Somewhere right now, two competing carriers are running the same language models, from the same vendors, on the same cloud, tuned by the same consultants. Both executive teams describe this as a competitive advantage.

It cannot be true for both of them. It is probably not true for either.

Insurance adopted generative AI faster than any technology in its history. Two-thirds of North American carriers now run it in production, per Celent's latest insurance industry survey, up from fewer than one in 10 three years ago. Yet in that same survey, only 3% say they are getting the value they expected. I argued in these pages last year that most insurance AI strategies would fail. I would sharpen that now: most of them are succeeding at the wrong thing.

The Uncomfortable Part

Here is the uncomfortable part. Advantage never lives in technology everyone can buy. When every carrier has the same models summarizing the same documents and drafting the same emails, the gains are real, and they cancel out. We have run this experiment before. A website was a strategy in 1999. A mobile app was a strategy in 2012. Each became plumbing within five years, and the winners were decided by what they built on top of it. AI inside the building is on the same clock, and this clock runs faster.

There is a simple test for whether an AI program is compounding or just converging: ask what your customer has noticed. For most carriers, the honest answer is nothing. Nearly every production deployment in the industry is employee-facing. In the same survey, not one insurer runs AI strictly for the customer. The claims process got faster on the inside. The customer still meets the same forms, the same hold music, the same "please allow 7 to 10 business days." We have taken the most human-facing technology ever built and pointed it at everything except the humans.

Where the Real Battle Lies

I argued last year that the real competitive battle is in the experience layer, the place where a client and their financial professional actually meet your company. A year of building on that idea has hardened it into something stronger. The transaction was never the product. Any carrier can process a withdrawal. What surrounds that withdrawal, the tax it triggers, the guarantee it quietly reduces, the conversation it should have started, is where a customer decides what your company is. AI is the first technology that can deliver that surrounding intelligence at scale, to every customer, on every transaction. Which is exactly why the layer it lives in is the one worth owning.

How a carrier builds and owns that layer is the strategy we are executing at Security Benefit, so I will keep the blueprint to myself. But the reasoning behind it fits in one question, and the question is free.

Sometime next year, one of your customers will ask a question about their money. Some intelligence will answer it. Maybe yours. Maybe your distributor's. Maybe a general-purpose model that has never heard of you. The carriers that matter in five years will be the ones that decided, on purpose, whose intelligence answers. That decision is still open. It will not be open long.

The grassroots approach to AI adoption has probably gone as far as it’s going to go.

Point solutions are impressive—but rarely scale. In-house pilots generate strong early results, only for performance to deteriorate for lack of consistent data.

As for copilots, people generally appreciate tools that draft emails, summarize documents, or eliminate busywork. But enthusiasm tends to wane when the tool stops assisting and starts deciding. Handing approval authority to a model can feel indistinguishable from handing over your badge and desk.

That's not irrational. Nobody organizes a grassroots campaign to accelerate their own obsolescence.

The next big push in enterprise AI, when it comes—and I suspect it will come sooner rather than later—will come from above: top down, through the capital structure.

Trillions of dollars are being invested in the infrastructure required to power the AI economy. The institutions providing that capital aren't investing on faith. They have underwritten a thesis: that vast investments in computing, data centers, power generation, networking and models will produce vast economic returns.

For that thesis to work, somebody has to create the demand.

Consider who is financing the AI buildout. BlackRock, Blackstone, Apollo, KKR, Goldman Sachs and other institutions that manage or deploy trillions of dollars are helping finance hundreds of billions of dollars in AI infrastructure. Many of those same institutions are among the largest shareholders of America's public companies, including insurers.

At Travelers, for example, Vanguard, BlackRock, State Street and Fidelity together account for roughly one-third of the company's shares. The financial system is simultaneously financing the supply of AI infrastructure and owning the enterprises whose adoption of that infrastructure will be necessary to justify the investment.

There's an almost poetic financial loop here worth noting: Blackstone's AI infrastructure investments are partly financed with insurance capital. Blackstone now manages $290 billion in insurance capital, while its infrastructure and data-center strategy explicitly connects digital infrastructure with its insurance-capital platform.

All of this means the capital funding AI infrastructure and the capital owning the companies expected to use it increasingly overlap.

Put more plainly: the same capital on your shareholder register may also be financing the infrastructure that needs you to become an AI customer.

The pressure will run downhill. Capital markets will lean on boards. Boards will lean on CEOs. CEOs will hang targets on business units. And eventually the questions will get very specific: What percentage of your workforce is AI-enabled? What's the productivity delta? Where's the ROI? What are your competitors doing with AI that you aren't? What is stopping you from moving faster?

Delivering solid business results is, today, enough. In the near future, how those results are delivered—through what combination of human effort, automation, and AI—will increasingly factor into executive evaluation.

Insurance leaders can prepare now. That does not mean recklessly deploying AI everywhere. It means creating the conditions under which AI can produce durable, defensible business results.

Governance is part of that infrastructure. Done badly, governance becomes an institutional “no.” Done well, it becomes a mechanism for saying “maybe” or “yes,” depending on the use case, the risk, and the evidence.

There is also substantial work between an AI pilot and a legitimate business result. Enterprise information must become usable by machines that can reason and act on it. That means building ontologies, pipelines, object models, knowledge graphs and relational data stores; connecting systems never designed to work together; and establishing the controls that make machine decisions auditable.

None of that is especially glamorous. All of it takes time.

Which is why waiting for questions from investors, analysts, and boards may be the wrong strategy. By the time those questions arrive, you do not want to be explaining why you haven't started. You want to be showing what you're building, what it does, and what it's worth.

The AI infrastructure industry needs demand. Build your own AI infrastructure to create demand for theirs. Build before they ask.

There is an uncomfortable question confronting the property and casualty insurance industry:

Are we experiencing a temporary backlash, or is insurance entering a new era of fundamental distrust?

Insurance has never been an easy product to love. Consumers pay premiums for something they hope they never need, governed by contracts that can be difficult to understand and fully appreciated only when a loss occurs. The relationship can change dramatically in a single claim. But something feels different today.

