Download

The Contradiction for Insurers Within AI

AI is retiring risks the market already prices and creating risks it does not. That imbalance matters more than any bottom line.

Insurance

AI is mitigating risks that are already modeled, already priced, and already reserved for. It is creating risks that are none of those things. Cyber loss, claims fraud, catastrophe exposure, and AML false positives all have loss histories behind them. Algorithmic discrimination, model drift, hallucinated advice, and correlated foundation-model failure do not.

That imbalance is why the industry's behavior looks contradictory. Carriers are deploying AI aggressively inside their own operations while moving just as fast to exclude it from their customers' policies. Both are rational. They simply belong on different ledgers.

WHAT AI CREATES

Three new exposures — and no existing control framework built to support them. 

The first is model risk migrating into regulated decisions. Roughly two dozen states and D.C. have adopted NAIC's AI model bulletin, requiring a written AI program, board-level accountability, and discrimination testing. New York's Circular Letter No. 7 goes further, naming fairness metrics and requiring written explanation of adverse decisions within 15 days. Colorado has extended its AI governance regulation to auto and health; nine states now restrict AI in prior authorization and claims review. In the litigation that follows — Huskey v. State Farm, the NH Predict class actions against UnitedHealth — courts have shown no appetite for treating an algorithm as protection. This is model risk with lingering negative consequences and a civil-rights overlay that traditional model risk management never had to price.

The second is third-party AI risk, which is concentration risk wearing a new coat. Swiss Re Institute warned in January that "concentration among cloud and AI service providers creates accumulation and systemic risk concerns"; the Bank of England's July Financial Stability Report modeled correlated disruption through shared providers. Gallagher Re made the insurance point: unlike a windstorm, an AI failure spreads through the shared use of a widely adopted model, producing simultaneous losses across industries and jurisdictions. That is aggregation, and aggregation is how insurers get hurt.

The third is the silent exposure sitting inside policies written before any of this. Gallagher's 2026 benchmarking survey found one in five insurance professionals reporting a client loss or claim tied to AI risk, only 51% of them fully covered. The market is closing that gap by subtraction. ISO's three generative AI exclusionary endorsements — CG 40 47, CG 40 48 and CG 35 08 — took effect Jan. 1, and Berkshire Hathaway, Chubb, Travelers and W.R. Berkley have all filed exclusions. Affirmative capacity exists — Armilla, Chaucer, Mosaic, Munich Re's aiSure — but tops out around $25 million to $50 million per insured. Anthropic settled a single copyright matter for $1.5 billion and paid it off its balance sheet. That is the protection gap in one line.

WHAT AI RETIRES

The mitigation case is strongest where the loss data is best. IBM's 2026 Cost of a Data Breach study, across 602 breached organizations, found firms using security AI and automation extensively cut breach costs by nearly $2 million and contained incidents 65 days faster than those using none. HSBC reports its AI transaction monitoring identifies two to four times more financial crime than the rules-based system it replaced, with 60% fewer false positives — the rare compliance result where accuracy and efficiency moved together.

In claims, Allianz Partners cut its travel claim lifecycle from 19 days to four, and Intact Financial has disclosed more than C$150 million in recurring annual AI benefits. And in catastrophe risk, the gains are peer-reviewed, not self-reported: Google DeepMind's cyclone model, published in Nature in August, delivers three-day track forecasts as accurate as the two-day forecasts that preceded it — a full extra day of warning for every property book on a coastline.

THE MEASUREMENT PROBLEM

Risk managers have registered the shift. In January's Allianz Risk Barometer, AI entered the global top three for the first time, at No. 2, up from No. 10 — the sharpest single-year climb the survey has recorded. But registering a risk and pricing it are different things.

Measurement makes the imbalance worse. Evident's 2026 AI Index found just 65 publicly documented AI use cases with disclosed outcomes across the 30 largest North American and European insurers — 75% productivity gains, 2% revenue uplift. FactSet reviewed seven insurers that discussed AI on second-quarter earnings calls; none disclosed a dollar or basis-point figure tying AI investment to results. IBM's own data adds a caution: even as security AI saved money, average time to contain a breach rose to 247 days, reversing five years of improvement, with one in four malicious breaches now AI-enabled. AI is arming both sides.

That is not an argument against adoption, but for treating AI the way the market eventually treated cyber. The precedent is on the books: Lloyd's Market Bulletin Y5258 forced underwriters, in July 2019, to state affirmatively whether cyber was in or out. No equivalent mandate exists for AI yet, and until one does, exclusions will keep arriving faster than coverage.

That correction may come. If it does not, if the exposure proves too correlated for loss history to tame, then the conversation goes somewhere the industry has been before. Nuclear got Price-Anderson, terror got TRIA, flood got the NFIP, each after the market judged a risk real but not privately bearable. One proposal would apply that logic to frontier AI: a mandatory, risk-priced federal indemnity pool with a sunset. The sharpest objection is that it solves the wrong problem. Ian Adams argues the bottleneck is not capital but legal uncertainty: no one knows what courts will hold AI developers and deployers liable for, and mandating cover over an undefined duty of care subsidizes that uncertainty rather than resolving it. His remedy is clearer statutory duties and safe harbors. Insurers can price a loss once they know what it is.

For risk managers, the work is unglamorous: an inventory of every place AI touches a challengeable decision, a close reading of the AI language in each D&O, E&O, cyber and GL renewal, and vendor contracts written on the assumption that the model will fail. The coverage conversation has started without them.

SOURCES

Allianz Commercial, Allianz Risk Barometer 2026 (Jan. 14, 2026; 3,338 respondents, 97 countries).

NAIC, Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, and NAIC implementation map (as of April 2026).

NY DFS, Insurance Circular Letter No. 7 (July 11, 2024). Colorado DOI, 3 CCR 702-10, as amended effective Oct. 15, 2025.

Huskey v. State Farm Fire & Cas. Co., No. 1:22-cv-07014 (N.D. Ill.); Estate of Lokken v. UnitedHealth Group (D. Minn.).

Swiss Re Institute, sigma insights 01/2026: AI adoption is reshaping the risk landscape (Jan. 13, 2026).

Bank of England, Financial Stability Report (July 7, 2026).

Gallagher Re, Global insurtech Report: Artificial Intelligence — Risks and Opportunities in (Re)insurance and Beyond (March 2026).

Arthur J. Gallagher & Co., 2026 AI Adoption and Risk Benchmarking Survey (Feb. 23, 2026; 1,200+ businesses).

Verisk/ISO endorsements CG 40 47, CG 40 48, CG 35 08, effective Jan. 1, 2026.

IBM Security / Ponemon Institute, Cost of a Data Breach Report 2026 (July 29, 2026; 602 breached organizations).

HSBC, "Harnessing the power of AI to fight financial crime" (June 10, 2024).

Allianz Partners (Sept. 2025); Intact Financial Investor Day (May 21, 2025).

Google DeepMind, WeatherNext cyclone model, Nature (Aug. 6, 2026).

Evident Insights, 2026 Evident AI Index for Insurance (June 16, 2026).

FactSet Insight, "Insurance AI Investments Emerge So Far as Qualitative Comments vs. Quantitative Drivers" (Aug. 14, 2026).

Lloyd's Market Bulletin Y5258 (July 4, 2019).

Gregory C. Allen, The Insurance Industry's Retreat from AI Threatens to Slow Innovation and Adoption, CSIS (Sept. 4, 2026).

Cristian Trout, Insuring Uninsurable Risks from AI: Government as Insurer of Last Resort (arXiv 2409.06672).

Ian Adams, AI Insurance and the Backstop Reflex: A Policy for Every Risk, a Taxpayer for Every Loss, Truth on the Market (Oct. 8, 2026).


Ron Shriberg

Profile picture for user RonShriberg

Ron Shriberg

Ron Shriberg is a client partner with NewRocket. 

He has over two decades of experience serving Fortune 100 clients in technology adoption. Before NewRocket, he spent 13 years as a senior director, ServiceNow Financial Services at KPMG US.

Shriberg holds a B.S. in business management with a minor in darketing from Bentley College.

