Some property/casualty insurers talk about experimenting with artificial intelligence to automate call center responses and back-office operations. But that’s not the road to turbocharging growth that AIG CEO Peter Zaffino envisions for his company.

AIG is all-in on AI that is directly focused on its core activities—underwriting and claims, Zaffino and Claude Wade, executive vice president and chief digital officer, told analysts during a four-hour Investor Day event late last month.

“We’re focusing on this as an end-to-end process—not on the fringes, [and] not just for expense savings,” said Zaffino, setting up a discussion of what the two AIG executives described as a strong “agentic AI ecosystem” where AI agent tasks involve data ingestion on one end of the underwriting process, all the way to figuring the propensity to bind on the other.

The ecosystem, which involves partnerships between AIG and well-known technology companies Anthropic and Palantir, will help AIG underwriters get through 500,000-plus E&S submissions to book at least $4 billion in new business premiums in the year 2030, Zaffino projects. Already, AI tools that are up and running in financial lines allow AIG to review 100 percent of every private and non-profit business submission that comes in, without adding underwriters, Wade said.

Anthropic CEO Dario Amodei and Palantir CEO Alex Karp praised Zaffino’s vision and leadership of AI initiatives for the insurer’s core activities of underwriting—already live for private and non-profit financial lines business, and in the works for claims processing—as they joined the AIG leader on stage for a half-hour panel discussion during the three-and-a-half-hour Investor Day event.

“Enterprises are having a lot of success deploying [AI] in the periphery,” said Amodei. “That’s common across all companies. All companies need customer service. They need internal productivity for their developers. That’s the same whether you’re in insurance or some other area. So, [there’s] a lot of success there, and a lot of aspiration to make progress in the core of what they do.”

Rather than experimenting with pilots, AIG has moved past aspirational goals, leaning into the core use cases, Amodei said. AIG “really had conviction on the claims cases and the underwriting cases, and has moved quickly to get a lot of this unstructured data [into the process] and really, really bet with conviction,” he said, noting that AIG started working with Anthropic when it’s AI Assistance Claude was Claude Version 2.1.

While AIG is now using Claude 3.5, it’s not a stronger AI model that is driving underwriting success for the carrier, Amodei said. “The technology can be great, but if there isn’t that conviction, if there isn’t that will to move quickly and that focus, it doesn’t happen,” he said.

“The real differentiator is not in the technology, as fast as that’s improving, but finding ways to deploy it in enterprises, finding ways to revolutionize an existing business,” Amodei said.

Palantir’s Karp also lauded Zaffino for starting to build the AI ecosystem ahead of the curve with a “singular and somewhat prescient focus” on that revolution of core business activities. Later, CNBC Analyst Sarah Eisen, who moderated the panel, asked Karp to talk specifically about what he had learned about AI applications for insurance.

Karp told her that insurance executives who ask, “Can AI outperform a human in what they do?” are asking the wrong question.

The right question is, “Can AI make one human [perform like] five humans?” he said, noting that in the case of underwriting, “the human is doing something very technical.”

“You have an expert who’s been trained in understanding not just data but proxy data … You’re looking at inferences for what it means based on years of experience,” he said. He drew an analogy to the analysts in attendance who can listen to executives of different companies saying the same thing to discern which will make money. Technical human underwriters have built similar abilities, he suggested.

“Can the large language model…make one human 10X more valuable, meaning they have 5X output in half the time?”

Getting to $4 Billion

Zaffino and Wade offered different numbers that would likely grab the attention of insurance industry peers, with Wade describing the measurable impact of AIG’s AI Underwriter Assistance on private and non-profit financial lines business, and Zaffino sharing his bigger vision of the growth of all E&S business for AIG’s Lexington.

According to Wade, with AI Underwriter Assistance already in use for North America private and non-profit business submissions:

• The underwriting timeline per submission has been slashed to less than one day, down from three or four weeks.
• The bind-to-submit ratio has increased from 15 percent to 20 percent.
• Data quality has increased more than 90 percent.

During his introductory remarks, Zaffino pulled back to create a bigger picture for E&S business growth. He reported that new business E&S submissions exploded to 300,000 in 2024 from 30,000 in 2018, with AIG’s Lexington binding just 2 percent (6,700) at an average new premium of $140,000—or roughly $1 billion total last year. Going forward, he expects a 10 percent compound annual growth rate in submissions, putting the total at about 500,000 new business submissions coming into Lexington in 2030.