Distrust is broader, louder and increasingly connected to public frustration over rising costs, corporate power, artificial intelligence, data collection, climate risk and the perceived behavior of large institutions.

For an industry whose fundamental product is a promise—we will be there when something goes wrong—trust is paramount.

A warning we wrote about before

In December 2024, Stephen Applebaum and I wrote “Broken Trust, Insurance Industry Included.” Our premise was straightforward: the public reaction following the killing of UnitedHealthcare CEO Brian Thompson was an alarm bell for the entire insurance industry, not simply the health insurance sector. We pointed to rising premiums, coverage withdrawals, privacy concerns, widening protection gaps and growing skepticism toward technology as evidence of a broader erosion of confidence.

Nearly two years later, that warning appears increasingly relevant.

The issue is no longer simply whether consumers trust their insurance company. It is whether the public increasingly distrusts the insurance system itself.

The symbol of a much larger problem

The December 2024 killing of Brian Thompson was horrifying. Yet the public reaction to the alleged perpetrator, Luigi Mangione, revealed something the insurance industry should not dismiss as simply an isolated social phenomenon.

Polling following the killing showed unusually strong sympathy for Mangione among younger Americans. The reaction was not really about one individual. It reflected anger toward a system that many people believe has become too powerful, too complicated and too disconnected from the people it serves.

UnitedHealthcare operates primarily in health insurance, not P&C. But to the public, “insurance company” can be a much more important category than the distinctions between health, auto, homeowners or commercial insurance.

That should concern every P&C executive.

Then came the claims controversies

The industry’s claim function has increasingly become the center of the conversation.

In 2025, a U.S. Senate hearing examined insurance claims practices following natural disasters, with executives from Allstate and State Farm appearing alongside policyholders, adjusters and consumer advocates. Allegations that claim evaluations had been manipulated or unfairly reduced received significant attention, although the insurers disputed them.

In June 2026, Oklahoma Attorney General Gentner Drummond filed another lawsuit against State Farm alleging that the insurer’s “Hail Focus Initiative” used undisclosed standards and other practices to reduce payments on hail and wind claims. State Farm disputes the allegations.

Just this week, Los Angeles County announced a lawsuit against State Farm over its handling of claims following the January 2025 Southern California wildfires, alleging delays, underestimation of losses and other improper claims practices. State Farm has disputed the allegations and pointed to billions of dollars in wildfire claims it has paid.

These cases have not established that the insurers acted improperly. Lawsuits contain allegations, not findings of fact. But public trust is rarely determined by the final disposition of a lawsuit – many of which are dismissed, in favor of the insurer or most often amicably settled. Either way, the narrative gets there first.

Profitability makes the narrative more difficult

The industry’s financial performance adds another complication.

P&C insurers have experienced a substantial improvement in profitability as premium growth, stronger underwriting results and investment income have combined to produce record-setting earnings.

Strong insurer profitability is not inherently evidence of consumer mistreatment, yet broad-brushing narratives paint a negative picture to make a point.

Insurance is a capital-intensive business. Insurers need adequate returns to support capital, absorb catastrophe losses, pay claims, invest in technology and remain capable of writing business through difficult cycles.

But consumers don’t necessarily see that complexity. Instead, the narrative can become:

My premium went up. My deductible went up. My claim was denied. In all cases, the insurer made more money.

Whether that conclusion is actuarially accurate is almost beside the point. It is emotionally powerful. And when consumers are already frustrated by inflation and the cost of housing, automobiles and repairs, insurance profitability becomes an easy target.

The “closed without payment” problem

A recent Wall Street Journal analysis of auto insurance claims illustrates the challenge.

The Journal reported that the percentage of certain auto liability and medical claims closed without payment had increased materially over the past decade. Analysis of NAIC data showed that 45% of such claims were closed without payment in 2025, compared with 35% in 2016. That is an important statistic.

But it is also an example of how a technically accurate statistic can create a broader impression that may not tell the whole story.

“Closed without payment” (CWP) does not equate to “wrongfully denied.” Claims can close without payment for numerous reasons, including coverage issues, duplicate claims, fraud, liability determinations or other factors. In fact, CWP rates are a singular gauge for purposes such as monitoring claim productivity and should not be isolated from average claim payments, reserve accuracy, re-open rates and other metrics to assess payments. Certainly not to accurately judge claim settlement fairness.

The industry’s response made an important distinction: the increase identified by the Journal was concentrated in liability and medical claims, while physical-damage claims were being paid at essentially the same rate as a decade ago.

A technically correct defense can still fail as a trust strategy.

If consumers hear “nearly half of claims aren’t being paid,” a subsequent explanation about claim categories, coverage triggers and statistical methodology may never overcome the initial impression.

The industry needs to communicate in a way that makes the underlying economics and claims experience understandable—not simply defensible.

Insurance has always had a trust problem

None of this is entirely new.

Insurance has several structural characteristics that make trust difficult. It is intangible, complex and often mandatory. The customer pays first and receives value later—sometimes years later. And when the customer most needs the product, the insurer must determine whether and how much it will pay, which can easily create friction.

A policyholder sees a damaged roof. An insurer sees a contract, causation, exclusions, depreciation, replacement cost, engineering evidence, estimating methodology, fraud indicators and applicable regulation. Both may believe they are acting reasonably. But one side has generally spent decades learning how the system works. That asymmetry creates distrust.

Four forces amplifying the problem

Affordability: Consumers have experienced dramatic increases in home and auto premiums in many markets. Those increases have legitimate drivers—repair costs, medical costs, litigation, catastrophe losses, reinsurance, inflation and changing risk. But consumers experience one part of the equation: the bill.

Claims: The claim is the industry’s moment of truth. Every difficult claim creates a potential advocate—or detractor. Social media can now turn an individual dispute into a national story almost instantly.

AI and data: AI can improve claims accuracy, detect fraud, accelerate settlement and reduce administrative expense. But from a consumer’s perspective, AI can also sound like: a computer decided not to pay me. The broader public debate over AI, privacy, data centers and technology companies suggests this skepticism will grow.