 

Need an Idea for Innovation? How About 1,208,580?

The founder of Idealab has used AI to generate more than 1.2 million ideas for innovative AI-based companies, including in insurance. They're thought-provoking.

Image
Insurance

Bill Gross recently announced a free database of AI-generated ideas for innovative companies, based on the 31 methodologies that he's developed over the past three decades and that have powered his Idealab incubator. 

The list of ideas — currently totaling a whopping 1,208,580 — may or may not contain an idea that you'll be able to turn into big bucks, but I think that exploring the database and interacting with the ideas is worth a bit of your time. 

Many of the notions I see for the insurance market are provocative. They can at least suggest where there are holes in insurance offerings today that you should work to fill, even if you don't launch the sort of standalone company that Gross would love to see join his pioneering incubator.

Let's have a look. 

Here is a link to the database. And here is a link to a Fortune article that describes the effort in detail, including Gross's argument that innovations driven by AI will create far more jobs than it will eliminate. The article says Gross's prompts to the AI break down into three categories: 

"Opportunities, which he says are things such as solving a business pain point, removing friction, removing waste, removing danger, and improving design, among others; enablers, such as new technologies, scientific advances, and new materials, processes, or changing cost curves that make solutions more feasible than before; and shifts, such as regulatory changes, as well as behavioral, economic, cultural, or geopolitical movements that make the timing for a particular business seem right. "

Gross also fed his AI "some 17 million source documents that included news stories, research reports, expired patents, public company filings, and threads from online community discussion boards" and came up with his 1.2 million ideas for new businesses.

But are the ideas any good?

I certainly found them thought-provoking. 

For instance, I selected the "fintech/insurtech" option, specified that I didn't want to invest more than $250,000, wanted to get to market within a year and hoped to "make the insurance claims process smoother for customers." The top suggestion, which the AI suggested was a 57% fit for me, was for a "claims trust operating system that... gives policyholders visibility into the evidence shaping settlements and repairs."

I'm not sure that's a business. It would take an awful lot of work to assemble the huge amount of settlement and repair data that would be needed to be truly definitive, but the idea raises an intriguing notion: Somebody will surely try to use an AI to give policyholders a quick handle on what the insurer will likely offer after a car accident, water damage to a home, etc. If the insurance industry provides the initial take, it will be able to lower expectations and control payments. If plaintiff attorneys fill the need, they'll be able to raise expectations, influencing policyholders to be disappointed... and perhaps hire an attorney to contest the offer. 

The competing interests would seem to create an opportunity for a trusted intermediary, but I, at least, have become skeptical about whether it could stay unbiased. I subscribe to the notion laid out by Cory Doctorow in a seminal blog post republished by Wired magazine in 2023 that describes the process of — excuse my French — enshittification. 

As I wrote at the time, Amazon is a great example. It initially built a huge base of customers by doing everything possible to make life simple for them. Once it had the customer base, it courted businesses — and the customer experience started to get gunked up by ads, attempts to keep customers from easily finding the best deals, etc. Once Amazon had built a marketplace where companies had to have a presence, it started milking them as much as possible on behalf of investors — a fee here, a fee there, maybe Amazon would even use its insights into sales data to launch products or businesses that would undercut a company with a hit offering on Amazon's platform. (Amazon has consistently denied doing so but has faced many lawsuits alleging it does.) Just about everybody still uses Amazon, but the experience is very different than it was a decade ago.

Even if I'm right that there isn't a business in a "claims trust operating system," the Idealab database has still pointed to a sore spot that the insurance industry needs to address. Somebody is going to use AI to shape policyholders' thinking about what to expect as a settlement, by providing near-instant estimates. So insurers should at least be thinking about how they show their work, and quickly, to keep plaintiff attorneys from setting unrealistic expectations and accelerating the trend toward more and more litigation. 

So I'd recommend that, whatever your role in the insurance ecosystem, you poke around in the database to see what friction points and opportunities it has surfaced that you might be able to address. There are more ideas than I can count about underwriting, pricing, services from agents and brokers, and on and on. 

If you're like me, your eyes glaze over after half an hour or so of an exercise like this, so maybe make a visit to the database an occasional excursion over the next month, perhaps at the end of the day as you're winding down or before a weekly meeting with colleagues so you can stir up discussions. We all get locked into our mindsets and into our day-to-day routines, and Gross has provided a whole host of prompts that can challenge us to do ever more than we are now.

Cheers,

Paul

When AI's Speed Endangers Software Development

As AI agents autonomously draft code and deploy changes, insurance carriers must anchor governance to human accountability before speed becomes a liability.

Insurance

Picture a product feature built almost entirely with AI. A product manager uses an AI assistant to turn a business request into user stories and acceptance criteria. Developers hand those stories to a coding agent. QA generates the test suite from the same stories. DevOps asks AI to update the pipeline that deploys it. Every test passes, every check is green, and the change ships on schedule.

Now suppose the original story misread one rule: A limit meant to apply to each coverage was written as applying to the policy as a whole. Every downstream AI faithfully built, tested, and deployed that misreading. Nothing failed, because every step checked its work against the step before it. The first sign of trouble would be a claim paid against the wrong limit. So where, exactly, was the governance?

In 20 years of building software, I have not seen anything disrupt the development and architecture lifecycle the way AI has. It is changing how we work, faster than most organizations can absorb, and that speed carries real risk. Every team is adopting AI in its own way, with its own patterns, tools, and agents, trying to find what works for them. The natural response is to standardize. But is standardization even the answer?

Why insurance raises the stakes

Insurance has very little tolerance for error. It is one of the most heavily regulated and litigated industries, and the cost of getting things wrong goes well beyond a bug fix: regulatory fines, market conduct actions, policyholder lawsuits, reputational damage, and in serious cases, a threat to the license to operate.

Rating logic, underwriting rules, and policy forms are often required to match what an insurer has filed with state regulators, so a code change can quietly become a compliance gap. Decisions that affect policyholders face growing expectations of fairness and explainability. Carriers also operate under regular audits and regulatory examinations that expect documented, accountable approvals.

Regulators are extending those expectations to AI. The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023, and more than half of U.S. states have since adopted it. The bulletin expects insurers to maintain a written governance program for AI that makes or supports regulated decisions, and regulators can examine that governance during market conduct reviews. Some states go further: Colo. and N.Y. apply their own frameworks, which add specific requirements such as testing AI systems and external consumer data for unfair discrimination.

The same logic should extend to AI that builds the systems making those decisions. If a model's output needs governance, so does the code that encodes a rating rule or a policy limit.

The autonomy ladder

A simple way to think about AI in software delivery is as a ladder of autonomy:

LevelNameWhat the AI doesWho owns the change
1SuggestOffers completions and snippetsA human writes and owns it
2DraftGenerates a full changeA human reviews and merges it
3Gated approvalAdvances change through automated gates, with humans holding defined approval boundariesGate rules plus named human approvers
4AutonomousPlans, changes, and deploys with no human approval boundaryThe process itself, unless a human owner is explicitly defined

The last column is the one that matters. As autonomy rises, it gets harder to name who owns the change.

Teams do not always move up this ladder deliberately. Many are already drifting from Level 1 to Level 2 as coding assistants shift from suggesting lines of code to generating complete changes, tests, and pull requests. Once agents can invoke tools, modify multiple components, or satisfy automated merge gates, the workflow is effectively operating at Level 3, even if the organization still thinks of the technology as a developer assistant.

Autonomy accumulates across the delivery lifecycle. An agent that drafts code, generates tests, responds to review comments, and prepares deployment changes may still appear to be operating under human supervision at each individual step. Taken end to end, however, much of the change has effectively been produced and advanced by the agent. Governance therefore has to consider the accumulated autonomy of the workflow, not just the apparent autonomy of each tool invocation.

High-impact insurance systems, the ones that drive underwriting, pricing, and claims decisions, can go as far as Level 3, gated approval, but no further. AI can do substantial work, but any change with material underwriting, pricing, claims, or regulatory impact should not cross the final approval boundary without human review. Autonomy is only half the picture, though. The same level of AI involvement means something very different for an internal reporting tool than for the system that applies policy limits. Governance has to weigh autonomy against the risk of the change itself.

Five anchor points for governance

Standardizing tools will not keep up, because teams will keep adopting new ones. What carriers can standardize is expectations: where governance anchors the delivery chain, regardless of which tools a team uses. Five anchor points matter most.

1. Anchor intent with accountability. Stories and acceptance criteria are the root of everything downstream, and if the source of truth is wrong, nothing after it can fix it. For high-impact rules such as rating, limits, underwriting, and claims payment, a named person or role with business authority should approve the requirements before any AI builds from them.

2. Keep verification independent of generation. Tests generated from the same AI-written stories as the code cannot catch a misread requirement. The strongest safeguard is a body of end-to-end tests that humans own. QA teams build them from the knowledge of underwriters, actuaries, and other business and technical experts, and from how existing systems actually behave, which is often the most reliable record of the real business rules. They do not share the code's AI-written origins, so they can catch errors that generated tests would pass.

3. Tier approval by system risk. Classify systems first, with rating and claims payment logic at the top and internal tooling at the bottom. Then let allowed autonomy and approval requirements follow the tier. If AI can modify the pipeline, it can weaken the very controls meant to catch problems, so pipeline changes need separate review. In practice, the most effective human gate is not reviewing every generated line of code, which quickly becomes a rubber stamp. Instead, it is requiring architects or subject matter experts to approve the artifacts with the widest blast radius, such as API specifications, contracts between systems, and CI/CD configuration, before anything merges to main or reaches a shared environment.

4. Preserve AI provenance with the change. Provenance means recording the origin of a change. Real work is often mixed: a human writes some parts, an agent writes others, and the human edits the result. Rather than a binary "AI-generated" tag, capture provenance metadata showing which tool or agent was involved and to what degree, through pull request labels, commit trailers such as "AI-Assisted: yes," and tags on stories, tests, and pipeline changes, each linked to the request that started it. This lets teams route AI-involved changes to underwriting, pricing, claims, or regulatory controls for the right level of review, demonstrate that certain tests remain human-owned, and show auditors not just what changed but how. Capture it at the point of creation: reconstructing it later is difficult, incomplete, and often unreliable.

5. Make the chain traceable, at build time and at run time. Any deployed change should trace back through test, code, and story to the person who approved the original intent. The running system must be traceable too: generated code should carry correlation IDs, structured logs, and business events that record which rule or limit was applied to which policy or claim, without capturing sensitive policyholder data. AI tools rarely add these unless required, so they belong in standards and templates that generated code must meet, with pipeline checks that fail when they are missing. When a regulator, auditor, or policyholder asks why a decision was made, the answer should be reconstructable from the record, not from memory.

Where to start

For carriers wondering where to begin, four steps make a practical 90-day start:

  • Inventory where AI is already used across product, development, QA, and DevOps.
  • Tier systems by risk and set the maximum autonomy allowed for each tier.
  • Name accountable business owners for high-impact requirements, and keep human-owned tests for the most critical rules.
  • Make provenance metadata, correlation IDs, and decision logging required standards for generated code, enforced in the pipeline.
The speed can come from AI

AI is not going away, and carriers should not be overwhelmed by it. It can dramatically speed up how we build software, but it can also be confidently wrong, and it does not understand our business the way our underwriters, actuaries, and engineers do. That domain knowledge is exactly what governance should anchor to. Teams will keep changing tools, so standardize the anchor points instead. When someone asks why a decision was made, "the AI did it" should never be an acceptable answer. AI can bring the speed. The accountability stays with people.


Soumya Udupi Ramesh

Profile picture for user SoumyaUdupiRamesh

Soumya Udupi Ramesh

Soumya Udupi Ramesh is a solutions architect with 20 years of experience in software engineering and enterprise architecture.

She holds a U.S. patent for an automatic application wizard. She is a senior member of IEEE and an active technical writer, peer reviewer, judge, and mentor.

Buy or Build? Yes.

AI has lowered barriers to building custom insurance workflows, but integration and adoption remain the hardest parts of implementation.

Insurance

One of my takeaways from ITC Vegas this fall was that a lot of vendors are playing defense. A lot of carriers are looking at what AI can do now and asking some version of the same question: Hold on a minute, the world changed, so should we just build everything ourselves?

It's a fair question. The barriers to building have come down a lot in the last year, and the benefits of owning your own technology are real. When a convincing demo takes a couple of engineers instead of a sales cycle, the old reasons to buy start to look pretty thin.

I suppose I'm not the most neutral person to ask. I've spent the last decade building and running Clearcover, an AI-native carrier, and now Dearborn Labs, an AI-native software development firm built specifically for insurance; if anyone was going to tell you to build everything, it would probably be me. But when people ask me whether to buy or build, my answer is...yes. We have components of our system that we bought and didn't build, and we're fine with that. We build more than we buy, though, and more every day.

Is "yes" a dodge? Maybe a little. Mostly it's that buy vs. build was never one decision. Every carrier already runs a mix of both, workflow by workflow. And even the build half of my answer comes with an asterisk.

Where should the line sit?

The choices sit along a spectrum, made one workflow at a time, and lately the mix has been shifting toward build. That makes sense to me, and I'd expect it to keep shifting for as long as building keeps getting cheaper.

Buy what's already solved well. If a problem looks roughly the same at every carrier, and someone has spent years getting good at it, there isn't much to gain from building your own version, and the time and money are better spent elsewhere. But that list gets shorter every month.

Build the workflows where being different is the point, like how you triage a new claim or decide which risks to write. Buying those means working the same way as everyone else who bought the same thing, which probably isn't what you were going for.

Where exactly that line falls depends a lot on the carrier and its lines of business, which is why a single company-wide decision on buy vs. build usually ends up wrong somewhere.

The asterisk

What qualifies as building, because you can put a demo together, is usually the smallest portion of the work. AI can change how a claim moves and which decisions get made without a person ever opening the file, but the systems underneath that workflow rarely change, or at least they rarely change quickly. Not everybody is doing a rip-and-replace of their whole company to launch a new AI workflow. A new claims-intake workflow still has to post to your core system and set reserves the way your actuaries expect, and an underwriting workbench still has to read from policy admin and write its decisions back. None of that shows up in a demo.

The interface is mostly a solved problem now, and that's good news. The hard part is the wiring underneath it. Most of the AI conversations I have with carriers wind up being about traditional software engineering. That's the work of fitting whatever the AI application or agent is into the rest of the company, so the systems supporting a reinvented workflow still work properly.

And the wiring is only half of the asterisk.

The other half is running the thing in production until people actually work that way. Most implementation projects are scoped to end at go-live, which is roughly where this half starts. At Clearcover, adoption of the AI tools across our claims team is 92%+. A number like that gets earned after go-live, inside real claims, over however long it takes for the new way of working to become just how the job gets done.

You might reasonably ask whether all of this is an argument for buying something that comes already integrated. For the work that's solved well, it is. But a bought workflow is the vendor's workflow, designed to fit as many carriers as possible, and in the workflows you'd want to run differently from everyone else, that's exactly the problem.

Build realistically.

Before you build a new workflow, ask three questions of it:

  1. Foundation — Was this workflow designed around constraints that no longer exist?
  2. Fit — Can it work with the systems you already have, without a rip and replace?
  3. Engineering — Do your engineers own the integration layer today, and can you test, govern and audit what runs on it?

The first is about not building on top of a bad foundation. A lot of insurance workflows were shaped by limits that made sense at the time, like a handoff that existed because two systems couldn't talk to each other, or a sampling audit that existed because nobody could read every file. If you automate that workflow as it stands, what you get is the old process running a little faster, and you've spent real money to keep a constraint you no longer have.

The second one is more encouraging. You don't have to fully rip out your systems to improve a workflow, so long as you have the software engineering chops to fit the new workflow into the systems you have.

The third is the one I'd push on hardest, mostly because a demo is very good at hiding it. If the answer is no, I'd want to know that before the demo goes in front of the Board rather than after.

