Thirty Years of Regulatory Resentment Turned Out to Be Agent Infrastructure

Deb Fortunato is VP of Claims Operations at a mid-size property and casualty carrier in the Midwest. Or she would be, if she existed. She's a composite: built from interviews with claims leaders, industry surveys, regulatory filings, and the operational realities of a domain where every decision might end up in a deposition. Her opinions are grounded in real data and real tensions. Her name is her own.

We spoke over what she described as "the worst coffee in any industry that moves this much money."

Insurance claims processing was built, over decades, to answer three questions: Who approved this? On what basis? Can we prove it? The infrastructure that resulted — case-linked identifiers, reserve thresholds tied to approval gates, documentation requirements designed to survive litigation discovery — was constructed for human adjusters operating under regulatory scrutiny.¹ Nobody was thinking about AI agents.

But when agents arrived, something unexpected happened. The rails were already there. Not perfectly. Not completely. But closer than almost anyone had anticipated.

Deb Fortunato has spent the last eighteen months discovering exactly where those rails hold and where they buckle.

You've described claims infrastructure as "accidentally well-suited" for agent oversight. What do you mean by that?

Deb: Every claim gets a file number. That file number ties together the first notice of loss, investigation notes, reserve changes, approvals, correspondence, settlement. Everything. It's a bounded unit with an identity. When regulators or attorneys come looking, they pull the file. The file is the record.²

We built that because if you can't produce a defensible file during a market conduct exam, you have a very bad week. But it turns out that's exactly what you need when an agent takes an action. A container that says: here's what happened, here's who authorized it, here's the state of the world when the decision was made.

We also already had exception routing. Claims that fall outside automated rules get escalated with defined SLAs.³ We had approval gates tied to reserve thresholds. We had a documentation culture where adjusters kept notes because they knew the file might end up in court five years later. None of that was designed for agents. All of it transferred.

That sounds almost too clean.

Deb: Oh, it's not clean. That's the part nobody puts in the deck.

What broke?

Deb: The biggest surprise was tacit knowledge. Our claims manual says one thing. Our adjusters do something slightly different. And the difference is twenty years of knowing which coverage question to ask first in which state, which vendor's estimates always run high, which policy endorsement changes everything about a water loss claim. None of that was written down anywhere. When we tried to write agent instructions, we discovered that the process we thought we had documented was maybe sixty percent of the actual process. The other forty percent lived in people's heads and in hallway conversations that happened in 2014.

The other thing — and this one still keeps me up at night — is stale state. A policy gets endorsed six weeks ago. The endorsement hasn't propagated to the primary database. The agent reads the policy, makes a coverage determination, and it's confident and wrong. The approver sees the agent's summary, not the underlying data. They don't know the state is stale.⁴

In a demo, the action and the context sit right next to each other. You can see everything. In production, the action is downstream of permissions, old data, batch processes that ran overnight, a contact address that changed last Tuesday. That gap between demo and production is where the failures live.

The Cigna case — 300,000 claims denied in two months, 1.2 seconds of "review" per denial⁵ — comes up a lot in these conversations. Is that the nightmare scenario?

Deb: It's the known nightmare scenario, which makes it useful. Everyone in the industry studied it. Cigna showed what happens when you design a system where human review is technically present but operationally meaningless. A physician clicking through fifty denials in ten seconds is not reviewing anything. That's theater.

What worries me more is the subtle version. You don't need 1.2 seconds to rubber-stamp. You just need an adjuster who's seen forty agent recommendations before lunch, and thirty-nine of them were fine, so they stop actually reading the fortieth. Approval decays under volume. That's not malice or negligence. That's Tuesday.

How do you prevent that?

Deb: You have to make the approval surface show real information. Not just "approve/deny." The policy language the agent relied on, the fraud score at that moment, the reserve figure it was acting against, whether any of those inputs changed since the last human touchpoint. If the approver can't see the state the agent acted on, they can't actually judge the recommendation. They're just confirming that the agent produced output. Which is a different activity entirely.