Distrust of institutions: Insurance is not operating in isolation. Banks, pharmaceutical companies, technology companies, healthcare organizations and other large institutions are experiencing versions of the same credibility challenge. Insurance is particularly vulnerable because its product is fundamentally built on trust.

The industry’s response

There are reasons for optimism. Some insurers recognize that restoring trust requires more than advertising.

Farmers has introduced a Coverage Review initiative designed to help consumers better understand what their policies do and don’t cover. Importantly, the service is available even to consumers insured elsewhere. This attacks the trust problem upstream by not waiting for a claim to explain the policy. Time will tell if this is simply clever marketing or possibly a new way to encourage discussions that consumers tend to avoid.

State Farm has also taken steps to return value to customers, including a $5 billion cash-back dividend for qualifying auto customers, while reducing auto rates in several markets. Progressive has returned excess profits to eligible Florida personal auto policyholders, while USAA has also reduced Florida auto rates and returned value to its members.

Rate increases are easing in several markets after several years of extraordinary increases. These actions matter. But trust is not restored by one dividend, one advertising campaign or one rate decrease. It is restored through repeated evidence that the organization behaves consistently with the promise it makes.

The ecosystem has a role to play

This is not solely a carrier problem.

Solution providers, claims technology companies, TPAs, adjusters, brokers, agents, consultants and analysts all influence the customer’s perception of insurance. Every automated decision and claims estimate. Every AI recommendation and vendor interaction. And each confusing communication. They all become part of the insurance brand.

That means the industry’s technology agenda cannot simply be about doing things faster and cheaper.

It also has to be about doing things in a way customers perceive as fair, understandable and trustworthy.

Instead of asking only, Can AI make this decision?

We should ask, Can the customer understand why this decision was made?

Instead of asking, Can we automate this claims process?

We should ask, Where does a human being add trust and judgment?

And instead of asking only, Can we reduce claims expense?

We should ask, Can we reduce expense without damaging the customer’s perception of fairness?

The uncomfortable opportunity

The industry should not respond to today’s distrust by simply defending itself more aggressively.

Some criticism is unfair, and some statistics are presented without sufficient context. Some claims disputes are considerably more complicated than the headlines suggest. And insurers are confronting genuine challenges from catastrophe risk, inflation, litigation, repair costs and capital requirements.

All of that is true. But another truth is equally important:

People don’t trust institutions simply because the institution can prove it is technically correct.

Trust comes from transparency, consistency, empathy and evidence. The industry has an opportunity to use the same technology driving transformation to address the trust problem. AI can make insurance more automated—but also more transparent.

Data can make underwriting more sophisticated—but also help consumers understand their risk.

Claims analytics can reduce leakage—but also identify where customers experience unnecessary friction.

And stronger profitability can provide the capital to invest in better products, better experiences and better claims operations.

The real question

Insurance does not need to become universally loved.

It needs to remain credible.

When someone buys homeowners insurance, auto insurance or commercial coverage, they are not really buying a policy document. They are buying confidence that when something goes wrong, someone will stand behind the promise.

That promise is the product.

The insurance industry has a trust deficit. The question is whether carriers and the broader ecosystem will treat that deficit as a public-relations problem—or recognize it as a structural business problem, a technology problem, a claims problem and ultimately a product problem.

The industry has spent enormous amounts of money making insurance more sophisticated. Perhaps the next investment should be making it easier to believe.

The policy isn’t the product. Trust is.

Paul Carroll

You’ve long been my go-to guy on the business implications of artificial intelligence. To start us off, how would you set the table for insurance agents and brokers in terms of how they should think about how AI will play out?

John Sviokla 

Well, first, I think it's important to reassert that salesmanship is going to continue to be needed—probably even needed more. How salespeople serve, what segments they serve, and so forth will be different, but the role remains.

The people who go too far toward automation are going to lose. The people who don't automate enough are going to lose. There's going to be a new balance point.

AI will not change the fact that a very small number of salespeople sell most of the stuff. It's not going to democratize selling, I don't think. In fact, the distance between the top people and the average is going to increase, not decrease. 

Technology is an amplifier. Getting a bigger amplifier does not make you a better guitar player. But if you're a really good guitar player, you'll likely have a bigger audience. Right?

Paul Carroll

You and I have talked about a few areas where agents and brokers should be adopting AI. Would you walk us through those? 

John Sviokla

The critical things a salesperson needs to understand in the early going are account planning, product and service knowledge, and sales skill. AI can help in each of those areas.

It can help you in account planning in terms of finding people and profiling them. It can help you with sales skill through simulation. You can actually tell an AI, "I'm going to sell this person. Here's their background. Go find out about them. Here are my products. Help me understand what's right for them." Then you can actually simulate the sales process. AI is a flight simulator for selling if you do it right.

The third thing is product and service knowledge, and AI is miraculous in this area. Even the general models know a ton. But certainly, if you have any kind of custom setup, it's trivially easy for you to load up all your product information, compliance forms, all that other stuff in a local database. 

AI can be a big cost driver because of the turnover rate in distribution channels that have high turnover rates. LIMRA says, for instance, that as many as 80% of new life insurance agents quit in the first 12 to 18 months. You sell to your relatives, and then you're out of the business.

Even if an agency doesn’t get better at keeping people, you’ll materially reduce your training costs, and you’ll be giving new agents tools and training that should make them more successful sooner. Will you go from 80% churn to 65%? No, but will AI take it from 80% to 75%? Yeah.

Most insurance companies and general agencies stink at onboarding. They think they're good because they spend money, but they aren’t. In an hour with the proper documents, I bet I could create a robot that would be much better than your best trainer on understanding product features, product regulations, and product documentation.

The other thing is that training is once-and-done, even though the half-life of the knowledge that I've been given at the beginning is short. Well, AI is there all the time. You can always ask, “Hey, I'm going to go pitch Paul. Tell me again about how to think about this product. What are the regulatory things again? Does this product work in this IRA, and all that?” That's just not happening right now. Insurance companies and general agents can't afford to keep staff around to answer those questions, to do the genius bar. AI is a genius bar for everybody all the time.