If you're asking me whether to buy or build, I'd say build more than you used to. Just go in knowing that the demo is the easy part, and that the wiring, plus the long stretch of running it until people actually work that way, is the job. So where does your line sit this year, and who's doing the wiring?


Kyle Nakatsuji

Profile picture for user KyleNakatsuji

Kyle Nakatsuji

Kyle Nakatsuji is the founder and CEO of Clearcover, an AI-native auto insurance carrier, and Dearborn Labs, which helps P&C carriers and MGAs operationalize artificial intelligence. 

Before founding Clearcover, he was a venture investor at American Family Insurance, where he led insurtech investments. He speaks regularly on AI strategy in insurance.

Independent Agents - A Forward View

Economic simulations now forecast fiscal health across 1,900 U.S. localities, helping agents anticipate consumer affordability shifts and community financial stress.

Insurance

Local agents swim in the waters around them, and while a rising tide lifts all boats, that sucking sound before a tsunami hits shore is a clear signal that locality risk is a constant. It's good to be lucky with a rising tide, but luck is not a strategy.

Simulations of the world are a way to describe observed reality and anticipate change, and modeling the changing economic tides of every locality in America (call that a list with 1,900 locations) is a good way to both reflect on what has happened since COVID-19 and to help those agents see what's coming next.

Modeling Communities Like We Model Weather

We simulate weather daily to anticipate storm tracks and estimate damage. We simulate climate change with variation in sea level, flood plains, and riverbanks. We simulate fire spread, hail fall, straight-line wind… and so, too, we should simulate the local economies of 1,900 communities across America.

In a consumer-driven economy, simulating how America spends is like watching a developing hurricane pattern off the shore – is it intensifying or waning, is it heading for land or staying off the coast, what amount of intrusion to interior geographies will a footprint create, and will there be additional risks set in motion as the dynamics unfold, including supply chain and demand patterns that may also be interrupted or surged?

Affordability and Fiscal Health as Key Signals

Just as a student loan may go unpaid if post-graduate incomes are low, the choices households and individuals make on debts to service and debts to default are mirrored at businesses, organizations, and local communities, cities, and municipalities. These local priorities of affordability affect decisions on risk transfer and affect premiums – agents sit amid it all.

(See this article on a data source to predict financial health, going beyond the backward-looking credit reports that insurers use.)

Affordability is a key signal, as is the fiscal health of a locality. When you combine these with the inflows and outflows of households and businesses, any agency can make a plan for the year or two ahead and set in motion a process for continually navigating their locales.

Looking Backward: The Affordability Crisis

People living in houses and driving cars work at businesses and spend money in their communities. Unsurprisingly, changes in the economies of those communities are both observable and predictable. That matters to local agents.

Looking backward, hindsight is usually 20/20 – higher uninsured driver populations occurred where affordability was lowest. More UM/UIM claims drive insurance costs higher, and rates may go up even further. In an economic downswing, rising costs and flat to slow wage increases force more budget stress on consumers. Where job loss escalates stress, things twist up even faster. Not only do consumers decide not to be insured, but they also change other spending habits. More businesses fail when customers can no longer afford them. Collections of tax revenue suffer when high earners migrate away and when business activity slows or falters. Taxes suffer further when sales slump, and property values deteriorate. These can lead to cuts in community services, and spiral from there.

Emerging Patterns: From Data Centers to Remote Work

Independent agents, brokers of all sizes, carriers, wholesalers, specialty, excess, and MGA/MGUs all compete in these predictable localities. And predictions for consumers, businesses, and even the localities are proving their value in understanding and pricing risk.

Some localities see changes that are from emerging patterns of life and risk, like hyperscale data centers. Other changes are more subtle, like a work-from-anywhere job landscape that allows high earners to live wherever they want. Or what happens when AI shifts work tasks in legacy workflows or creates new ways of working?

Boomtowns to ghost towns and everything in between paint the map of America -- parcel-by-parcel, building-by-building, business-by-business, organization-by-organization, job-by-job, and so, too, insurance and risk management dynamics. Some things change very slowly, while others shift abruptly, but the profound effect of "work anywhere" from COVID may be the most significant overarching impact in our generation and the next.

Anticipating Economic Consequences

We anticipate sudden changes to life patterns in a variety of peril-driven scenarios – hurricane, tornado, fire, severe convective storm – as well as in environmental scenarios like sea level change, cyclical patterns like El Niño, and even major disruptions like power outages, transportation interruptions, and cyber interference.

Anticipating economic consequences, especially shifts in contribution to cash flow in community capital assets and liabilities, is another risk in the underlying infrastructure where all the other risks interact. Thriving localities add services and amenities, which attract more growth. Dwindling localities lay off public workers and renegotiate long-term liabilities while trying to improve local revenues with more taxes and fees. Debt offerings and the ability to service debt obligations are then most comparable to Ernest Hemingway's famous line on personal bankruptcy – which happens "gradually, then suddenly."

New Analytics for Local Planning

Analytics now exist for consumers, businesses, and their communities across the 1,900 localities that make up the investable and insurable economies of America.

This observability of a locality, its finances, its threats/opportunities, and its ability to service its debts and obligations is summed across the consumer and business mix of their constituents and forecast against those changing elements to portray current and future years' planning horizons. Local agents can take heed.

Situations with imminent insolvency are expected to make certain choices, while growth-busting upside creates better problems.

Adapting to Local Reality

While agents can always relocate, they are more likely to need to adapt to their local reality. While premium forecasts and business mix are constantly used in distribution planning and customer appetite focus, the pace of change has accelerated inside of the existing process steps. Using a locality stress index along with existing forecasts and affordability measures will make agents smarter in their marketplaces and quicker to react to change.

This means that in troubled areas, expect more business failures, lapses, and non-payment of premiums in a dwindling premium pattern, yet a rise in risk for fraud, theft, and malicious mischief, including a heavier threat of arson. This includes household stress, too, where skyrocketing base rates have pushed daily budgeting stress onto most households for car and homeowner insurance into an affordability squeeze where many drivers are underinsured, overcommitting to large deductibles, or going all the way to lapsing insurance entirely and becoming uninsured drivers.

What does luck look like?

A common quote from successful businesses – "It's better to be lucky, even if you're good."

Never has that been more relevant than today, as changes in local luck have magnified in recent years. Some places have a deep downside – rising taxes, failing businesses, public service degradation, municipal financial stress, drops in quality of life generally, and the ensuing changes in household composition from thriving to just holding on. Some places have all sunshine -- growth in incomes, formation of new jobs, infrastructure improvements, and engaged public services.

To live in an uninvestable locality is the hallmark of being unlucky.

The endpoints are always locality-based. It is where people live, where businesses operate, and where municipal authorities administer. For an agent, the forward look in these markets is flat to down. Fewer businesses for fewer booming households means fewer insurable assets. Even if premiums rise and commissions float with them, the total insurance value opportunity won't be in a boom cycle. The larger concern is a doom cycle.

Investors are aware of the general migration trend outlined in this recent Census article, whereby large metro areas are losing residents, while many smaller metro areas are growing. Details matter, and this is where data and analytics add critical value:

  • High earners are migrating away from large metro areas in fiscally troubled states.
  • They are migrating to small and mid-sized metro areas: Boise, Idaho; Frisco, Texas.
  • High earners are responsible for the majority of personal and property taxes, new business formation, and capital investments.
Scoring the 1,900 Localities

Across the observed 1,900 cities, the municipalities, consumers, and businesses are scored based on their future fiscal viability and likelihood to default across a set of debts (everyday bills, credit cards, car loans, mortgages, business loans, pension commitments, municipal bonds, etc.).

Ranked near the top of the list is St. George, Utah, with a short-term debt ratio of 0.31, which represents total liabilities divided by total assets. St. George is representative of 60 mostly mid-sized "under the radar" cities that have low legacy liabilities and sustainable budgets and are growing their populations with a healthy taxpayer base. This is where you can feel comfortable doing business for the long term. Lucky local agents!

