The insurance industry has been rapidly adopting generative artificial intelligence, but many organizations struggle to move from pilot to full implementation.
James Thom, Chief Product Officer at Vertafore, said at an InsurTech Summit panel that a focus on outcomes can help many carriers move on from the pilot stage. He was joined by Craig Weber, Head of Insurance Strategy at Cognizant, and William Steenbergen, Co-Founder and Chief Technology Officer at Federato. Here are where many projects fail and how carriers can move AI projects onward to implementation and business transformation.
Where AI pilot projects typically stall
AI projects often stall when carriers are focused on using AI to solve interesting problems rather than important ones. Carriers start investigating AI by asking what certain tools can do. Over time, one test leads to another, and teams can end up with a growing list of experiments that never become part of the business. The result is pilot fatigue.
That exploration can pull focus away from the more important question: What business problem should it solve?
Pilots also have a real cost attached, both in employee time and direct spend on licenses, vendors, or AI tokens. Most AI costs are calculated based on usage, not number of users. Every prompt, document, and follow-up has a cost attached, and vibe-coded projects and experimental pilots can add costs up quickly as they grow in complexity.
Insurers that move beyond the test-and-learn mode find focused use cases that change their production workflows. They spend less time and money on experiments and see real returns as they bring AI into their work.
How to move AI projects past the pilot stage
To break out of the cycle of experimentation, carriers should ask where an AI tool can create measurable ROI and then move forward aggressively.
Real ROI requires choosing a focused business problem. The use case should have clear value attached to it. Focus on actual bottlenecks that are making staff less effective than they could be: improving communication with insurance agents, automating credentialing, or helping underwriters turn incoming submission material into structured work.
Thom’s advice is to take a leap of faith with a use case that makes sense. “Start with one, and you’ll find at least a two, and at least a three, and at least a four very quickly,” he said.
That first use case gives the organization something concrete to implement, measure, and learn from. Once carriers prove value in one workflow, the next opportunity becomes easier to see and easier to justify.
Keys to success
Integrate AI into processes
One key to scaling artificial intelligence across an insurance business is making sure it improves the work people are already doing. A tool might be impressive in a demo, but it will struggle in production if it adds steps, creates another login, or forces employees to work outside their normal process.
AI needs to be integrated into the core workflow, not bolted on beside it.
According to an often-cited MIT report published last July, 95% of enterprise AI projects fail to produce ROI. But the report also mentions that the reason so many projects stall is “due to integration complexity and lack of fit with existing workflows.”
“If you’re doing a lot of swivel-chairing, if you’re having to force new tool adoption outside the existing workflow, you’re going to struggle to achieve scale adoption,” Thom said. “People resist that type of change.”
The test is whether the AI makes the workflow easier, faster, or better. If employees have to leave their system of record, copy information between tools, or change behavior without seeing clear value, adoption will be limited. But when AI is embedded where decisions are already made, it has a much better chance of becoming part of the business.
Identify why to build vs. buy
As with all technology solutions, insurers also need to make a build-versus-buy decision for their AI applications. Building internally may make sense when control is essential, especially if the use case depends on proprietary processes, data, or competitive differentiation. But if control is not the primary need, building AI tools from scratch can create more burden than value.
The challenge is the pace of change. AI models, infrastructure, governance needs, and integration patterns are still evolving quickly. An internal build is not a one-time project. It requires ongoing investment to maintain, update, and adapt. The question is not just whether a carrier can build the AI capability but whether they can keep it current, govern it effectively, and make it useful inside the business.
“We’ve set up AI infrastructures to handle things like LLM changes and complicated changes to keep up with the moving target that is AI,” Thom said. “You have to think carefully about whether you want to take that on yourself.”
Buying is not automatically easier. A vendor solution still needs to fit the workflow, connect to core systems, and avoid becoming another silo. By working with a trusted partner with AI solutions built specifically for insurance, insurers can trust that an AI solution will be secure and compliant and that it will work with the specific data they feed it.
How to move insurance AI projects forward
To move an AI project forward, Thom suggests carriers shift away from the mindset that projects are multi-year experiments to see ROI. Instead, AI innovation is happening quickly, and AI adoption should, too.
According to a study by Anthropic, the company’s Claude large language model reduces task completion time by 80%. “Seeing an 80 to 90% reduction in time savings is very real,” he said. “People are getting ROI out of implementation right away if you drive it hard.”
That is the difference between testing AI and changing a workflow. A limited pilot may prove the technology works, but it may not go far enough to create a real impact. Carriers can spend heavily and still see little return if they only “tiptoe around the edges.”
Thom’s guidance was to push harder: “You have to drive a truck straight into the middle of it.”
The practical move is to pick the use case you want to attack and go after it. There is value across many insurance processes, but carriers do not need to pursue all of them at once. They need conviction around one workflow, enough commitment to implement it, and a clear way to measure the result.