For account management, you look for people in your area, and you put names into the AI. When I pitch top management, there’s enough information out on the internet that I can actually do informal psychological profiles of the individuals and then of the whole team. I ask what points in my presentation are most important for each person, then ask the AI to simulate our interaction. Right now, I’m doing that with text, but in time I’ll be loading pictures of the team and will be talking with them. 

Paul Carroll

Going back to our work together in the early internet days, we saw how it takes a while for people to figure out how to adopt something that's radically new. What tips would you offer agents and brokers on AI? Should they take a top-down approach, bottom-up, or both?

John Sviokla

It's absolutely both. But you have to understand something first—and I'm going to make a blanket proviso about something I know is more complicated than the way I'm expressing it. 

AI is not a technology. AI is a capability: a capability that you have to grow, that you cannot buy.

You buy the beginning of it, but it's like buying seeds or buying bulbs. You've got to grow them. The people need to learn, the machine needs to learn, and then that group together—the people and machines—needs to learn.

The proper way to think about it is what we call our RISE adoption model, which is the old S-curve adoption model with some tweaks. 

The R part is research and education, and that's a pure investment. The proper measures of that stage—do you succeed or not—are two: Do people use it, and do they feel confident while they're using it? You bring people in and train them up.

The I is islands of automation. You invest in a portfolio of innovations. That cost should be net zero because you lose a little money here and make a little there as you build a capability.

Then you go to scaling, the S part. You manage that the same way as any other initiative. Here's the time frame. Here are the inputs. Here are the outputs. Here's the accountability. Here are the goalposts. All that stuff. 

This is where lots of organizations fall off, for three reasons. One, a lot of executives basically want you to do the scaling work off the side of your desk. "Hey, Paul, that looked great. Why don't you scale it?" No, to scale, I have to deal with data, process, security, scale, updates—all that stuff. 

If you want people to reinvent things, organizations need to give people a way to raise their hand and note that they don’t have the necessary time and resources. Most don’t. That’s the second reason for failure.

The third reason—this is the most complicated—is that organizations in general no longer have the people who actually know how to do process redesign, who can do real R&D, and who know how to buy small. Very, very few large companies know how to buy small. How do you work with small vendors? Nobody knows how to do that.

Finally, the E is emergent intelligence. This is where AI capability becomes self-sustaining and strategically differentiating. AI models improve autonomously. Agentic systems coordinate across the enterprise without constant human direction. Hybrid human-AI decision-making becomes the operating norm. 

Paul Carroll 

What should someone reading this do tomorrow morning to get started?

John Sviokla

First, assess where you are on the RISE framework.

If you're at islands of innovation, what's the process you're going to use to figure out what you're going to scale? If you're at scaling, do you have the three competencies I talked about in enough supply to succeed? 

The measurements of success don’t have to be that complex, by the way. Once you’re scaling, you should see revenue per employee increasing. When you get to emergent intelligence, the revenue per employee should be increasing at an increasing rate.

Paul Carroll

This is great, John. For those who want to learn more, here is a link to your big AI conference this month in Boston, the 4th Annual GAI World 2026. Want to give us the 25-second pitch?

John Sviokla

Thanks. The confererence is about helping business leaders in financial services and healthcare understand where they are, what they need to do next, and how they scale for economic value. The first two days are speakers and some hands-on activity. For those who want to stick around on the third day, we go deep on how to use Claude in your personal and professional life. All of GAI World 2026 takes place Sept. 28-30 at the Hynes Convention Center in Boston.  We would love for members of the ITL audience to come and learn from practitioners who are scaling and getting value.

Paul Carroll

Super. Thanks, John.

About John Sviokla

Dr. Sviokla is the chairman and co-founder of GAI Insights. Throughout his career, he has explored the practical implications of leading technologies. He is widely published and was a partner at PwC, vice chairman of Diamond Technology Partners, and a professor at Harvard Business School, where he pioneered AI research and AI courses. Dr. Sviokla has his doctorate, master's, and BA from Harvard University. He was named an executive fellow at Harvard Business School to develop cases for the MBA and executive ed programs and is a Forbes contributor.

Organizations today are operating in an era of continuous transformation. Artificial intelligence, automation, shifting workforce expectations, and evolving customer demands are reshaping how companies create value and sustain performance. For insurers, this reality is especially pronounced as carriers modernize core systems, explore AI-enabled workflows, improve customer experiences, and accelerate digital payment.

In this environment, change management can no longer be viewed as a one-time project discipline. It must become an organizational capability. As McKinsey notes in The State of Organizations 2026, the era of returning to “business as usual” after a transformation initiative is over; the new normal is continuous change.¹

This has important implications for leadership. The organizations most likely to succeed are not simply those with the best technology strategies. They are the organizations with leaders capable of building trust, creating clarity, and helping people adapt to new ways of working.

Technology may enable transformation, but people determine whether transformation succeeds.

Change Starts With Leaders

According to McKinsey, today's leaders must take an “inside-out” approach to leadership, recognizing that leading others also requires leading oneself.² In a business environment shaped by AI disruption and continual change, leaders are being called to redefine leadership in more human-centric terms.

This means balancing performance with empathy, confidence with vulnerability, and execution with reflection. Leaders must still deliver results, but they also need to create environments where people feel safe to learn, experiment, and adopt new behaviors.

The business case is compelling. McKinsey found that organizations adopting human-centric leadership practices report increased employee satisfaction and retention, strengthened trust, improved decision-making, and greater organizational adaptability and resilience.³

As organizations integrate AI into everyday work, this leadership shift becomes even more important. Research from MIT Sloan's The EPOCH of AI: Human-Machine Complementarities at Work identifies five distinctly human capabilities — Empathy, Presence, Opinion, Creativity, and Hope (EPOCH) — that complement rather than compete with artificial intelligence.⁴ Organizations that invest in these capabilities help employees work alongside emerging technologies rather than fear being replaced by them.

Leaders who cultivate these traits within themselves and their teams will be better positioned to help their organizations adapt, innovate, and thrive.

