The Rise of the AI Colleague: What Agentic Systems Mean for P&C Insurance

The conversation around AI in insurance has shifted.

For years, the question was whether AI would find a meaningful role in underwriting and claims. That debate is largely over. The question now is how quickly insurers can move from experimentation to production, and what it actually takes to get there without losing control of the decisions that matter.

On July 22, Insurance Insider brought together four senior executives from across the P&C value chain to examine where agentic AI stands today. Nemanja Jokic, CTO of BoundAI, joined Will Hyams, Head of AI at Howden, Kathleen Ziegler, COO for the Americas at AXA XL, and David Hurst, Chief Technology Innovation Officer at CRC Group, for a discussion that covered everything from governance and human oversight to where the technology is already delivering measurable results.

What follows are the key themes from that conversation.

From Experimentation to Production

The panel opened with a question that cuts to the heart of where most insurers find themselves right now: what does success with agentic AI actually look like?

The consensus was that the industry has moved past the pilot phase in at least some areas, but that moving from a contained experiment to something running in production at scale is a different challenge entirely. The gap between a working proof of concept and a reliable operational system is where most implementations run into difficulty.

Kathleen Ziegler framed it in terms of where the technology is genuinely delivering versus where it is still finding its footing. "We are seeing real value in specific, well-defined workflows where the inputs are structured and the outputs are measurable. The challenge is that insurance has a lot of workflows that don't fit that description."

The distinction between tasks that are ready for semi-autonomous systems and those that aren't came up repeatedly. Claims triage and submission intake were identified as areas where agentic AI has demonstrated consistent value, largely because the logic is definable and the volume is high enough to justify the investment. Underwriting judgment, by contrast, was treated with more caution.

David Hurst pointed to the importance of starting with the right problem. "The organizations that are getting results are the ones that picked a specific, high-volume process and went deep on it rather than trying to automate everything at once."

Which Processes Are Actually Ready

Not every insurance workflow benefits equally from agentic AI, and the panel spent considerable time on where the technology is genuinely ready versus where enthusiasm is running ahead of the evidence.

Submission intake and document processing were the clearest examples of processes that have crossed the threshold into reliable production use. The logic is consistent, the volume is high, and the cost of manual processing is well understood. When a system can read an ACORD form, extract the relevant fields, check for duplicates, and route a cleared submission into an AMS without human intervention, the value is immediate and measurable.

Nemanja Jokic was direct about where BoundAI has seen the most consistent results. "The workflows that work are the ones where you can define what good output looks like. Submission intake, policy checking, inspection review - these have clear success criteria. You either extracted the right data or you didn't. You either caught the discrepancy or you missed it."

Claims triage was identified as another area with strong near-term potential, particularly for high-volume, lower-complexity claims where the triage decision follows a defined logic. Distribution and placement were treated as more complex, given the relationship-driven nature of the broker market and the degree of judgment involved in market selection.

Will Hyams noted that readiness is often less about the technology and more about the data and process discipline underneath it. "Before you can automate a workflow, you need to understand it well enough to describe it precisely. A lot of organizations discover, when they try to implement AI, that their processes are less defined than they thought."

The Human Oversight Question

If there was one theme that ran through every part of the conversation, it was this: agentic AI works best when human judgment stays in the loop at the right points, not at every point.

The distinction matters. Requiring human review of every output defeats the efficiency argument. Removing human review entirely creates accountability gaps that insurers, regulators, and capacity providers are not ready to accept. The challenge is designing systems that know the difference between a decision they can make reliably and one that needs a person.

Kathleen Ziegler put it plainly. "The question is not whether humans should be involved. They should be. The question is where their involvement adds the most value and where it is just adding friction without adding judgment."

The concept of confidence scoring came up as a practical mechanism for managing this. When a system processes a submission or reviews a policy, it can assign a confidence level to its own output. High-confidence results move through automatically. Low-confidence results get flagged for human review. The human's time goes to the cases that actually need it rather than being spread across everything equally.