Then there are plenty of typical "middling" cities, with budgets that are not sustainable over the long term due to underfunded pension and benefits liabilities. Steady as she goes for the next few years.

On the other side of the scale – consider a poster child for fiscal mismanagement to the extent it appears on a list titled Uninvestable Cities. These cities are laying off public sector workers today as they struggle with revenue shortfalls. Most of these cities meet the technical definition of insolvency. They appear to have accelerating financial deterioration over the next 24 months, resulting in a highly likely outright fiscal crisis.

Envision trying to do business in a city where high earners (lowest risk, rational decision-makers) are leaving. The businesses are shuttering, as well. New business formation is almost non-existent. The city struggles to fund basic services, while raising taxes. The residual population has lower FICO scores. Crime, including costly fraud, rises.

The Resilience of Independent Agents

Every year another study predicts the demise of independent agents. They are getting too old; they can't scale to digital; they can't compete with changing technology… add in AI.

Yet every year independent agents remain a dominant force in commercial insurance and have never failed to impress in personal lines. Watch how they adapt across 1,900 localities as a way to keep insuring America.

ADDITIONAL THOUGHTS PARKING LOT

What has changed is almost everything all at once.

Blame it on remote work options, but look at how many high-earner households opt to relocate for better taxes, better weather, better schools, better anything.

Blame it on migration, but the businesses and services they need will follow them, while the ones that used to serve them will feel the loss.

Blame it on anything you like, but the undeniable facts are that no one relocates to a state; they move to a specific location. The imbalances from losing localities to gaining localities are seen in shifts in income, housing, business formations/failures, and municipal financial resilience. To assess health, a current approach ties to the capital markets where municipalities trade their debt.

Using a locality lens, it is possible to observe the condition of consumers, businesses, and municipalities. With modern data and analytics, we can track changes over time. Did all three (consumer, business, municipality) improve, stay the same, or deteriorate in the last few years? What is the effect of these local changes, and how fast are changes now occurring? What is in store for the next couple of years?

The US is in the early stages of an economic transformation, by locality, and that means insurance agents have both local threats and opportunities.

Like Detroit filing Chapter 9 bankruptcy in 2013, it is conceivable that other major cities may follow suit. No city is immune to stress in changing local economies, and dozens of large municipalities have a reckoning coming. The tax base within these cities (high-earner consumers and meaningful business tax bases) is departing at rates never seen before.

How Document AI Becomes Usable at Scale

Enterprise AI must move beyond convenient document access and develop structured processing pipelines that extract, verify and validate content reliably.

Insurance

The latest wave of enterprise AI integrations makes perfect sense: People should be able to work with their documents in the tools they already use. Box's recent move to make its content available directly in ChatGPT is one more sign of how quickly that expectation is becoming normal.

But access to a document and reliable processing of that document are two very different things.

We've all been through it. Drop a contract, an invoice, or an audit report into a general-purpose assistant, ask a sensible question, and get back something that looks remarkably good. The system finds a few key terms, drafts a summary, perhaps even suggests what to do next. Then someone who knows the file reads it closely. A clause in an annex was missed. The total on the invoice does not match the purchase order. The answer treats a proposal as an approved decision. The work is half done, and its polished presentation makes the missing half harder to spot.

That is not an argument against general-purpose assistants. For exploration, drafting, and a quick feel for what is in a file, they are genuinely useful. The problem begins when a convenient interaction is mistaken for a dependable business process.

A document is rarely one question

Take a supplier contract. A user might ask, "What are our obligations?" A useful answer may require identifying the parties and the effective date, reading the pricing schedule, reconciling amendments, finding exceptions to termination rights, and distinguishing signed terms from an earlier draft. Some of those facts are on different pages or in separate documents. A single fluent answer can hide whether each of them was actually checked.

Or think about an invoice. Reading the amount is easy: determining whether it should be paid means comparing it with a purchase order, a delivery record, agreed tax treatment, and any prior invoices. The unit of work is not the PDF. It is the decision, with evidence from several objects.

This is why quality improves when we put a modest processing pipeline behind the document. First, identify what has arrived and preserve its source and version. Extract the content into useful parts: line items, clauses, dates, tables, signatures, attachments. Give the AI a specific job on each part, with an expected output, rather than asking it to improvise over the whole bundle. Then bring the results together, check relationships across documents, and apply the relevant business rules.

Finally, verify the result. Some checks are exact: amounts add up, required fields are present, a referenced purchase order exists. Other checks need AI: does the conclusion contradict a clause, omit an exception, or rely on a superseded version? The important decisions still need a person who can see the source, understand the proposed result, and correct it. A second AI pass can catch problems; it cannot certify its own answer as true.

The tedious part is the most valuable

This may sound less exciting than "chat with all your documents." It is also how document AI becomes usable at scale. Teams can repeat the same steps, measure where errors occur, and improve one stage without rewriting the whole process. When a reviewer corrects a field or rejects a conclusion, that correction belongs to a specific document, rule, and processing step. It becomes possible to explain not only what the system answered, but how it arrived there.

There is no universal recipe. A strategy document needs its assumptions and decisions separated; an audit response needs claims tied to evidence; an insurance case needs facts gathered across a changing file. The pipeline should reflect the job. Smaller pieces help the model focus, but they must be reassembled carefully so a local reading does not lose the wider context.

This is the shift I think enterprise buyers should ask for. Don't stop at "Can your AI see our content?" Ask: What exactly does it extract? Which version did it use? What is checked before a decision is proposed? What happens when two sources disagree? Who reviews an exception, and can that person get back to the passage that supports the answer?

The goal is not to make every interaction elaborate: The goal is to make the important ones repeatable, inspectable and good enough to trust.

Putting content within reach of an assistant is a welcome step. The next step is to treat what happens after the file is opened as a product problem in its own right. That is where the quality lives.

Automation Helps Claims Teams Handle Surges

Claims can surge after quiet periods, such as this hurricane season, and strain leaner teams that must review complex restoration invoices thoroughly.

Insurance

The 2026 Atlantic hurricane season has been remarkable for how quiet it has been. The first hurricane did not form until Oct. 8, the latest arrival in the satellite era. September passed without a hurricane for the first time since 1994. Thus far, the season has produced just one hurricane, compared with the annual average of seven.

For property claims teams, a quieter hurricane season can offer some breathing room. But it also raises a question: After a prolonged lull in claims activity, how will the industry handle the surge when a major event does occur?

For claims teams, that question extends well beyond the first wave of reported losses. Damage to hotels, apartment communities, retail centers and industrial properties can lead to large restoration projects and complex invoices that arrive over the weeks and months that follow. The capacity to respond to new claims is only part of the challenge. Teams also need the capacity to review the bills that come later.

The industry has been operating through a prolonged stretch of slower claims activity. Industry data puts commercial property claims at 710,000 in 2025, down from 910,000 in 2023. A separate measure, U.S. property claim assignments, fell 12% year over year in the second quarter of 2026 and ran 13% below the five-year average, extending a four-year decline. These are different Verisk datasets, but both point to a slower claims environment. More of the work that remains also arrives in bursts: catastrophe claims made up 43% of second-quarter volume, up from 34% five years earlier.

No claims organization can staff indefinitely for a catastrophe-level workload during a stretch like this. Nor should it. Teams that have adjusted staffing and resources to match lower demand face a particular challenge: claims volume can rebound much faster than review capacity can be rebuilt. When restoration invoices pile up, teams operating with fewer resources can face a difficult trade-off. Bills can wait for a thorough audit, extending claim cycle times, or move through with less scrutiny, increasing the risk of leakage. Automation offers a way to absorb more work without having to rebuild the entire manual review operation.

A surge can also have a long tail. After Hurricane Sally in 2020, Florida insurers had closed about two-thirds of residential claims six weeks later, but only 13% of commercial property claims. Hurricane Ian made landfall in Florida on Sept. 28, 2022. In the Florida Office of Insurance Regulation's March 4, 2025, snapshot, 96% of the 789,066 reported Ian claims were closed. Yet 6,628 of the 32,850 commercial property claims remained open nearly two and a half years after the storm. Those figures do not establish whether the delays involved coverage, reconstruction, invoice review or disputes. They do show how long commercial claims work can continue after the initial surge.