Trust Is the Foundation of Adoption

When employees encounter major change, they do not simply evaluate a new technology. They evaluate what the change means for them.

• Will I have the skills to succeed?
• Will my work become easier or more complicated?
• Will leadership support me through the transition?

Recent insurance industry experience reinforces this reality. A 2026 Carrier Management article profiling Indiana Farmers Insurance described AI transformation as “a culture problem dressed up as a technology and training problem.” Rather than focusing solely on technology, the organization emphasized leadership alignment, employee engagement, skills development, and behavioral change — significantly increasing both usage and confidence among leaders.⁵

The Prosci success story involving McCarthy Holdings highlights a similar lesson. By focusing on communication, training, feedback, and manager support, the organization achieved 90% adoption of its AI-powered work platform within 30 days.⁶

The message is clear: adoption happens when people understand the purpose, see practical value, trust the process, and feel supported by leadership.

Adoption Must Be Designed, Not Managed After the Fact

One of the most common mistakes organizations make is treating change management as a launch activity. Technology is deployed, training is rolled out, and teams are expected to adapt.

Leading organizations take a different approach.

Human-centric leaders recognize that adoption starts long before go-live. They build organizational readiness during planning, engage stakeholders early, identify barriers, and communicate how change will affect daily work.

This principle is particularly important in digital payment transformation. Success depends not only on technology implementation, but also on how effectively insurers prepare claims teams, billing departments, customer service representatives, treasury functions, vendors, and policyholders for new workflows.

The most effective transformation programs embed change management throughout the project lifecycle — from discovery and design through implementation and optimization. When education, communication, and adoption planning are addressed early, organizations reduce disruption, accelerate adoption, and create a better experience for employees, vendors, and customers alike.

Successful adoption is rarely accidental. It is designed.

This reality also changes how organizations should evaluate technology providers. The most effective partners do more than deliver software; they help organizations build readiness, educate stakeholders, and create the conditions necessary for successful adoption from the beginning of the transformation journey.

Managers and Influencers Matter More Than Ever

Executive sponsorship remains critical, but employees often turn first to their immediate managers for guidance and reassurance.

Human-centric leaders recognize that managers must be equipped — not simply informed. They need the tools, context, and confidence to help their teams navigate uncertainty.

Leadership visibility matters as much as leadership sponsorship. During Indiana Farmers Insurance's AI transformation initiative, CEO Wes Sprinkle actively participated in training sessions and used the technology himself.⁷ When leaders model desired behaviors, they create credibility that accelerates adoption throughout the organization.

Trusted influencers across the organization also play an important role in reinforcing the purpose of change and building momentum from within.

As transformation becomes continuous, responsibility for successful adoption must extend beyond project teams. It must be shared across leaders, managers, and employees alike.

A Final Thought: Transformation Requires More Than Technology

As insurers continue investing in AI, automation, digital payments, and customer experience modernization, the people side of change will increasingly determine the return on those investments.

Technology implementation may mark the beginning of transformation, but adoption determines whether transformation delivers lasting value. And adoption depends on trust, communication, leadership, education, reinforcement, and a clear connection between change and meaningful work.

Human-centric leadership is not optional. It is the foundation for resilience, innovation, and sustainable growth.

It also changes how organizations should think about implementation partners. The most successful digital transformations recognize that outcomes depend as much on people, processes, and adoption as they do on technology itself. Organizations that address all three dimensions are far more likely to accelerate adoption, minimize disruption, and realize the full value of their transformation investments.

Sources:

1. McKinsey & Company – https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-state-of-organizations
2. Ibid.
3. Ibid.
4. MIT Sloan Management Review – https://mitsloan.mit.edu/press/new-mit-sloan-research-suggests-ai-more-likely-to-complement-not-replace-human-workers
5. Carrier Management – https://www.carriermanagement.com/features/2026/06/11/288942.htm?bypass=1030be34e07ce01533730034b2c8aa20
6. Prosci – https://www.prosci.com/resources/success-stories/mccarthy-building-companies-achieves-90-ai-adoption-with-prosci
7. Carrier Management – https://www.carriermanagement.com/features/2026/06/11/288942.htm?bypass=1030be34e07ce01533730034b2c8aa20

When AI gets something wrong, who bears the loss?

For the past few years, that question has mainly appeared in discussions of AI governance and legal liability. But as AI moves into core business processes—and as agents begin to execute tasks autonomously—another market is being forced to confront the same question: insurance.

An IT Pro article on AI insurance notes that a new category of coverage is beginning to emerge around AI-related risks. These may include financial losses caused by hallucinated advice, algorithmic discrimination, copyright infringement, models that fail to meet promised performance levels, and even losses caused by autonomous AI agent actions.

This is not yet a mature or standardized insurance product. Coverage is still scattered across standalone policies, endorsements, and exclusions. But the more important development is not simply that insurers are beginning to cover AI. It is that once insurers agree to absorb losses caused by AI, they must answer a harder question:

When an agent can act autonomously, is the insurer underwriting model risk, system risk, or task-governance risk?

What kind of AI deserves to be insured? That question reveals a deeper shift: AI governance is becoming part of the infrastructure of insurability. And it may, in turn, change how companies govern AI. As agents gain greater autonomy, a company’s ability to show that their risks can be understood, constrained, and traced will increasingly shape whether those systems are insurable.

For insurers willing to underwrite AI risk, that challenge is also becoming a market opportunity.

1. From Risk Governance to Insurability

In mature risk markets, insurers first ask whether a risk can be identified, controlled, and estimated. Only then can they decide whether to insure it and at what price.

AI is beginning to enter that process.

Gartner predicted in 2026 that by 2030, property and casualty (P&C) insurers may require robust AI risk controls as a condition for providing affirmative AI liability coverage. Companies with stronger AI governance and risk-management capabilities may not only find it easier to obtain coverage, but may also receive more favorable pricing.

That means companies building AI governance will need to answer more than “How do we manage AI?” They may also need to answer: Can we prove to an insurer that we actually have AI under control?

