Nearly half of insurers have moved generative AI out of the lab and into production, according to a Celent report on the state of generative AI in insurance. The firm surveyed insurers and found 48% are now live with the technology, with that figure expected to top 50% by the end of the year.
The shift marks a turning point. After years of pilots and proof-of-concept work, firms are now asking a harder question: where does the technology actually deliver measurable value?
Putting AI in front of experienced people
Matthew Twist, vice president of EMEA at Earnix, said the conversation has changed dramatically over the past year. "A year ago, everyone wanted to know what generative AI could do. Today, they are asking where it delivers measurable value," he told FinTech Global.
The answer, Twist said, is not in replacing underwriters or claims handlers, but in giving them tools that make better decisions faster. An underwriter doesn't need another dashboard. They need a tool that can read 200 pages of submission data and flag what matters. A claims handler doesn't need to be automated out of the process. They need relevant information collated so they can act on it.
"That's where I'm seeing genuine traction," he said.
Twist also cautioned that the hype around Generative AI and LLM tools can push firms toward the wrong solution. Some problems are better served by a rules engine or straightforward workflow automation. "I've seen organisations reach for generative AI simply because it's the technology everyone wants to talk about," he said. "Sometimes a rules engine or straightforward workflow automation is exactly the right answer."
The hidden risk on insurers' books
Melanie Hayes, co-founder of KYND, sees value in a different place: understanding how clients are already using AI, often without declaring it. AI is now embedded across businesses of every size, frequently through third-party tools rather than deliberate strategy, and much of it never appears on a proposal form.
"Carriers, brokers and MGAs writing AI cover tend to describe a similar picture: exposure building steadily, largely undeclared, and concentrated in a handful of widely used models," Hayes said. "It looks like an accumulation question, and cyber risk analytics that map a live technology footprint can help bring it into view at the point of underwriting."
Where generative AI fits best
Both Hayes and Twist agreed that the biggest value sits in processes built on established data. Broker submissions, engineering surveys, policy wordings, claims files, and customer correspondence are all areas where generative AI can make sense of unstructured information at scale.
But Hayes warned that the quality of the underlying data determines whether the output is useful. A model summarising a month-old scan "will produce a fluent description of a risk that may no longer exist, and it will do so convincingly," she said. The processes best suited to generative AI are those already built on live, complete risk data.
Twist added that AI should not make insurance decisions on its own. "Insurance decisions affect people's livelihoods and businesses," he said. "There still needs to be accountability."
Why this matters for insurance professionals
The takeaway is not to chase AI for its own sake. The firms that win will be those that pair specialist expertise with technology to get a clearer, more dynamic view of risk, pricing, and exposure. For underwriters, brokers, and claims teams, the practical path forward is to identify the language-heavy, high-volume tasks where generative AI can save time, while making sure the data feeding those tools is current and complete. Those working in AI for Insurance should focus on augmenting experienced judgment, not replacing it.
Your membership also unlocks:
