Across P&C insurance, agentic AI is shifting the central promise of artificial intelligence from recommending the right next action to taking that action autonomously, pursuing optimized outcomes in the process. But that’s only if carriers can trust it to work safely, accurately, and at scale. In this three-part series, I’ll examine this leap from intelligence to action and what carriers need to make agentic AI valuable, governable, and worthy of that trust.

Part One focuses on claims, where I started my career as an adjuster. Part Two will turn to underwriting, where agentic AI can orchestrate risk assessment and pre-bind clearance checks aligned with carrier appetite, authority, and compliance requirements. And Part Three will pivot to distribution, including agentic customer acquisition across direct, producer, embedded, and conversational channels.

When the File Starts Moving the Claim Forward

Back when I was an adjuster, a claim moved only as fast as one person could read, call, calculate, document, and follow up with the right next move. Claimants like Andrew Wessels know the cost when that work stalls. Nearly a year after the January 2025 Eaton Fire damaged his home in Altadena, California, his family had been forced to move a dozen times and take on tens of thousands of dollars in debt. As he told NPR, “We haven’t taken one step forward.”

Catastrophes amplify the problem, but don’t create it. Capgemini found that 64% of adjusters struggle to meet turnaround targets amid everyday claims volumes. At one carrier processing more than 200,000 homeowners’ first-notice-of-loss (FNOLs) annually, adjusters spent up to 15 minutes per claim reviewing policy documents across multiple systems. Dead time like this has helped push average homeowners’ claims payout times to more than 40 days. In markets like the UK, a 2026 survey found one in four policyholders had switched carriers due to claims processing and settlement delays, and nearly a third planned to do so. On a global basis, that kind of sentiment puts billions in annual renewal premium at risk.

Agentic AI can help carriers avoid all this—compressing that dead time by autonomously extracting data from unstructured documents, retrieving coverage details, routing claims and document packages to relevant systems and people, and consulting a claims professional the moment human judgment is needed. But first, carriers need clarity on what an AI agent is and what it should be allowed to do.

What Makes an AI Agent ‘Agentic’?

Agentic AI uses individual AI agents, data sources, and approved tools to autonomously pursue a defined outcome across multiple steps: interpreting new information, planning, taking permitted actions, checking the result, preserving claim state, and escalating exceptions. A large language model might interpret documents; a predictive model might score severity or fraud risk. The AI agent uses probabilistic reasoning to turn that output into controlled work using claim context, approved tools, permissions, and escalation rules, while adapting its path as the claim changes.

Keys to Effective Agentic Claims

Agentic AI initiatives benefit from a modern core infrastructure that gives probabilistic AI agents trusted context and room to adapt while preserving deterministic control over policy and coverage records, reserves, payments, and audit history. The core supplies authoritative claim state and operational orchestration for workflows and transactions, enforcing financial controls, permissions, and regulatory rules. Agents may interpret, plan, recommend, and initiate permitted actions. Those that pass core controls and human approval, where required, become authoritative transactions.

Safe operation at scale is attainable via a cloud-native, API-first infrastructure. Agents need to read, write, and react concurrently through secure interfaces without bypassing core controls. An agentic framework connects selected models and agents to live insurance context and approved tools, providing testing, evaluation, audit tracing, and governance over agent activity. In this scenario, the core remains authoritative for claim state and transactions.

Build a Portfolio, Not a Super-Agent

With a proper framework in place, carriers don’t have to build every agent. Pre-built agents are emerging for common work such as conversational FNOL and claim summarization. Carrier-created agents encode differentiated practices; third-party agents contribute specialized capabilities. They should operate through the same core controls. In my view, this portfolio of agents should span key capabilities such as assignment and workflow; intake and triage; coverage and liability; fraud and SIU; investigation and evaluation; reserving and payment; negotiation and settlement; communications; recovery; and litigation. The critical element is an orchestrator agent that acts as a conductor by using shared claim state to control the workflow, invoke the right specialist agent at the right time, and trigger governed work through the core. Here’s how.

Give Every Agent an Operating Contract

Before an agent enters production, its operating contract can define:

• What claim outcome should it advance?
• What authoritative claim context may it use?
• Which approved tools may it call?
• What actions may it take and within what authority limits?
• When must it escalate?
• How will its actions be reconstructed and its performance evaluated for accuracy, cycle time, leakage, and customer outcomes?

The answers determine where it enters the workflow, what it can advance, and when a claims professional takes over. An agent may gather records and calculate exposure, but claims professionals retain final authority over coverage decisions, material reserve changes, fraud allegations, settlement offers, and large payments. The contract turns governance into agent guardrails.

Give the Portfolio an Orchestrator

Operating contracts govern what each specialist agent may do. For this operating model, an orchestrator agent determines when each should act. Like a conductor working from the carrier-defined life cycle, it maps work, monitors state and events, invokes the right specialist, passes authoritative context, and routes outputs through framework governance and core transaction controls before advancing work or involving a claims professional. Consider a liability recommendation: once a claims professional approves it, the orchestrator can initiate reserve review, recovery analysis, negotiation prep, and customer communication—each handled by a specialist agent within defined limits. When a medical report enters the claim, the orchestrator may invoke a summarization agent, then route relevant findings to liability or reserving agents. The sequence changes with the claim type, but the guardrails remain intact.

Measure Business Impact, Not Just Agent Adoption

As McKinsey’s Tanguy Catlin put it on my InsurTalk podcast, AI leaders are “obsessed with achieving impact” and work backward from the outcomes they seek. For claims, establish a predeployment baseline and measure changes in cycle time, payment latency, reserve accuracy, leakage, recovery, litigation, complaints, retention, and total cost per claim. Balance those results against overrides, escalation accuracy, and reopened claims so speed doesn’t conceal lower quality or added risk. Calculate value across the full life cycle after technology, integration, oversight, and change management costs. The question isn’t how many AI agents are live, but whether claims resolve faster, more accurately, and more cost-effectively—without compromising policyholder trust.

From Proof Point to Practice in the Agentic Age: What’s Next

Aviva has shown the value of treating AI as a claims operating model. Its 80-plus claims models cut complex liability assessment by 23 days, improved routing accuracy by 30%, reduced complaints by 65%, and saved $82 million in motor claims. It’s now pursuing larger, end-to-end agentic opportunities. Meanwhile, another carrier’s 12-country multiline rollout used agentic AI for straight-through settlement with minimal human intervention, handling times measured in minutes, and an eight-figure P&L impact.

Would agentic capabilities have changed the outcome for claimants like Andrew Wessels? It’s hard to say. What I do know is that AI agents would mean fewer claims stranded by handoffs, backlogs, and inaction while preserving human judgment for moments that matter. In P&C’s new Agentic Age, leaders need adaptable intelligence, trusted context, granular authority, a framework that traces agent activity, and a core that governs consequential transactions.

Wessels eventually received a contents payout “well into the low six figures” that he described as “a huge relief.” But relief shouldn’t require extraordinary persistence. Claims is where a carrier’s promise becomes either real or rhetoric. Agentic AI can help ensure that this promise is real.

In Part Two, I follow agentic AI upstream to underwriting—before risk reaches the book.