This is not a paper exercise. An AI policy, a set of ethical principles, or even an AI governance committee is not enough to demonstrate that risk is truly under control. What insurers should care about is whether those governance practices have become operational mechanisms before an incident occurs—and what, in practice, the company has changed because of them.

To assess AI risk, insurers will need to know which models a company uses, which decisions those models influence, and who is responsible for those decisions; whether automated actions are logged; whether high-risk situations trigger human intervention; whether model drift and anomalies are detected; and whether incidents can be reconstructed after something goes wrong.

As one expert quoted by IT Pro put it, governance that exists only on paper, without being translated into operational mechanisms, will not satisfy a prudent underwriter.

AI governance is therefore acquiring a function that has received far less attention: it is becoming part of the infrastructure of AI insurability.

2. What Insurers Underwrite Will Extend Beyond the Model

There is still an unresolved problem. Much of today’s AI risk management remains model-centric. Companies maintain model inventories, test accuracy, monitor bias and model drift, and preserve output records. For generative AI, hallucinations, discriminatory behavior, and incorrect outputs can all create direct financial loss.

But once AI becomes an agent that can execute tasks autonomously, the nature of risk changes. An agent may correctly execute a task that has already become irrelevant. It may continue pursuing the original objective after new information changes the situation. Or every individual action may comply with the rules while the combined result still creates an outcome the company cannot accept.

At that point, model capability alone is no longer enough to assess the risk. The exposure insurers need to understand begins to extend from model risk to task-governance risk.

The underwriting question therefore becomes:

Under what conditions is this AI allowed to do what it does? And if the context changes, will it know when it should stop?

For companies, that raises another question: how do they establish the insurability of an AI agent?

3. Agent Insurability Requires Evidence

Imagine a company that allows an AI agent to process customer refunds autonomously. From a model-governance perspective, we can test whether its answers are correct, whether data is secure, and whether its outputs show bias. But an insurer may also need to know:

How much money can the agent refund? When must it reassess what the customer actually needs? If suspected fraud, a new complaint, or an abnormal transaction appears, is the original task still valid? Who determines whether the agent may continue, pause, or hand the case to a human? And what evidence shows that the task was actually completed as intended?

These questions go beyond conventional model governance. They point to four task-governance mechanisms that can also serve as underwriting evidence:

Task design — defines what the AI is actually authorized to accomplish. For an underwriter, it helps establish where the exposure begins and where it ends.

Permission boundaries — define which data, systems, assets, and actions the AI may access or change, and how far its authority extends. These boundaries directly shape the scale of potential loss.

Understanding loop — determines whether new signals cause the system to reassess the task itself rather than simply continue executing the original objective. This is critical when a change in context should trigger a pause, escalation, or human intervention.

Outcome verification — defines what counts as task completion: merely finishing a process, or producing the business outcome that was actually intended. It also creates the evidence needed for auditability and traceability.

Companies may think of these mechanisms primarily as ways to control agents. From an insurance perspective, however, they serve another purpose: they are evidence that an agent’s risk can be understood, constrained, and traced.

4. Insurance Is Becoming an External Constraint on AI Governance

Internal AI governance has a familiar problem: governance rarely generates revenue directly. Under pressure to move faster, increase efficiency, and deliver business results, it can easily deteriorate into policies, checklists, and compliance documents.

Insurance may change the incentives for companies to govern AI well.

A 2026 AI-insurance study accepted by Harvard Data Science Review points to a fundamental difference between insurers and ordinary AI assurance providers: insurers actually bear the cost of AI-related claims. That gives them a direct economic incentive to demand more rigorous testing, monitoring, and validation—and, through premiums and underwriting conditions, to encourage safer AI governance.

That mechanism is not new. Fire protection, automotive safety, industrial risk controls, and cybersecurity have all been shaped by insurance pricing and underwriting requirements, gradually turning safety practices into standards that markets can verify.

AI may now be entering a similar stage. If a company cannot explain an agent’s task boundaries, operating permissions, intervention mechanisms, outcome records, and post-incident traceability, it may face more than regulatory scrutiny. An insurer may simply conclude:

I cannot price this risk.

Or:

I can insure it—but at a higher premium, with lower limits, and with certain autonomous agent actions excluded.

Once AI governance capability begins to affect whether a company can obtain coverage, the terms of that coverage, and the premium it pays, governance has entered the company’s risk pricing. That may be one of the most important futures for AI insurance.

Conclusion: Are Insurers Ready to Judge AI Insurability?

Discussions of AI liability often end with a familiar principle: no matter how advanced AI becomes, humans remain ultimately responsible.

The principle is sound. But as agents gain greater autonomy, the harder question is no longer whether someone can be found to bear responsibility after the fact. It is whether the company can demonstrate in advance that human responsibility has been translated into operational control mechanisms inside the system.

Who defines the task? Who configures the permissions? What signals require the system to reassess? Under what conditions must it stop or hand control back to a human? And how do we prove that the AI completed the right task—not merely that it executed correctly?

These questions once belonged mainly to AI governance. AI insurance is turning them into underwriting questions.

That could mark an important shift. Once the insurance market begins to price AI governance capability, governance is no longer something a company simply believes it “should do.” It may increasingly determine whether the company is entitled to grant an agent greater autonomy, whether outsiders can trust its AI, and whether the resulting business risk can be transferred.

What ultimately differentiates enterprise AI capability may therefore be more than model strength or the number of tasks an agent can perform. It may be whether a company can prove that, even when AI begins to act autonomously, the resulting risk remains understandable, controllable, and insurable.

Seen from the other side, insurance in the AI era is not only about covering losses caused by AI. It is also about assessing whether a company can turn AI risk into something understandable, priceable, and subject to intervention.

That brings the question back to insurers themselves: as risk moves from models to autonomous agents, do insurers already have the capability to judge whether those agents are insurable? If not, AI insurance will struggle to develop beyond the concept stage.

That capability is not only a prerequisite for AI insurance to become a real market. It may also become one of the insurance industry’s next essential areas of expertise.

