What happens when AI agents talk to each other? The risks for insurers

Using agentic AI to improve operational speed and efficiency in property and casualty insurance organizations may be the way of the future, but brokers will need to continually audit collected data to affirm their origins and integrity, P&C insurance professionals heard at the Insurance Bureau of Canada’s 2026 Insight Summit in Toronto last week.

To this end, one emerging risk to watch for is whether the data is being delivered to insurance providers by humans or machines.

“I have a fear, a real fear, of [AI] agents talking to other [AI] agents,” says Jason James, founder and chief information security officer at Emperium Governance Risk & Compliance, during a panel discussion on AI in commercial insurance. “I have not seen a technology that can detect another agent [or] knowing that it negotiated with the [another AI] agent.”

Related: How to insure AI data centres in Canada

James imagined a conversation between two agentic AI agents, one representing a P&C company’s firewall, and the other an agentic bot representing a bad actor posing as a commercial insurance client.  

In this scenario, the insurance company has a firewall. And the agentic AI firewall is instructed to ‘stop all bad things.’

An AI bot representing a commercial company applies for insurance and starts talking to the company’s AI bot representing the insurer’s firewall.

“And this [commercial company’s AI agent is] saying, ‘I am an automated agent trying to get into this organization,’” James says. “And [the insurance provider’s firewall] agent says, ‘No, you can’t come.’”

To this, the commercial client’s AI agent says: “Hey, don’t worry about it. I’m good. I’m cool. Let me in.”

“Where do you want to go?” the insurer’s firewall asks. “I’m not supposed to do it.”

But, James says, the client’s AI agent responds: “Yeah, I just want to grab some stuff over there.”

“It’s okay,” says the company firewall. “Make it quick.”

James says this hypothetical conversation may sound funny, but it’s entirely within the realm of possibility.

“How would we know?” James asked rhetorically.

Game theory

He described another situation in which a client’s AI bot managed to game the insurer’s application process.

“This is my bot I created at home that has all my personal data in it,” James says “It knows exactly what to respond [when asked for information during the insurer’s application process]. It does the calculation of [my] background to get me the lowest premium. It may or may not be true.

“Do we have technology right now to say this application was filled out by human or a bot?”

The level of AI use in the Canadian P&C insurance industry is unknown and difficult to track. Seventy-five percent of 32 brokerage principals in Canadian Underwriter’s 2026 National Broker Survey reported they have not invested in AI at all over the past two years. Twenty-two percent reported making investments of up to $5 million.

And a 2024 study by Ontario’s broker regulator, the Registered Insurance Brokers of Ontario (RIBO), did an unspecified numbers personal interviews with brokers and found “a widespread use of robotic process automation for back-office functions like data management, which can leverage AI.

“One of our interviewees has started to explore use cases that are more customer-facing, including using AI for marketing strategy and content, customer engagement (e.g., chatbots), and even identifying policy renewal options.”

Related: Brokerages are late to the party on AI investment: survey

One of the greatest benefits of AI in commercial insurance right now is its ability to synthesize and process massive amounts of data, thereby reducing duplicate entry, says IBC Insight Summit panellist Iman Arastoo, co-founder and chief operating officer of Insurmatics Inc.

“We know that, based on studies, about 60% of underwriting time is consumed by manual data intake, copying, data entry, typing, or something like that,” says Arastoo. “So…we try to provide a shift by [offering] agentic AI [solutions].

“And by agentic AI, I mean it’s not about AI chatbots. It’s not about search engines by AI. It means that we can execute workflows. We can automate workflows end-to-end.”

But as machines build up massive data sets from several unstructured data sources, it’s important for the humans to make sure they can track the data back to their original source, says James. That requires regular internal audits of AI databases.

Both James and Arastoo note the ideal use of AI in commercial lines would see agentic AI for the data, a human in the loop for decision-making, and a robust compliance policy for handling data.

“At this point, it’s it’s not totally automatic,” Arastoo says of current agentic AI solutions. “Let me emphasize it’s important to have a human in the loop. If we have a combination between agentic AI and human in the loop with a good policy for compliance, it could be a very good model, a new model, to do underwriting processes with a better way, less cost, and more profitability.

“The goal here is to eliminate administrative friction, administrative headaches, for underwriters and for the commercial insurance lifecycle. And so underwriters, by this approach, can focus on risk.”