What AI Actually Feels Like Inside Insurance Right Now

In the spring of 2026, an AIG executive stood before analysts and described what AI had done to his underwriting operation: time-to-quote down 55%, binding rate up 40%, more submissions reviewed with better information in less time. The numbers were real. The slide was clean. If you read the transcript, it sounds like the future arrived ahead of schedule.

Around the same time, a veteran adjuster posted something on Reddit that stopped a lot of people mid-scroll. He'd been using AI to summarize medical records — genuinely useful, he said, cutting through hundreds of pages in seconds. But then he described a claim where the AI missed a single detail: Metformin in the medication list. A laceration on the foot. A diabetic patient whose risk profile changed completely once you knew that one fact. The AI had handed him a clean summary. He'd almost moved on. "It doesn't pick up little nuances," he wrote. "That's the stuff that changes everything."

Both of those things are true at the same time. The official story of AI in insurance and the frontline story are not the same story. If you work somewhere in that gap — which is where most insurance professionals actually live right now — this article starts there. Not in the boardroom. Not in a vendor pitch deck.

To understand how that gap opened, and whether it's going to close in your favor or against you, you have to look at what AI is actually doing to specific jobs, one workflow at a time. The picture is more uneven than either the optimists or the doomsayers admit.

The Administrative Layer Is Going First

AI is automating the administrative layer of insurance work — first notices of loss, document summarization, coverage checks, routine correspondence. That's not a prediction. It's already happening at measurable scale.

Lemonade is the clearest example of what this looks like when it runs at speed. Their cost per claim dropped from $65 in 2020 to $19 by 2025, with 55% of claims now fully automated from start to finish. That's what the efficiency curve looks like when AI handles high-volume, standardized work. The boardroom loves this number.

But here's the number the boardroom doesn't lead with: the industry average for straight-through processing across U.S. P&C insurers sits below 10%, and nearly 60% of insurers report zero straight-through processing in their claims operations. For most insurance workers, the "AI everywhere" headline is running several years ahead of actual floor deployment. The daily reality is bolt-on tools, copy-paste between systems, and rising KPIs that assume the AI is working better than it is.

The deeper problem isn't the technology. It's what automation does to the path into the profession. Ten years ago, a first-year adjuster closed 40 soft-tissue claims before touching a complex bodily injury file. They weren't just learning procedure — they were learning what normal looks like, building the mental baseline that lets an experienced adjuster feel when something is off before they can articulate why. Today those 40 files are automated. The new adjuster's first real case is their first hard case. As one industry observer put it bluntly in Claims Journal: "Automation didn't eliminate the role. It removed the path into it."

This pattern isn't limited to adjusters. Underwriters are watching AI pre-screen submissions and score risk appetite. Agents are seeing conversational tools handle intake and FAQ calls. Customer service reps are being backed up by bots that answer "where is my check?" at 2 a.m. In every function, the same question applies: what does your role look like once the routine is gone?

Knowing what's automating is only half the picture. The harder question is: who inside these companies is actually getting ahead of it — and what separates them from the people who are waiting to see what happens?

The Early Adopters Who Don't Look Like Early Adopters

The most instructive figures navigating AI inside insurance right now are not the insurtechs built from scratch on machine learning. They're the veteran practitioners who got curious and started building before anyone told them to.

Ty Robben spent 14 years in casualty underwriting before joining Palomar as Senior Vice President. He's described on The Insurance Podcast as someone who "leans more to the old-fashioned way of doing things." He didn't have a data science background. He didn't have a corporate AI initiative behind him. He started building his own process solutions for casualty underwriting anyway, journaling the process publicly, and bucked the industry trend of outsourcing AI development to vendors who had never read a complex liability submission. He isn't a techie. That's the point.

On the other end of the spectrum, Jeff Sutton, SVP at Markel Canada, leads a team that uses AI to strip the drudgery from submissions specifically so underwriters can "just pick up the phone and call." Markel literally placed orange rotary phones in offices to reinforce the behavior. The technology serves the relationship, not the reverse. It's a different approach from Robben's, but the underlying logic is the same: AI is a tool you direct, not a wave you ride.

I believe Gen AI has the potential to humanize insurance by consolidating and querying information, but the critical step is still the human interaction, something that's becoming a lost art in the Canadian P&C industry.
— Jeff Sutton, SVP Sales & Marketing, Markel Canada

What makes Robben's approach especially relevant is a gap that shows up in the data. According to MoneyGeek's 2026 industry analysis, 92% of insurance workers say they want AI training. Only 4% of carriers invest in reskilling at scale. Waiting for your employer to hand you an AI strategy and train you on it is a documented losing position. The workers navigating this transition well are treating AI literacy as a personal professional project — the way earlier generations treated getting a CPCU or a CPA.

This applies equally to the agency manager testing EZLynx's AI autofill tools, the claims manager learning to interrogate a Claim Summarizer output, and the underwriter who starts running their submission pile through an AI pre-screen. The entry point is a workflow you already own, not a course you have to take first.