Salesforce and Anthropic announced Claudeforce in late August. Here’s the pitch, stripped of the press-release glaze: Claude’s reasoning now sits inside Salesforce’s workflows, and Salesforce’s data now sits inside Claude. Ask who last spoke with a client, and you get a clean answer as fast as you can type. What used to take an hour clicking through six screens now takes one sentence. 

Great for users. But for executives, that’s not the story.

The story is what happens when you ask a harder question: “What does our churn risk look like over the next 90 days, and who’s at risk?” The system shrugs. That shrug is the whole ballgame.

For 30 years, insurance technology strategy has had one move: consolidate. First we digitized—paper became PDFs, phone calls became email. Then we platformed, on systems like Guidewire, Duck Creek, and Majesco, on the theory that if everyone were on the same platform, the enterprise would finally speak one language. Both eras shared a quiet assumption: the human is the worker, and the computer is the tool that helps the human do the job faster.

Claudeforce is the first real crack in that assumption. It forces executive committees to answer a question that used to be hypothetical: Are we still building technology for people to use, or are we starting to build an operating environment where machines increasingly do the work?

At root, that’s a technology problem with a technology answer. Palantir CTO Shyam Sankar put it well: “If you try to fix the system, you’re going to lose. The goal is to win without fixing systems.” Translation for insurance: stop trying to force every legacy system into one warm data lake. Build the semantic layer instead—the ontologies, object models, permissions, and APIs that let an agent understand what your systems mean, not just what they contain. You don’t need one database. You need a distributed operating environment that machines can read and act on.

That has obvious capital allocation implications. The win isn’t consolidating data for humans to navigate; it’s making distributed data intelligible for machines to act on. Unified systems still matter where economics or regulation demand it. But unified experience no longer requires unified data—it requires a semantic layer good enough that an agent can move across your fragmented systems the way a good analyst does. The center of gravity of your technology investment plan shifts.

It also has unavoidable cultural implications. Building a machine-readable layer over how work gets done is, whether you say it out loud or not, an investment in machines doing the work. The human-to-machine ratio stops being an automation metric buried in an ops review and becomes the operating model itself.

The transition from workers having computers to workers being computers requires leadership and a spine, not just a rollout plan.

I’d position the transition as great news—which I sincerely think it is. The jobs on the chopping block are mostly the digital rockpile: the manual lookups, the re-keying, the 12-screen scavenger hunts that human beings hate and that quietly corrupt information as it moves through the enterprise. Nobody’s proudest work is chasing down which system has the current address on file.

Look at Amazon—the app, the checkout, the returns flow, the whole digital experience. When’s the last time you talked to a human to get something resolved there? Not the co-branded credit card; that’s a bank. The actual Amazon experience. For most of us, the honest answer is never, and we don’t miss it. That’s not an absence of humans (Amazon employs 1.6 million people). Plenty of people built, and keep improving, the system that makes it work. It’s an absence of rockpile jobs. Nobody at Amazon is manually checking your order status or re-keying a return into a second system, because the digital experience was built end-to-end for machines to run, with no human standing in the loop as middleware or connective tissue. The humans create the value upstream; the system delivers it downstream, on its own.

That’s the shift Claudeforce points to for insurance: not fewer systems, but the right systems—semantic infrastructure that lets machines navigate the mess so humans don’t have to. Your next technology dollar shouldn’t buy another layer of software for people to learn, but the machine-readable layer that makes the existing stack feel smaller. More technology, yes—but the kind that finally makes it feel like less.

If you lead architecture, pricing or compliance in personal lines, you have probably been handed two mandates in the same quarter: get predictive pricing into production faster, and be ready to show a state examiner that it does not discriminate. The standard answer to the second mandate is explainable AI. I spent a year testing whether that answer holds up, and it does not, not because XAI is bad but because it arrives too late to do the job carriers are assigning it.

The truth is that you cannot explain your way out of algorithmic bias. You have to intercept it before the rate exists, and interception costs far less in latency than most architecture teams assume.

Where the current approach breaks

Personal lines pricing has moved from static actuarial tables to multi-layered predictive models fed by telematics and third-party data. The risk stratification is better. The governance exposure is different in kind from anything traditional model validation was built to catch: a deep neural network can construct proxies for protected classes on its own, from variables such as garaging ZIP code or credit history tenure, without a protected attribute ever appearing as an input.

Most teams answer this with post-hoc explainability. Those methods do real work: they describe how a model reached an output and make audit response tractable. But explanation is retrospective by construction. It tells you why a rate was biased after that rate exists. It cannot stop a discriminatory rate from reaching a consumer. That is a control gap, and interpretability does not close it.

The consequences are no longer theoretical. Colorado amended Regulation 10-1-1 effective Oct. 15, 2025, extending its governance framework for external consumer data and information sources beyond life insurance to private passenger automobile insurers and naming telematics explicitly as ECDIS. Interim progress reports were due Dec. 1, 2025; compliance reports became due July 1, 2026 and recur annually. New York's Insurance Circular Letter No. 7 expects insurers to assess whether AI systems function as proxies for protected classes, and makes clear that responsibility does not transfer to a vendor. More than half of U.S. jurisdictions have adopted the NAIC Model Bulletin or substantially similar guidance.

What changed this year is examination capability rather than new obligations. The NAIC has been piloting an AI Systems Evaluation Tool across 12 states since early 2026. The pilot runs through September, with a revised version expected to be considered for adoption at the Fall National Meeting in November. It gives examiners a structured framework for reviewing insurer AI governance during market conduct and financial examinations. The question is no longer whether your program is documented. It is whether it can be demonstrated on request.

The three-layer architecture

The fix is structural rather than statistical. Put a deterministic layer between the model and the rating outcome.

Demote the model to a data generator. The predictive model calculates a risk score. It does not hold authority to issue a premium. This is a configuration decision more than a technical one, and it is the step most programs skip.

Give the rules layer the authority. Route model output into a rules engine that encodes state-specific constraints before anything reaches policy administration. Where a recommendation depends on latent proxy variables, the engine rejects it in-line. Encode the constraints per state rather than at the enterprise level, so the mapping from a control to a regulatory obligation lives in configuration instead of in prose.