The bottleneck moves as the claim progresses

An initial wave of assignments is only the beginning. Adjusters must address coverage, communicate with insureds, inspect damage, review scopes and estimates, and coordinate recovery. Restoration invoices arrive as work progresses, often while those adjusters are still managing both new and existing claims.

Large restoration projects can involve multiple crews, weeks of drying equipment, substantial material charges and several subcontractors. The resulting invoices can be complex even under normal conditions. When many arrive together, claims teams face two pressures at once: more invoices to review and more work required to review each one thoroughly.

An invoice is not simply a total to approve. Time-and-materials bills can require checking labor hours and classifications, equipment usage and duration, agreed rates, materials, subcontractor charges, markups and arithmetic. A reviewer may need to compare the invoice with supporting records and the work performed. That takes time, and the work does not become less necessary because the team has fewer people available. Both delayed reviews and unsupported payments carry consequences for the insured, the carrier and the restoration contractor waiting to be paid.

Adding more reviewers can help, but it takes time to find, train and coordinate them. It also raises the cost of handling a temporary spike. The more useful question is which repeatable parts of review can be completed consistently without requiring a few more hours of specialist time for every additional invoice.

How automation can absorb a surge

Automation can help teams review restoration invoices faster by extracting and organizing line items, applying programmed audit rules and identifying charges that need a closer look. During a surge, these capabilities help maintain consistent, thorough review as volume increases.

AI can help turn invoice documents into organized data, while programmed audit rules check calculations and compare charges with agreed rates or benchmarks. The adjuster or specialist still has to consider the work performed, evaluate the findings and resolve disagreements. Automation reduces the manual checking that precedes those decisions. For example, NovaData’s Rest-Assured platform combines artificial intelligence, programmed audit rules and human expertise, and property insurers have used the platform to audit nearly $1 billion in property restoration invoices.

Measure the effect on the claim

The right measure is not merely how fast an audit comes back. Claims leaders should look at the time from invoice receipt to a documented review, how long exceptions take to resolve, the rate of confirmed findings, the cost per review and whether payment disagreements are cleared sooner. Those measures show whether faster processing is actually helping the claim move toward resolution.

What to put in place before the invoices arrive

The quiet period is the time to prepare. Once a major event occurs, mitigation invoices can begin arriving within weeks, followed by larger reconstruction bills over the months ahead. Putting the review process in place now gives teams time to establish the workflow before volume tests it. A few decisions made in advance can keep review from becoming the bottleneck later:

  • Confirm the rate agreements, pricing guidelines and benchmarks reviewers will apply, including how post-storm pricing will be evaluated.
  • Set one intake path for restoration invoices and define the supporting documentation each one needs.
  • Identify which repeatable review steps, such as line-item extraction, arithmetic and rate checks, can be automated.
  • Decide which findings can be resolved routinely and which require an adjuster or specialist.
  • Decide in advance where overflow review goes if volume exceeds internal capacity.

Claims teams cannot maintain catastrophe-level staffing through every quiet period. They need a review process that can expand when demand returns. Automating time-intensive invoice checks can help leaner teams absorb a surge without forcing a choice between longer cycle times and greater leakage. The next event will test more than the industry’s initial response. It will test whether teams can sustain a thorough review as the restoration invoices arrive. That is something claims leaders can prepare for now.

Source notes

  • Verisk ClaimSearch Trends Report, 2025 Year-end Analysis, p. 4. Commercial property claim counts are not directly comparable to the separate XactAnalysis assignment measure.
  • Verisk Quarterly Property Report, April–June 2026, pp. 3–4.
  • Florida Office of Insurance Regulation, Hurricane Ian catastrophe claims data, as of March 4, 2025. Commercial property claims reported: 32,850; open with payment: 5,332; open without payment: 1,296; closed: 80%. Aggregate insurer-reported figures are unaudited and claim statuses can change.
  • Isaias timing: CBS News, citing NOAA records, Oct. 9, 2026 (first hurricane of the season on Oct. 7, the latest date in 121 years).
  • Isaias insured loss estimates: Verisk, $1 billion to $2 billion onshore (Oct. 13, 2026); Moody's RMS, $2.5 billion to $4.5 billion private market (Oct. 16, 2026). AM Best and Vexcel figures as reported by Insurance Journal, Oct. 15, 2026.
  • Florida Office of Insurance Regulation, Hurricane Sally claims data, as of Oct. 26, 2020. Residential property: 57,373 reported, 65.6% closed. Commercial property: 3,649 reported, 13% closed. Figures are unaudited and cover Florida only.

Should We Trust AI With Medical Decisions?

As AI adoption in healthcare surges, organizations must prioritize targeted implementation and transparency to maintain patient trust in medical decision-making.

Insurance

Given its immense and growing capabilities, more and more healthcare organizations are, understandably, looking to implement AI. Adoption is up, with 75% of US health systems using the technology in 2026 – a steep increase from 59% last year. But as AI moves into most corners of American healthcare, the industry must take a step back. AI use for the sake of AI use is not the goal. Rollout of the technology must be targeted, especially as trust in AI is called into question.

This is true across the entire sector, and I've seen it up close across the field I operate in: contested medical review and health insurance disputes. Here, it's vital to insist that AI only be used where it meaningfully addresses a problem and, crucially, that it never makes an important medical decision.

Where AI can help: The No Surprises Act case study

Take the No Surprises Act (NSA) Independent Dispute Resolution (IDR), for example. When the No Surprises Act took effect in 2022, it set off a crescendo of disputes between insurers and providers, which reached 394,140 monthly cases as of July this year. With hundreds or even thousands of pages of documents per case, independent dispute resolution entities (IDREs) have ramped up efforts to handle the sheer volume of evidence. But such enormous numbers remove the adjudication process from the realm of human capacity alone.

That's an area where AI can play an important role, keeping costs and capacity crunches under control. By trawling through documents, pulling out necessary information, and comparing and contrasting evidence from payers and providers, AI can strip away a lot of the administrative burden that takes up so many hours, allowing arbiters and experts to make informed decisions on more cases.

Where AI doesn't belong: Final medical decisions

Equally, though, it also demonstrates a perfect example of where AI simply doesn't belong: it should never make the actual judgment in any insurance dispute. That decision must always remain with a human expert. Just because AI can make the call, and might do it quicker, doesn't mean it should. Taking that step would undermine something even more important: trust.

Rising scrutiny and accountability

Maintaining and building that trust is even more crucial at this moment, because scrutiny of AI's use in IDR is undoubtedly ramping up. In September, the office of New Jersey Rep. Frank Pallone sent letters to six IDREs, asking a series of questions about their operations. One of those questions zoomed in on AI in the IDR process.

This scrutiny is reasonable. These processes carry real implications for insurers, doctors, and hospitals. An IDR decision is a real healthcare decision that determines what insurers pay for procedures, and how much providers receive. In my view, that is enough reason to keep AI out of the direct medical decision-making altogether.

Pallone's letters hint at a vital question many have about all aspects of health AI – one that extends far beyond the healthcare insurance space: is it making decisions about my healthcare without my knowledge?

Growing skepticism across healthcare

Because the skepticism is rising across the board. Federal initiatives such as Medicare's AI-led WISeR program and new patient chatbot initiatives, for example, have come under fire for making key decisions and even prescribing drugs to patients. Irrespective of whether these uses of AI are appropriate, there's no denying that narratives around them are shifting. They are proof that building trust and being targeted with the technology's use at this moment in time is vital.

What patients want: Transparency and control

There is clearly still appetite for AI. Patients themselves are increasingly turning to chatbots for medical advice alongside that of their doctor. What matters is the execution. You can see that in polling by the Pew Research Center, which suggests that an overwhelming 72% of Americans believe it is very or extremely important to know when AI is used in their healthcare – including in administrative decisions, or even note-taking. It's clear that patients see a difference between AI nebulously injected into every aspect of their regular care, and targeted, careful use that adds another component, or extra layer of support, to it.

