staff use these tools daily, which I’d consider a clear sign of success.”

Train to Adopt

While many interviewed for this article tout broad adoption of AI tools within their organization, that kind of uptake doesn’t come by accident. In fact, a key challenge, especially for larger brokerages, is driving AI adoption throughout their organization.

Satyawadi says to drive AI adoption at Marsh, it brought together its technology, data, and operations teams under one roof, forming Business and Client Services.

The organization has also invested heavily in upskilling its workforce through initiatives like the Marsh University and its AI Academy, ensuring colleagues understand what generative AI is and how to use the various tools within the LenAI suite.

At Acrisure, Funk says that employee adoption is accomplished in part through forward-deployed engineering, which means sending technical experts into business teams to fundamentally understand their processes, identify high-value pain points, design AI solutions that integrate deeply into those workflows, and ensure the delivery of results.

“Our Auris [AI] platform has served as a bridge between our technical teams and our client advisors, assisting broad AI adoption across our more than 19,000 employees,” Funk says.

After initially relying on culture to drive adoption, ReSource Pro is considering a more formal approach. “We’re a fairly innovative group of people,” Watkins says. “We have a culture of innovation and trying to always do better, and so the 50% adoption of a tool like AIChat by our employees has really been through email campaigns and some webinars, but there hasn’t been a full training program. But with AI tools getting even more complex and even more beneficial, we are starting discussions around establishing a formal training program.”

Measuring Gains

Many analyses have highlighted that hype over corporate uses of AI has outpaced ROI in the insurance industry and beyond. As McKinsey stated in a July 2025 report, “Only a few insurers have extracted outsize value from AI to gain a competitive edge.”

“There’s so much hype here that oftentimes people are more focused on the activity of doing an AI project than actually getting something out of it,” says Watkins. “I think most of these AI efforts have negative to questionable ROI.”

Many brokerage and IT vendor executives interviewed for this article said that precise measurements of overall AI benefits at brokerages, including for middle- and back-office use cases, are often hard to come by. In addition, a number of brokerages declined to discuss their approach to the technology.

“There is real value in the technology, but the hype is saying it’s greater than the reality,” Watkins says. “I can’t give specific dollars, but we’re looking at tasks where the average handling time can be reduced 10% to 20% whereas the hype is saying that costs can go to zero. But that’s not what we’re seeing or really believe will happen.”

The majority of AI uses are often not sufficiently refined to develop hard dollar revenue savings estimates, but process and human hours are good proxy metrics, Watkins emphasizes.

“People should be looking at measuring the short-term ROI by looking at average handling, time reduction, and the speed [of] processes before and after AI,” he says. “Or, can a company grow without increasing headcount long term? I think eventually, as AI practices mature, there needs to be more of a focus on measuring ROI based on the business outcome, whether that’s time to bind reductions or policy accuracy rates, things [that] are directly tied to the business outcome, but we are a long way away from that right now.”

ReSource Pro measures value through process efficiency and human hours saved, Watkins says. In policy checking and certificate of insurance management, plus AI-added operations such as loss runs, the company has observed approximately 20% to 30% reductions in human hours, while maintaining accuracy through structured workflows and human oversight, Watkins says.

Watkins also believes the value of ReSource Pro’s proprietary AIChat tool is clear. “The cost of deploying based on open-source technology and manag[ing] it is really [in] the thousands of dollars a month,” Watkins says. “And then we see the amount of emails that people are using to write and translation and the usage across that many people. While we haven’t done a specific ROI for that case, it’s clearly very positive. I think we’re paying about $10,000 a year to maintain this tool. Everyone in the company is using it.”

Cullen-Meyer says IMA has a process to scale use of new AI technologies that helps properly vet them before their financial and time costs increase.

“We take a disciplined proof-of-concept approach, where we test pilots and measure ROI within specific teams, and those teams validate whether efficiency gains or quality improvements justify broader deployment,” she says. “Business leaders commit to offsetting investments before we scale companywide.”

AI investments at IMA must make economic sense relative to our labor costs, turnaround speed, and output quality.

“In some use cases, particularly for policy and document comparison and contract reviews, the benefits clearly outweigh the costs, driven by measurable time savings and productivity gains,” Cullen-Meyer says. “Qualitative impact is equally important. We measure our associates’ satisfaction and the increased time they spend on client relationships rather than repetitive administrative work as part of the ROI.”