Nemanja Jokic described how this plays out in production. "The goal is not to remove the underwriter from the process. It is to make sure that when an underwriter's time is required, it is because there is a genuine judgment call to make, not because the system encountered a document format it had not seen before."

David Hurst added that the design of the oversight mechanism shapes how much trust the organization can place in the system over time. "If your exception routing is well-designed, you build confidence in the automation incrementally. Every time a human reviewer looks at a flagged item and confirms the system was right to flag it, that is data that improves the next decision."

Governance When AI Acts on Your Behalf

The accountability question is one the industry has not fully resolved, and the panel did not pretend otherwise. When an AI system flags a risk, routes a submission, or drafts a recommendation letter, someone is still responsible for the outcome. The governance frameworks that define who that is, and how they can demonstrate they fulfilled their responsibilities, are still being built in most organizations.

Regulators are paying attention. The panel noted that supervisory bodies in both the US and UK markets are increasingly focused on how insurers document AI-assisted decisions, particularly in underwriting and claims. The concern is not that AI is being used, it is that insurers may not be able to explain what the system did and why when a decision is challenged.

Kathleen Ziegler was direct about the organizational implications. "You need to be able to show your work. If an AI system contributed to a coverage decision or a claims outcome, you need an audit trail that a regulator or a reinsurer can follow. That is not optional."

Will Hyams framed governance as a design requirement rather than a compliance afterthought. "The organizations that will get this right are the ones building auditability into the system from the start, not the ones trying to retrofit documentation onto a system that was never designed to produce it."

For delegated authority frameworks specifically, the documentation requirement is acute. MGAs operating under binding authority need to demonstrate to capacity providers that their underwriting process is consistent, controlled, and auditable. An AI system that produces structured, logged outputs at every decision point is a stronger position than one that relies on individual underwriters to document their own reasoning.

How Agentic AI Reshapes the Value Chain

Beyond individual workflows, the panel turned to a larger question: what happens to the relationships between carriers, MGAs, and brokers as agentic AI becomes more embedded in day-to-day operations?

The short answer is that speed and data quality become competitive differentiators in ways they haven't been before. When one MGA can clear and respond to a submission in under a minute and another takes two days, brokers notice. When one carrier receives clean, structured data from every submission and another receives whatever arrives in the email, the underwriting quality reflects that difference over time.

David Hurst suggested the shift is already visible in parts of the market. "The brokers who are getting the best terms fastest are the ones whose submissions arrive in the best shape. That is not new, but AI is widening the gap between the operations that have figured this out and the ones that haven't."

The panel also touched on how agentic AI changes what carriers and MGAs can reasonably expect from each other. When data moves through a structured pipeline rather than email and spreadsheets, the tolerance for inconsistency drops. Systems that can validate every field at intake make manual workarounds harder to justify and easier to eliminate.

Nemanja Jokic connected this back to where BoundAI sits in the value chain. "We are not trying to replace the relationship between a broker and a carrier. We are trying to make the work that happens between them faster and more reliable. The judgment and the relationship stay human. The coordination around them does not have to be."

Conclusion

The conversation that took place on July 22 covered a lot of ground, but the through-line was consistent. Agentic AI in P&C insurance is past the point of being theoretical. It is running in production, delivering measurable results in specific workflows, and creating real competitive separation between the operations that have embraced it and those still working through the decision.

What the panel made clear is that the technology alone is not what determines whether an implementation succeeds. The process discipline underneath it, the governance design around it, and the human oversight built into it matter just as much. Organizations that treat agentic AI as something to bolt onto existing workflows will get limited results. Organizations that redesign the workflow first and deploy the technology on a clean foundation are the ones seeing the numbers move.

Will Hyams summed up where the industry stands. "We are still early, but we are not as early as some people think. The gap between leaders and laggards is already opening up, and it will be harder to close in two years than it is today."

For carriers, MGAs, and brokers trying to figure out where to start, the panel's collective advice was consistent: pick a specific, high-volume process, define what good output looks like, build the oversight mechanism in from the beginning, and measure the results. The organizations doing that are already ahead.

You can watch the full webinar here.

‍