Insurers are raising their AI budgets and getting choosier about which projects they take beyond a pilot.
Once a tool goes into daily use, the cost of protecting customer data and monitoring its output can change whether it’s worth rolling out across the business, according to Chris Raimondo (pictured), EY’s global and US insurance consulting leader.
KPMG’s US insurance practice reported this year that 90% of the insurance executives it surveyed said their companies had increased AI budgets from the prior year. In a separate survey of 44 North American property and casualty insurers, Celent found the average respondent expected overall IT funding to rise nearly 10% in 2026, with AI and the data needed to support it among the investment priorities.
Raimondo told Insurance Business that the conversations he’s having with carriers have changed as they move past early trials. The first rounds of testing helped insurers see where the technology could be useful. This year, he said, clients are asking how they can run it securely at scale and what they’ll get back for the money they put in.
“They definitely see the opportunity, but they realize that the path to getting there is continuing to be more complex and probably requires more investment,” Raimondo said. “So, they’re going to be much more selective moving forward in how they channel those investments to the highest-ROI opportunities.”
Read more: Why your agentic AI pilots fail to deliver returns
Deloitte’s 2027 insurance outlook, published November 12, offers a snapshot of the returns so far. Seven of 35 insurance technology leaders in its global survey rated the returns on their AI investments as high or very high. Deloitte said operating costs could become more variable as insurers use different models and as usage grows.
A pilot can show how much time a tool saves on a particular task, but when the same tool goes into regular use, the insurer has to decide who can access the data and how to reconstruct what happened if someone questions the output. Those costs come on top of running the tool itself, Raimondo noted. “To do it effectively, you have to put the right investment around governance, safety and security,” he said. “But at the same time, the industry is challenged for growth. The industry needs to operate at a more profitable baseline.”
Among EY’s clients, adoption has moved faster for employee productivity tools and AI features built into software staff already use, such as claims systems.
Custom agents built for specific tasks are still more often the subject of experiments. Once those agents start doing work on their own, insurers also have to decide where someone needs to check the result.
Raimondo sees more scope for agents to act independently in internal IT. In underwriting, pricing and claims servicing, where errors can affect what a customer pays or how a claim is handled, he expects people to keep working alongside them. The carrier also has to watch for model drift, when a system’s output changes over time.
A 2026 WTW survey of 59 P&C insurers in the US and Canada found that 16% were using AI to support human underwriters and 60% planned to prioritize that use by 2028. An underwriter using AI to assess a risk can check the result against the submission. Giving an agent more discretion means the carrier has to work out how it will catch errors before they affect a quote or a claim.
“The closer you get to the core insurance value chain, the higher the bar is going to be to automate that,” Raimondo said.
As carriers bring those tools into daily use, Raimondo said modern systems and usable data can give a smaller insurer a head start over a larger rival still working through older technology. He said senior leaders also need to agree on where AI fits and how staff will use it in their day-to-day work.
“You have to start embedding it into how you’re going to work in the future, how you’re going to serve customers in the future, how you’re going to operate the business in the future,” he said. “You can’t think of it as a bolt-on in your existing landscape.”