Some brokerages and services vendors are trying to produce hard benefits calculations, particularly those like Applied Systems that sell AI-enhanced solutions as discrete services. Applied Systems estimates that AI could save brokerages 40% to 50% of their time by automating most manual tasks, such as policy comparisons, and that automated task validation could reduce operational errors by up to 90% in account reconciliation, data extraction and entry, and policy and submissions management. Gupta says such statistics are based upon quantitative and qualitative measurements.

“We are doing user experience research all the time, which means we are walking the halls of the agencies and observing how they work and asking anecdotally about, ‘Hey, how much time would this save you,’” Gupta says. “From our observations, we know how much time they’re spending in different workflows. So that one’s qualitative. Then we are also doing quantitative input [regarding savings] from hundreds and hundreds of agencies on an ongoing basis around many specific questions.”

Applied says it anticipates agencies boosting revenue by 30% by doing effective cross-selling through AI, a number that is based upon its deployment of AI solutions, Gupta says. Applied’s products already feature time-use tools that allow clients to measure efficiency gains. “They’re all instrumented so that you see how much time is being spent in different workflows and how much more efficiently you are doing things,” Gupta says.

Business process outsourcing savings are another hard number many point to in support of artificial intelligence. “A lot of those situations where we’ve used AI in back-office situations is where we’ve done the outsourcing and we’ve been able to pull more and more of those use cases in internally with AI,” DeGusta says. “We have a whole process where we’re just sort of looking at all those workflows and saying, ‘Hey, which needs can we now use AI for?’ And we’ve already gotten large savings by switching a lot of those manual, outsource[d] processes to internal AI.”

Thom takes a similar view, saying that “things like commission statement ingestion, policy data entry, policy checking, solutions that are highly manual today,” are ripe for AI, though he notes that savings must be compared to the costs of developing and managing the internal AI tool. “AI is making a significant impact in how our clients manage data ingestion. We’ve built this capability directly into our core solutions. It enables users to automatically ingest documents like emails, statements, and policies directly into the system, improving both the quality and completeness of their data.”

“I think there’s two types of customers out there. First, there are those who are big on experimentation— they’re investing heavily, exploring many different use cases, and searching for that ‘needle in the haystack’ that will prove a strong ROI,” Thom says. “Then there are others who take a more disciplined approach, focusing their AI investments only on large, meaningful problems where they’ve identified a clear ROI from the start. The companies that have seen real success are those that deeply integrate with an agency’s existing systems, whether with Vertafore solutions or competitors, through APIs and other integrations. That kind of seamless connection allows AI to fit naturally into existing processes and deliver immediate value, especially in back-office operations, where ROI can be proven quickly.’”

Quality of AI output is another way to measure benefits, some note.

“AI-generated code, for example, is not a new use case; people have been talking about it the last 12 months,” Watkins says. “But the capabilities of the model, particularly with ChatGPT 5 and Anthropic’s Claude, function much better than they did even a few months ago. Three months ago, the output of the code would be very buggy and would be sort of missing the intent of what you were trying to do, and it would probably take more effort to generate something that was actually functional, what you meant, then just writing it from scratch, if you had the capability to do that because you were a sophisticated engineer. But now it’s getting to a place where someone who has a reasonable knowledge of how to architect software, but who isn’t necessarily a software engineer, can write pretty sophisticated applications that are functional and that work, even if they are not well-written code.”

The importance of accuracy depends on function, according to Gupta. The bar can be higher for data-intensive back-office functions.

“There’s not one rule of thumb across all use cases,” he says. “For example, if you’re doing financial reconciliation, the bar is higher—you need 95% plus accuracy, even higher in some cases, for this to be used on an ongoing basis. In contrast, if you’re summarizing an email, how do you measure that kind of accuracy? The way we measure that is: how many times did someone actually make an edit? How many times did they give a thumbs-up or a thumbs-down on the suggestions that the AI put in front of them?”

The Costs of Efficiency

The benefits provided by AI don’t come for free. A particularly sensitive aspect is the potential for some human jobs to become obsolete, and organizations will have to deal with the administrative and reputational costs of mass layoffs if those occur.