Log every evaluation, not just every interception. A record written only when something is blocked cannot demonstrate coverage. A record written for every transaction turns your compliance evidence into a query rather than a project.

Two things to watch. A constraint set that is not updated as filings change becomes a source of error rather than protection. And overly restrictive constraints suppress legitimate model lift, so monitor the false-positive rate of the control layer as deliberately as the interception rate.

What the numbers showed

I tested this design and published the results in Cureus in July 2026. The trial used 10,000 synthetic multi-state personal auto underwriting profiles, generated by Monte Carlo simulation and deliberately seeded with historical proxy correlations.

Before: the standalone predictive model produced a proxy pricing disparity of 14% against protected demographic groups.

After: routing identical model outputs through the deterministic rules layer reduced that disparity to 0.14%, a relative reduction of 99%.

Three secondary findings mattered as much. The rules layer intervened on 1,976 of 10,000 transactions, or 20%. That is how much non-compliant recommendation volume a model can generate before any human sees it. Mean processing latency moved from 120 milliseconds to 125 milliseconds under simulated enterprise load, which is the finding that answers the performance objection this design reliably attracts. And a decision record was generated for 100% of transactions.

These results come from synthetic data under simulated load. They demonstrate the mechanism; they are not a production benchmark. Live environments introduce third-party API orchestration and vendor data latency that a controlled trial does not reproduce.

What to do next

Bias mitigation belongs in the architecture, not on the data science backlog. The carriers that hold up under examination will not be the ones with the most complete policy binder. They will be the ones whose systems produce the answer as a by-product of running.

Take one question into your next cross-functional review: can we produce, for any rating period, the population of AI-influenced pricing decisions and the subset where a compliance control fired? If answering it requires a project rather than a query, your evidence architecture is retrospective, and retrospective evidence is exactly what a structured examination is designed to find.

Five milliseconds is what the guardrail cost. Decide whether that is the expensive part.

Sabine VanderLinden, a keen observer of venture capital in insurance, recently wrote a startling sentence: "50% of every insurtech dollar invested in the first six months of 2026 went to companies that will never sell you a policy."

That number is a far cry from what proponents expected when the insurtech wave began a decade-plus ago. At that point, predictions were rife that some Big Tech company such as Google or Amazon would do a cannonball into insurance and change the game entirely or at least that some startup would figure out a way to leapfrog incumbent carriers and make them play catch-up. 

But VanderLinden's analysis provides a guidepost about where investment in insurtechs is today and where I think it's going.

Let's have a look. 

Artificial intelligence still takes up the vast majority of the headspace for most incumbents as they try to innovate. They're spending enormous effort to look for efficiencies in processing claims, in underwriting, in sales, and so on. They're also experimenting with ways to set up autonomous agents and to coordinate their actions while staying within crucial guardrails.

VanderLinden's analysis found significant funding for AI-based insurance startups, too, but they were just the third biggest category in the first half of the year. First was: "risk data. Satellites, sensors, and driving behavior.... This is happening because proprietary risk data has become the scarcest asset in the value chain. Models are abundant. Compute is abundant. Ground truth is not. The ventures that own a persistent, hard-to-replicate view of physical risk are commanding late-stage checks."

She highlights ICEYE, which raised a $500 million Series F "to expand its radar satellite constellation for natural catastrophe monitoring," and Cambridge Mobile Telematics, which raised $350 million for its insights into driving behavior. She also mentions mea platform ($50M), Fulcrum ($25M) and Axle ($17.5M).

Her observation certainly dovetails with what I'm seeing. I've long argued that the surest insurtech winners would be what we called "arms dealers" during the early internet days. Just as Sun Microsystems made bank by selling servers to startups, whether they thrived or, more likely, crashed and burned, companies that developed important, proprietary data sources were always likely to thrive. 

And there have been impressive advances in risk data, as evidenced by any number of articles we've published recently at ITL. This interview I did with Eagleview lays out a vision for how aerial views of properties will enable continual monitoring of property risks. This piece, from Nearmap, describes how the condition of roofs and other aspects of properties can be tracked long before a claim surfaces. This describes advances in "hyperlocal" weather intelligence. This explains how catastrophe modeling is moving beyond static pictures of disasters and toward images that show how floods, wildfires, etc. develop over time. We've also published on new ways to track maintenance records of commercial properties to better understand the likelihood of a claim, to monitor for the next pandemic, and so on. 

VanderLinden says the second biggest category of venture investment was in digital-first insurers and MGAs and offers a key insight: "Not one of them is a generalist." She writes:

"Alan raised $116M for digital health in France. Corgi closed a $106M Series B insuring technology companies. Counterpart took $50M for small-business liability, Shepherd $42M for construction, Lassie $75M for pet, Zego $28M for gig drivers. Stoïk and Mitigata both raised for cyber, on two different continents.... The funded insurance provider of 2026 is specific, defensible, and priced for its niche."

Third was AI: "$216M went to AI-led claims, underwriting, and operations automation.... These ventures do not compete with insurers. They sell digital labor to them. Claims handling, prior authorizations, underwriting triage, document verification: the workflows where a human-agent ratio can shift fastest and the savings land on the P&L within quarters, not years."

The insurtechs that went after incumbents head-on -- notably Lemonade, Hippo and Root -- are still around and seem to have stabilized after years of struggle but are way down from their peaks in early 2021. Lemonade shares are off some 65%, Root is down 85%, and Hippo has fallen 90% even as the S&P 500 has nearly doubled. So I continue to believe that the sorts of "arms merchants" VanderLinden describes are the future of insurtech.

Cheers,

Paul

P.S. If you'll permit me a proud papa moment....

My older daughter made her debut in the Wall Street Journal over the weekend. She trekked the 500-mile Camino de Santiago in Spain this spring and wrote an essay [free link] about the trip that has generated some 500 comments and emails and spent some time on the "most popular" list. It's a splendid piece. I'm delighted for her.

I also love that the Carroll family is now on its third generation at what we joke is the family business. My father spent a year at the WSJ. My younger brother and I combined for 59 years. Now Shannon....