“Can we keep doing 10 percent and stay with the industry on a CAGR basis? Absolutely we can. That would be very good by normal business standards. But I want to challenge us to think differently. I want to think about 20 percent-plus,” Zaffino said.

“Can we keep doing 10 percent and stay with the industry on a CAGR basis? Absolutely we can… But I want to challenge us to think differently. I want to think about 20 percent-plus.”

Peter Zaffino, AIG

Envisioning that AI Underwriter Assistance can push the bind rate up to 6 percent in 2030 from 2 percent in 2024, the volume increases to $4 billion (assuming the same average premium), he said, reporting a premium volume figure in line with the 20 percent-plus per year aspiration.

“You get four times the amount of new business… And I think when we do that, it’s not going to be just adding volume, it’s going to be better risks,” he said with an excited tone. “I don’t even know what happens with the market. I can give you a view [which suggests] the 500,000 [submissions] will be light. We’re going to be able to bind more.”

“This is the direction we’re going. I give Lexington as an example because it is the one that is right there in front of us, but there’s a lot of other businesses with an AIG [where] we’re going to be rolling this out and adopting it. It’s for real. And for companies like us, I don’t believe this is a choice. We’re doing this. We have to do this and drive change,” he asserted.

Getting Started and Building Layers

While the E&S example reveals a four-times growth scenario, like Palantir’s Karp, Zaffino more often referred to a “five times” ideal when describing the early steps in moving to the AI Underwriter Assistance solution.

Starting his talk with a slide headed “Accelerating the End-to-End Underwriting Process by 2X – 5X,” Zaffino emphasized two things. “This was end-to-end, and we wanted to focus on growth. That was the entire premise of our strategy. And the two-times or five-times reflects that there were two piles. If you have a strategy for Gen AI that can reduce cycle time or improve revenue by two-times, it goes in one bucket. If it’s five-times, it goes in another.”

“Everybody was involved,” he said, going on to focus on AIG underwriters. “Today, there is so much time still spent gathering data, gathering financial statements,” he said, noting that creating the path forward started by giving underwriters the opportunity to identify the 125 data elements they would like to have a perfect submission.

AIG also aimed for consistency here. “When you’re extracting that data, does it come from a bunch of different sources? If you ask underwriters, typically they will start to get the same [data elements] but they get it different ways. It makes the process very inefficient,” he said.

Zaffino also mentioned a needed cultural shift. “You can’t just overnight say, ‘OK, now you’re going to do five-times the amount of underwriting today. Here you go. [We] can’t wait for the outcome.'”

“You’ve got to start to shift,” he said, alluding to training, preparation and engagement with distribution without delving into the process of shifting the culture in detail at that point. Later, however, in responding to a question from Eisen during the panel discussion about how LLMs change the business of underwriting, Zaffino admitted, “There was not necessarily, early days, as much belief” in the power to change. A simple exercise helped shift the culture, he suggested.

“We said to the underwriters, I’ll give you an infinite amount of time if you’d like, and we did. And within two to three weeks they said, ‘OK, I have what I need.’ And they got probably 75 percent of the data that we had outlined. And then with what we were doing with Palantir and Anthropic, we got 92 percent of the data in three hours.”

“So, all of a sudden it was a shift in [the] dynamic that is cultural. We’re not going to waste time in underwriting, pulling data. We’ll have more data, more insight and better decision-making. It’s going to be profound for us over time,” he said.

Still, one thing that’s not changing is broker behavior, the former Marsh executive said, noting that brokers submit data in countless ways. “It’s structured. It’s unstructured. It’s PDF. It’s text,” he said, noting that the challenge of gathering sources containing the 125 needed pieces of information in real time to get risk insights to the underwriters was solved with Palantir Foundry. Zaffino further explained that the data isn’t just ingested from the broker submissions but is supplemented with internal AIG data (for renewal business or for accounts that used to be with AIG) and data from 30 approved third-party sources.

Next, large language models like Anthropic’s Claude 3.5 are trained to extract the needed pieces of data from the ingested sources and also trained to prioritize certain risk characteristics—an industry group, geography or size group that AIG might be targeting, Zaffino said. “So, our underwriters are actually reviewing where we believe [there are] the best risk-adjusted returns,” he said, noting that the bind rate should go up as a result.