Building trust through transparency

The answer for all organizations, then – from healthcare providers and healthtech firms to governments – is to make trust a top priority. That means hard-and-fast rules about transparency and accountability. For insurers, it means clearing up what prompts are used to guide generative AI, and, in the context of IDR specifically, where they are relying on human experts to make key judgments about payment offers. In some cases, third-party AI audits might be necessary, but whatever the measures, they must be openly communicated to the public.

Crucially, every organization must be targeted with the technology's use. In every application, the use case needs to be clear from the start. Does AI complete the work better than a doctor, a claims specialist, or an underwriter, or just faster? Speed and volume – AI's strengths – should not necessarily be the priority across all areas of healthcare, especially when they come at the cost of accuracy, reliability, accountability and transparency.

A critical juncture for healthcare AI

This feels like a critical juncture for public trust. Moving forward, we must implement the technology only where it meaningfully solves a problem, and never, ever, to make an important medical decision on its own. Indiscriminate use risks permanently damaging trust and, if we let that happen, we'll lose so many of the meaningful improvements AI has the potential to bring to patients across the world.

When an AI Calls, Who Should Answer?

Blocking Meta's AI agent on your website won't stop it from calling. Insurers need a plan to handle these calls.

Insurance

A consumer's AI agent can now call an insurer and ask for a quote. The possibility of such calls leaves carriers with two bad options: putting a licensed person on the line with a machine, or risk hanging up on a real person.

Both options are wrong. There's a third.

The AI phone calls are already happening. Reuters reported that people using Meta's Muse, the personal AI agent launched in the U.S. this past September, can have it call businesses to book appointments, check inventory or get quotes from contractors. A Meta employee testing this feature said an insurer kept hanging up on Muse after realizing it was an AI caller.

At InsureTech Connect in Las Vegas, Steve Rullo, who is Guardian Life's chief digital and technology officer, told Insurance Business, "We are already seeing providers calling in for verification of benefits with AI." He described healthcare providers using AI to verify patient benefits, often for an entire list of patients in a single call.

The Debate Is About Websites

Since Muse launched, carriers and marketplaces have been deciding whether to allow it access. Insurify quickly blocked Muse, saying it "risks stripping carrier quotes of critical contextual information." Co-CEO Giorgos Zacharia explained, "A quote without its context is not a fair comparison. It is a number."

Shortly afterward, Kin announced that it would let Muse crawl its site so customers can shop for quotes, and Zebra announced an integration the following day. Employers, a workers' compensation insurer, told Dig-In it doesn't allow third-party AI agents like Muse to access its operational systems or customer data.

Each of those decisions is about access to a website or system. None of them covers phone calls, and blocking Muse online won't stop it from calling. I expect insurers will get a lot more of these calls. Handled the old way, they're a cost. Handled well, every one is a quote in front of a real shopper. That's where the real opportunities are.

The Talking Is About 5% of the Value

The answer is simple: AI for AI, humans for humans. Let AI answer calls from other AI, and let people answer calls from people. This way, a family dealing with a flooded basement won't be stuck on hold while a licensed agent talks to a bot.

Most voice AI answers the phone, then sends the caller through a phone menu or to a person. That's 5% of the value. The other 95% is the work behind the call: pulling up the policy, filing the claim, writing the result back to the core system. If AI can handle a machine caller's request at this level, no person needs to pick up, and licensed people can focus on judgment calls, like denied claims.

Most carriers aren't there yet. According to KPMG International's report Unlocking AI value in insurance, only 29% of respondents use AI agents or automation for end-to-end processes. Rullo said Guardian will keep people involved in its AI processes: "We'll still have a human in the loop." For some risk profiles, he said, Guardian wants alerts on exceptions instead of reviews of every transaction.

People should own the decisions. They shouldn't have to answer every call.

You Can't Always Tell Who Is Calling

Carriers also can't always tell if a caller is a machine or a person. Reuters reported that Meta tested having human contractors make some Muse calls without telling the other party. A Meta vice president said the test without proper disclosures "was a miss," and Reuters said Meta has rolled back that test for now. So a call that seems to come from an AI agent might actually be from a person. Zacharia told Dig-In that Muse got around Insurify's bot detection and didn't identify itself.

A system designed to resolve requests can help both human and AI callers. A system that only routes calls can't.

Policyholders Already Draw the Line

In Liberate's September survey of 1,001 U.S. claimants, 50% said they would prefer a claim paid in 60 seconds with no human contact over talking to a person within 10 minutes and getting paid in three days. Another 41% preferred a person on the phone. Of those who chose the fast payout, 55% still want a human involved in every denial. 

Fast on yes. Human on no. 

Across the whole survey, 61% say a human must review every denial. In the same survey, 86% want to be told when AI touches their claim.

A Hi Marley survey of 550 senior property and casualty decision-makers and 2,000 policyholders in the U.S. and Canada highlights the gap carriers need to close. Eighty-five percent of executives think AI will build customer trust, but only 40% of policyholders agree. Ninety percent of policyholders said it's very or extremely important to know when they're interacting with AI, and two-thirds said their biggest worry is not being able to reach a human when it matters.

Three Rules for Carriers

First, use insurance-specific AI that works within your core systems to answer calls from machines. Sending a bot to a phone menu or a person wastes time and doesn't solve the problem.

Second, keep a licensed person on every denial, no matter what else the AI handles.

Third, be transparent about AI on both sides. Let people know when they're talking to AI, and expect AI callers to do the same.

If carriers follow these rules, they can answer every call, whether it's from a person or an AI. The shopper using a machine gets a response, and the family with a flooded basement gets to talk to a real person. Insurance has always wanted to answer everyone who calls. Now it can.


Amrish Singh

Profile picture for user AmrishSingh

Amrish Singh

Amrish Singh is co-founder and CEO of Liberate, which builds insurance-native AI agents for P&C carriers and agencies. 

He previously spent nearly four years in back-office operations and technology at Metromile.

Insurers Await AI Integration in Core Platforms

Insurers aren't avoiding AI due to fear—they're waiting for core platform vendors to deliver native capabilities instead of risky integrations.

Insurance

We've all heard some version of the same story this year: Many carriers and brokers are still on the fence about AI because of regulatory risk, errors and omissions (E&O) exposure, or "black boxes" they can't explain to a state department of insurance (DOI) examiner.

All of this is true. But none of it tells the whole story.

Spend time with almost any carrier's IT and operations leaders, and eventually they'll say the quiet part out loud: They're not avoiding AI; they're waiting for it to arrive inside the systems they already run. Those systems are the core platforms, such as Duck Creek, Guidewire, Majesco, Sapiens, and Insurity, that sit beneath a large share of the policies and claims on the books today.

The Real Bottleneck Isn't Fear. It's Architecture.

For most carriers, policy, billing, and claims data live in a core system: a multi-year, multimillion-dollar investment that the business treats as its "single source of truth." The vendor varies from carrier to carrier, but the obstacle their IT teams face is the same.

Bolting a standalone AI tool onto that architecture, even an elegant and streamlined one, isn't a plug-and-play decision. It's a multi-team integration project that typically stretches across several quarters, exposing the organization to delays, cost overruns, and shifting priorities. One claims technology vendor estimates that AI vendors lose roughly 40% of their projected timelines to integration complexity, whether the carrier runs Guidewire, Duck Creek, Majesco, or a custom legacy core.

That kind of complexity is often the difference between a pilot that launches and one that quietly dies in the backlog.

Naturally, carriers also scan the standalone AI market. They see dozens of point solutions for underwriting triage, claims scoring, and first notice of loss (FNOL) intake, and wonder: "Why integrate a third-party layer today when the core vendor is about to deliver the same capabilities natively?"