Beyond the human capital cost, the cost of ensuring the accuracy, refinement, extensiveness, and consistency of an organization’s data—which is necessary for AI applications—is enormous, particularly for brokerages that have grown through acquisitions and integration.

Many brokerages have not refined that data, particularly those that have grown through acquisition of smaller companies with different IT systems and data storage practices. “When you talk to these consulting firms, you learn that a lot of the work they’re doing on AI projects is really just trying to gather internal data in the right form,” says DeGusta.

“We’ve made long-term technology investments and created a unified platform that houses documents and data that can’t be accessed in the public domain. Our initial investments in data weren’t made with AI in mind, but they have served as an accelerant. We can use our in-house data to train AI models and test the accuracy of AI tools,” DeGusta says.

As agencies expand the use of AI across back- and middle-office operations, this foundational work becomes a strategic investment rather than a hurdle, according to Gupta. Clean, well-governed data enables AI to automate reconciliation, standardize workflows, and reduce manual handoffs at scale, creating long-term operational gains. In practice, agencies that prioritize data normalization and integration early are better positioned to realize sustained efficiency gains and faster returns from AI-driven automation, Gupta says.

Likewise, the cost for ReSource Pro to deploy AI across various use cases was limited because the technology merely augmented earlier IT investments, Watkins says.

Another notable expense is developing and running large language models—like the ones being deployed in numerous brokerage firms. Generally, training the model by teaching it what it needs to know to address user questions costs far more than running the model to perform tasks such as retrieving information or writing documents or code. That is because training the model involves ingesting a far greater volume of documents and resolving discrepancies through adjustments and reruns.

Specific expenses for these operations will depend on a brokerage’s digital maturity. There is no rule-of-thumb cost across the industry.

Building an in-house AI solution may not be the most efficient approach for many brokerages, particularly smaller firms, Watkins says. Rather, they may prefer to rely upon experienced third-party providers.

Cost containment is key even for AI-forward businesses, DeGusta says. For example, applying AI to ingest and analyze longer documents generally costs more, which presents opportunities for savings, he says. “If the first 180 pages of a document is boilerplate, only send the last 20 pages with the important data that varies and that people care about through the model,” he says. “Approaching it that way can help reduce the cost [of running the model].”

Likewise, applying AI to analyze preexisting, organized data, such as that in a spreadsheet, is much less costly than extracting and processing data that must be mined from large, less data-analysis-friendly documents, such as a report or insurance policy documents, DeGusta says.

For brokerages of any significant size, it takes specialized IT talent, including data scientists, to deploy AI effectively, DeGusta says. Amwins employs about 20 IT employees to work full time on AI-related projects, with an additional 60 working on those projects in some capacity.

“Part of the reason we made the investment in a true data science team is to enable us to build our own insurance use case models,” DeGusta says. “We train our models from scratch, and while this is a different cost model with more upfront costs, the ongoing costs are lower because we’re not paying OpenAI or Anthropic to run the model.”

Smaller brokerages should have at least one AI expert on staff, DeGusta says: “It is critical for at least one person to be fully versed in and have a complete understanding of AI. That person needs to be in every meeting, particularly in discussions of traditional software, to provide the necessary feedback on how to implement AI.”

Training employees on AI use can also be a large cost. “As a rule, for every dollar spent on developing digital and AI solutions, plan to spend at least another dollar to ensure full user adoption and scaling across the enterprise,” the McKinsey study advised.

Costs aside, AI adoption doesn’t seem to be slowing. And many interviewed say it will become an even bigger part of their work moving forward.

Some say AI will eventually hold a key role in strategic decision-making. “Looking forward, we’re hyper-focused on embedding AI in core decision-making workflows—not just using it for summarization or chat,” Funk says.

While the MIT report found only a certain amount of value for many AI deployments, the McKinsey study found that insurance carriers that are AI leaders are already outshining their peers, creating six times the total shareholder returns of laggards over the past five years (compared to two to three times in most other sectors). It also found that “domain-level rewiring with AI has had a measurable impact on key parts of insurance businesses, including a 10 to 20 percent improvement in new-agent success rates and sales conversion rates, a 10 to 15 percent increase in premium growth, a 20 to 40 percent reduction in costs to onboard new customers, and a 3 to 5 percent accuracy improvement in claims.”