And here's the part people don't want to hear: you have to accept that meaningful review is slower. You cannot have both "the human meaningfully evaluates every recommendation" and "we process claims 400% faster." At some point those goals collide, and someone in a conference room has to pick one.

You mentioned the adjuster's job changing. How?

Deb: [long pause]

The routine work is going away. FNOL triage that used to take four to eight hours now takes minutes.⁶ Document extraction, coverage verification on straightforward claims — agents handle that. What's left is the hard stuff. Coverage disputes, fraud indicators, aggressive attorneys, public adjusters inflating scope. The kind of claim where you earn your salary or lose the company a quarter million dollars.

So the job didn't disappear. It concentrated. And here's the problem nobody wants to talk about: the routine work was also the training ground. That's how junior adjusters learned judgment. By handling hundreds of straightforward claims until the patterns became intuitive. You develop a feel for when something's off. You learn to read a contractor's estimate the way a copy editor reads a sentence — you just see the error.

We're running about twenty percent annual attrition on adjusters. Each one who leaves takes roughly six years of institutional knowledge with them.⁷⁸ And the people replacing them don't get the same apprenticeship, because the apprenticeship work is automated. So you end up with fewer adjusters handling harder claims with less preparation. I don't love that math.

What do you still worry about?

Deb: Accountability passing. When the agent makes the recommendation and the adjuster clicks approve and the outcome is bad — who owned that decision? The agent vendor says the recommendation was sound given the data. The adjuster says they were reviewing too many claims to inspect each one deeply. Management says the process was followed. And I'm the one explaining to the state commissioner who was responsible.

If you had to name the one thing that surprised you most—

Deb: That the regulatory burden was the advantage. Every insurer I know has spent decades complaining about documentation requirements, audit trails, the whole compliance apparatus. Thirty years of resentment. Turns out that apparatus is exactly what you need to run agents responsibly. The domains where agents will actually work in production aren't the ones with the best models. They're the ones where the institution already knows how to say no, prove why, and survive the challenge.

We were accidentally building agent infrastructure for thirty years. We just called it "regulatory compliance" and resented every minute of it.

Deb Fortunato does not exist, but her problems do. The tensions she describes — approval quality under volume, stale state, the apprenticeship collapse, accountability passing — are drawn from industry surveys, regulatory filings, and the operational realities of claims processing in 2026. The coffee, however, is entirely fictional and probably better than whatever's actually in that break room.

Footnotes

1. Claims documentation requirements are designed to survive market conduct examinations, litigation discovery, and regulatory review. See Claims Journal, April 2026. ↩

2. Reserve thresholds and approval steps are tied to status transitions and governed for auditability. See Expert Insured — Claims Workflow and Statuses. ↩

3. Exception routing systems automatically assign exceptions to the right teams with defined SLAs and escalation paths. See ARDEM — Claims Workflow Optimization, 2026. ↩

4. The demo-to-production gap in claims includes policy endorsements not yet propagated, reserve figures out of sync across systems, and state-specific regulatory changes not reflected in agent data. ↩

5. Cigna medical directors were alleged to have denied over 300,000 claims in two months using the PxDx tool, with an average processing time of 1.2 seconds per denial. See MDEdge, August 2023 and Alignmt.ai, April 2026. ↩

6. Carriers deploying agentic workflows report FNOL-to-triage times dropping from 4–8 hours to under 5 minutes. See Insurance Thought Leadership, April 2026. ↩

7. Deloitte's interviews with P/C chief claims officers found average adjuster attrition around 20%, with turnover significantly higher among early-career adjusters. See Sedgwick, August 2025. ↩

8. Each adjuster departure costs roughly six years of institutional knowledge, with onboarding costs of $8,000–$10,000 per new hire — the majority going toward compliance and systems training rather than judgment development. See Openly, May 2026. ↩