It isn't a hypothetical question. Nearly every major core platform vendor is moving to close this gap. Consider a few examples:

  • Duck Creek launched its Agentic AI Platform in April 2026.
  • That same month, Guidewire embedded ProNavigator, an AI assistant that gives adjusters context-specific guidance, into ClaimCenter and its other core applications. In August, Guidewire followed with its Qusar release, which added Guidewire-built claims and underwriting agents, including an agentic FNOL tool.
  • Majesco announced that it's quadrupling its AI investment in 2026 to build an AI-native P&C platform and expand Majesco Copilot and its agentic AI capabilities.
  • Sapiens has been rolling AI-based functionality into its CoreSuite releases.
  • Insurity has committed an additional $100 million over two years to embed AI in its core systems, on top of the more than $50 million it had already invested in AI and R&D.

In each case, the pitch is essentially the same: The AI isn't a separate purchase; it's part of the system carriers already pay for.

That point is missing from many of the "why insurers are hesitant" stories in the media. Carriers aren't approaching AI cautiously simply because they're risk-averse. They're using their position as large, long-term platform customers to create a buyer's advantage. That leverage lets them be deliberate about their "build-versus-buy-versus-wait" calculations—and for many carriers, waiting has so far proved the more disciplined option.

Why Waiting Is a Deceptively Difficult Decision

The problem is that waiting has a cost, too, and that cost doesn't fall evenly. One AI underwriting vendor that sells a core-agnostic product has argued that carriers and managing general agents (MGAs) can't afford to wait for a full core system transformation before deploying AI and can't afford to be locked into a tool that works with only one platform.

Both concerns are valid, which makes waiting a strategic bet rather than an obvious answer.

There's also a quieter signal that undercuts the "carriers are simply being cautious" narrative. A May 2026 IA Capital Group survey of national and regional carriers found OpenAI's models in 91% of respondents' AI stacks, often delivered through Microsoft Copilot. Carriers aren't AI-averse. They're AI-selective, using their scale to adopt productivity tools quickly while holding the line on systems of record until their core vendors catch up. For many of them, that moment has either arrived or is very close.

The Exception That Proves the Rule

So why do carriers buy standalone tools at all? Because when a tool solves a problem that general-purpose AI, or the AI now arriving in core platforms, doesn't handle well, the calculus flips.

That's where most insurance-specific AI vendors stake their pitch: not "we're AI," but "we're AI trained on insurance data and workflows, which a horizontal model wrapped around your files isn't." It's an argument that still holds water. Shift Technology has made this case to carrier IT leaders directly, arguing that generic horizontal AI players often lack the data foundations and expertise needed to move a pilot into production. It's a real distinction, not just positioning. A general-purpose model can summarize a claims file. Reasoning reliably across coverage language, state-specific regulations, and loss histories is a different problem, and it's one that a system trained specifically on that data is, at least in theory, built to solve.

What sustains the "insurance-specific always beats generic" argument, then? The products carriers keep buying from these vendors, despite their preference for waiting until such capabilities arrive inside a core system, share one trait: They're widely believed to do the job better because of deep domain training, proprietary data, or a workflow too specialized for a core vendor to prioritize anytime soon.

Everything else, such as a general chat interface bolted onto a claims queue, is the kind of functionality most likely to show up natively in the next core release, whichever vendor delivers it. That's why carriers aren't rushing to pay for it twice.

What This Means If You're Evaluating Tools Right Now

A few things are worth doing before you sign anything:

  • Separating "edge" AI from "core" AI. Marketing copy and service-desk triage carry low regulatory exposure and can run successfully on general-purpose tools today. Underwriting and claims determinations that touch policyholder outcomes are a different risk category, where the wait-for-native instinct makes more sense.
  • Asking vendors for proof of successful integrations, not just a demo. A clean demo says nothing about how a tool behaves once it's wired into a decade-old core system. Ask for a named carrier reference on your exact platform and version.
  • Pushing for a real audit trail. If a vendor can't produce output in a format that your internal compliance team and state DOI examiners would accept, that's a governance gap that will surface at the worst possible time.
  • Ensuring pilot programs don't require a multi-year lock-in. Any AI vendor worth considering should be comfortable running a three- to six-month pilot on a limited book of business before asking for a long-term commitment.
  • Watching your core vendor's product roadmap as closely as the pitch deck. If your platform vendor plans to deliver the capability you're evaluating in its next release cycle, that changes the decision, even if the standalone tool works well today.
  • Testing whether the value is domain-specific. If a vendor's edge could be replicated by connecting a general-purpose model (for example, an enterprise assistant such as Microsoft Copilot, or an AI workflow that retrieves and summarizes documents) to your own policy forms, endorsements, claim files, and underwriting guidelines, don't count on that edge lasting once your core system catches up. If it depends on proprietary data or workflow depth that a core vendor is unlikely to build soon, that's the case worth paying for.
  • Knowing the legacy exception. If a legacy system will continue to house and service existing policies for years before any platform migration, a general-purpose AI layer built over that system's documents and data, such as a search-and-summarize workflow for policy forms and claim notes, can still earn its keep as a bridge, even without a lasting edge.
The Differentiator

Most coverage of AI hesitancy in insurance treats reluctance as a problem of mindset: Carriers are cautious, regulators are strict, liability is scary. That may be true, but this framing overshadows a more relevant (and more interesting) story.

Carriers that are being deliberate aren't necessarily behind the curve or tech-averse. Many are making a sequencing bet: Let the core platform vendor absorb the integration risk, compliance burden, and audit-trail engineering, and then adopt AI as native functionality instead of a fragile add-on.

That strategy holds up well, with one condition: Waiting has to remain an active decision rather than a default. A carrier that sequences AI adoption around its core platform should still revisit that choice whenever a standalone tool offers an edge its core vendor can't match soon, using the tests outlined above. Handled that way, the real story in 2027 won't be whether insurance is "ready" for AI. It will be which carriers turned strategic patience into a well-timed rollout.

Sources:

- Decerto, AI Claims Processing: The Complete 2026 Guide for US Carriers, April 2026: integration complexity and the estimate that AI vendors lose roughly 40% of projected timelines to it.

- PRNewswire / Convr, Convr Delivers AI Underwriting Across Guidewire, Duck Creek, Sapiens, and Beyond, June 2026: the case against both waiting for core transformation and single-platform lock-in, from a core-agnostic AI vendor.

- Duck Creek, Duck Creek Launches Insurance-Native Agentic AI Platform and Unveils New Applications to Transform Underwriting and Claims, April 28, 2026: launch of the Duck Creek Agentic AI Platform.

- Guidewire, Guidewire Launches ProNavigator, Embedding Expert AI Insights into Insurance Workflows, April 16, 2026: ProNavigator embedded in InsuranceSuite (including ClaimCenter) and InsuranceNow.

- Guidewire, Guidewire Introduces Qusar Release to Help Insurers Build and Control AI Agents, Aug. 3, 2026: Qusar's Agentic Framework and Guidewire-built claims and underwriting agents.

- Zfort, Guidewire ClaimCenter vs Duck Creek Claims, August 2026: comparison of Duck Creek's Agentic AI Platform and Guidewire's Qusar release.

- Hesper AI, Claims management systems compared, May 2026: Guidewire's ProNavigator embedded directly into ClaimCenter.

- Business Wire, Majesco Accelerates P&C Market Leadership with 2026 AI Investment Surge, December 2025: Majesco quadrupling its AI investment and deepening Majesco Copilot and agentic AI capabilities.

- SEC filing, Sapiens Launches Enhanced CoreSuite for P&C, September 2025: Sapiens rolling AI-based functionality into its core platform releases.

- Business Wire, Insurity Named to the AIFinTech100 List of Leading AI Innovators in Financial Services, July 2026: Insurity's additional $100 million commitment, on top of more than $50 million already invested in AI and R&D.

- The Insurer, OpenAI dominates AI stacks as insurance industry moves from pilot to production, May 2026: IA Capital Group survey data on foundation-model presence in carrier AI stacks.

- Shift Technology, Making Agentic AI Work for Insurance webinar page, 2026: the case against generic horizontal AI models for production-grade insurance agents.


James Ballot

Profile picture for user JamesBallot

James Ballot

James P. Ballot is an insurance research, thought leadership, and content strategy leader with more than a decade of experience helping industry, regulatory, business, consumer, and higher education audiences understand and navigate complex industry transitions.