Another wrinkle—underwriters can see the source of the extracted data, he explained, noting this helps to minimize errors caused by AI hallucinations. An underwriter seeing that a data source is a rating agency, for example, can go check it and feel confident in the reliability of that information, he suggested.

Wade provided a more technical discussion about how AIG’s Underwriter Assistance was designed to improve response accuracy, referring to a Retrieval Augmented Generation framework. Offering an example of the RAG framework in action, he described the retrieval of a client’s total 2024 revenue, with an initial step of the framework essentially pointing to a reliable data source—the 2024 income statement—and a next step sending the prompt and relevant context “Review and return Total Revenue from 2024 Income Statement” to an LLM. The LLM generates a response “including the source document chunk from which the 2024 revenue was found, which is then reviewed, confirmed or corrected by an AIG underwriter—”as we call them, the human in the loop,” Wade said.

“This approach substantially minimizes hallucinations by limiting the source context for the large language model, but it also drives better efficiency in terms of response time and token costs because we’re only sending the most relevant data and context to the large language model. And as a core design principle, the human in the loop has to validate the response and determine the next best underwriting action,” he said.

Providing some more technical details, Wade also spoke about the componentized or “modular” architecture of AIG’s Underwriter Assistance, which allows the assistant to evolve along with any new AI technologies that develop, and to rapidly scale to new lines of business, products and geographies.

Like Zaffino, Wade provided less technical background of the overriding company and industry context that led to the development of the RAG framework, often coming back to the human-in-the-loop approach.

“We are not a technology company. What we are is an underwriting and claims company,” Wade said, early in his presentation. “We set out to leverage technology to turbocharge our highly experienced knowledge workers, not to replace them with technology,” he stressed, outlining the challenge of having “an army” of underwriters pore through nonstandard documents to find the 100-plus data elements to assess submitted accounts against underwriting guidelines. Without AI, “we do not consistently get to every submission we receive. That leaves profitable business on the table,” he said.

“This isn’t just an IT or data science project,” he stressed. “We [convened] an underwriter council and then we co-located them with our digital teams so we could ensure that the solution we built aligned with real-world workflows,” he said. “The underwriter is at the core of our business and at the core of our gen AI strategy.”

“We set out to turbocharge the underwriter, our knowledge workers, not to replace them. Our goal is to enhance their human expertise.”

Claude Wade, AIG

Noting that AI Underwriter Assistance moved from a concept to full deployment in private and not-for-profit financial line within a 10-month timeframe, Wade said, underwriters in the division “now arrive at their desk to find each submission has been ingested, augmented and prioritized.”

“All the pertinent data—unstructured qualitative data from those submissions, [and] structured data from our relevant internal sources—are all immediately available in a curated summary ready for them to begin underwriting.

For ingestion, LLM-powered document classification identifies and categorize the submission documents, such as applications, financial statements and loss histories “consistently at an accuracy level of 97 percent,” Wade reported. “Then using data extraction tools, we capture the specific data points such as revenue, number of employees or physical location. This leverages very context-specific language models like Anthropics’ Claude 3.5 and Palantir AIP, he said.

Next, data is augmented with AIG’s internal data, such as our policy and claims history, and third-party data and external research from trusted sources like Dun and Bradstreet, he explained.

“Finally, we apply AI and machine learning to estimate the propensity to bind,” he said, referring to prioritization. “What that means is it’s ranking the submissions by their likelihood to convert from quote to policy,” he said.

Wade said that with AIG Underwriter Assistance live in production for North America private and nonprofit business, the insurer is now rolling it out to all North America business lines and international with a target date of December of 2026.

Later this year, AIG will also launch AIG Claims Assistance, leveraging the same capabilities of ingestion, augmentation and prioritization built in Underwriter Assistance to accelerate claims adjudication and payments, improve claim outcomes and client experience—”all designed with our philosophy of human in the loop with turbocharging our claims knowledge workers and not replacing them,” he reported, without noting any specific business segment targeted for the launch.

“This is happening for us—and it’s something that is real and something that’s going to be a big part of our company going forward,” Zaffino enthusiastically told listeners, concluding his remarks about AIG’s use of AI.

Featured image: AI-generated/ Adobe Firefly