Agentic artificial intelligence is forcing insurance boards to rethink how they allocate responsibility and oversight as decision‑making shifts from people to software, according to Paulo Salomao (pictured), national lead of strategy and consulting at Accenture Canada.
“In the past, we used a lot of technology, but it was very passive,” Salomao said in an interview with Insurance Business. “With agentic AI, it’s a little bit different because now it’s active, it can decide what to do and when to do it,” he said. “It actually gains decision authority that used to sit with people.”
Most insurance governance is still built around human roles and titles – who holds the job, who signs off, who is accountable for an end‑to‑end process. Salomao said that premise starts to break down when AI systems own discrete parts of the work.
“If our current model of governance, that is predicated on the job, the title, the accountability, if that no longer holds, then I think the boards need to adjust accordingly,” he said. “And that’s where I think the biggest change is happening.”
Rather than replacing entire positions, agentic AI is being embedded into specific, repeatable tasks across underwriting, claims and servicing, he said. That changes the “unit” that governance frameworks have to track.
“The difference is AI isn’t taking over the job beginning to end,” Salomao said. “AI takes over tasks that make up the jobs. And that shift in the unit of accountability is what drives the entire upheaval, if you will, in governance. Because now you have to govern at the task level as opposed to the overall job level.”
Salomao said the shift is already underway in core insurance functions, not a future scenario.
“I think it is happening now,” he said. “For example, underwriters used to spend a lot of time gathering information and then making a decision. Today the information is consolidated mostly by AI. And the underwriters just make the final decision based on the data that is presented to them.”
On the claims side, he said, simple, high‑volume loss types are increasingly being handled with little or no human touch.
“If you look at claims, we’re seeing more and more of simpler claims become auto‑adjudicated,” he said. “And many times the auto adjudication happens with the help of AI, and the humans focus on the more complex claims or exceptions.”
He pointed to life insurance and group benefits to illustrate how regulation sets the limits. In life insurance, a licensed individual still has to make the final call, whereas in group benefits, a carrier can run a high share of claims straight through with minimal human touch, depending on what the rules allow in each segment.
The technical building blocks aren’t new, he noted. Carriers have been using rules engines and workflow tools to automate simple decisions for years. The change now is in the complexity of tasks that can be handed off.
“The idea is not new,” Salomao said. “In the beginning, you could automate based on simple rules. Let’s say your auto adjudication was in the single low digits. But now you’re able to increase the complexity because you have these agents that can perform more complex tasks.”
He said some Canadian players are already moving ahead, though they still trail the US market.
Salomao said insurance is both more exposed to and better positioned to benefit from agentic AI than many other sectors because it ticks three key boxes: it runs on highly repetitive decisions, it sits on large volumes of data and it has clear economic signals that make it easier to judge where automation makes sense versus where humans should stay in charge.
By contrast, he noted, industries such as transportation have messier, less standardised decision flows, while insurance operations are comparatively modular and easier to break down into discrete, machine‑executable tasks.
That modularity is what pushes the governance problem down a level, he said. Boards and executive teams used to map accountability by function – head of claims, chief underwriting officer, CIO – and assume that those leaders controlled relatively coherent blocks of work. In an agentic setup, boards have to understand which tasks are being executed by which classes of models and what controls sit around them, Salomao said.
The issue becomes more acute as agentic AI extends from high‑volume retail products into small commercial business, where data is richer and risks repeat more frequently than in large corporate accounts.
“Small commercial has meaningful data and we’re starting to see some application on that, especially when it comes to adjudication or underwriting,” he said.