But curiosity alone isn't enough. There's a specific reason why insurance's AI story is harder to navigate than most industries — and it has nothing to do with technology. It has to do with the law.

Your File Is Now a Governance Document

Insurance's dense regulatory environment is the single biggest reason AI cannot simply replace human workers in this industry. It is also creating a new category of job risk that most frontline workers don't know they're carrying.

The documentation trap works like this: plaintiff attorneys are now hunting for moments where an adjuster relied on an AI output — a denial recommendation, a fraud flag, a damage estimate — without documenting how the system works or who validated it. A claim file with an AI recommendation and no human override notation is a discovery liability. The question your file must answer is not just "what did the AI say?" but "why did the human agree — or disagree — and on what basis?"

Florida's HB 527, which passed committee in December 2025, makes this explicit. The bill prohibits using an algorithm or AI system as the sole basis for denying or reducing a claim. A human professional must independently analyze the facts and certify that AI was not the lone decision-maker. That certification lives in your file notes. Whether you know it or not, you are already signing something every time you close a claim.

The regulatory perimeter is expanding fast. As of mid-2026, 28 states have enacted some form of AI regulation for insurance — requiring fairness testing, explainability, and documented human oversight. The NAIC's AI Systems Evaluation Tool is being piloted in 12 states, with full adoption expected at the 2026 Fall National Meeting. Regulators are coming to examine these files.

For the adjuster writing denial letters, the underwriter approving an AI-scored submission, or the agent relying on AI-generated coverage recommendations, the practical question is the same: if this decision were examined in discovery or a market conduct exam, what does my file show about my independent judgment? Answering that question is now a core job skill, not a compliance formality.

Regulation explains why humans aren't going anywhere. But it doesn't explain who will thrive in the roles that remain — and who will find themselves doing more work for the same pay while the AI gets the credit. That comes down to what you do with the time AI gives back.

The Honest Ledger

The AI disruption in insurance is genuinely split — real efficiency gains in high-volume standardized work, real failures in nuance-dependent judgment — and where you land on that ledger is not fixed.

Where AI is delivering: the AIG metrics from the opening aren't a projection. They're in production, across eight lines of business. Gallagher Bassett's Claim Summarizer and Email Sentry are reducing the time claims managers spend on routine file review. Lemonade's cost-per-claim trajectory shows what the efficiency curve looks like at full scale. These are real wins, and pretending otherwise doesn't serve you.

Where the gap is real: 82% of insurers believe AI will dominate the industry's future, but only 14% have fully integrated AI into their financial operations. The integration gap described earlier — average STP below 10%, bolt-on tools, fragmented data — is the actual daily experience for the majority of insurance workers in 2026. The floor is not the earnings call.

We start by measuring what our clients care about, and then putting innovations in place that actually move the needle.
— Joe Powell, Chief Digital Officer, Gallagher Bassett

Joe Powell, Chief Digital Officer at Gallagher Bassett, captured what separates the deployments that work from the ones that don't. His team involved former adjusters as AI specialists to ensure the tools reflected real-world workflows. The workers who shaped those tools trust them. The workers who had them handed down don't. "We start by measuring what our clients care about," Powell said, "and then putting innovations in place that actually move the needle." That's a management insight, but it's also a worker insight: the people who put their hands on the tools early — who test them, critique them, and document what they get wrong — become the ones shaping what comes next.

The workers who will come out of this transition ahead are not necessarily the most technical. They're the ones who use the time AI saves them to do the one thing AI demonstrably cannot: build relationships, exercise judgment in genuinely ambiguous situations, and carry institutional knowledge that no model trained on historical claims data can replicate.

Which brings us back to where we started — and to a 14-year underwriting veteran who decided not to wait.

What "Getting Ahead of It" Actually Looks Like

Ty Robben didn't have a data science degree. He didn't have a corporate AI initiative behind him. He had 14 years of underwriting instincts, a specific workflow problem, and the decision to start tinkering before anyone told him to. That's the version of this story that's available to almost everyone in this industry — not the Lemonade version, not the AIG version, but the version that starts with one workflow you already own.

The workers who will navigate this transition well aren't the most technical. They're the most curious and the most willing to put their hands on the tools before anyone tells them to. The skills that AI cannot replicate — judgment in genuinely ambiguous situations, institutional knowledge, human relationships under stress — are worth more now than they were three years ago. But only if you're building them deliberately, not just defending them rhetorically.

This week: pick one workflow in your current role that involves summarizing, sorting, or drafting information — a coverage review, a client status update, a submission pre-screen. Spend two hours testing whether an AI tool changes it. Document what it gets right and what it misses. That documentation is your starting point. That's the whole assignment.

The gap between the earnings call and the adjuster's desk is real. What you do inside that gap is still yours to decide.

Recommended Tools & Resources