As soon as I heard about Flock, the now-huge network of cameras that can automatically read license plates and instantly report their location to authorities, I wondered how long it would take before we read about police officers using the network to track girlfriends, wives and exes. 

It actually took longer than I thought, but here we are: The Washington Post reported a week ago that at least 50 officers throughout the U.S. have been accused of stalking by using Flock and similar systems. All sorts of national media have now followed the Post story, and we're in an environment of heightened sensibilities about surveillance and about how scary AI might turn out to be, so this story isn't going away any time soon. 

In the swamp that is today's social media, some conspiracy theorists are trying to drag in the insurance industry, saying it uses Flock cameras in nefarious ways. As the scandal unfolds, we need to not only make sure we avoid any unethical uses of camera networks but can respond quickly and convincingly to any online trolls alleging abuse.

Let's have a look. 

Many years ago, a good friend was driving an old car in Washington, DC, when he stopped at a stop sign and saw a pedestrian motioning to him that he needed directions. Tim rolled his window down, and the guy stuck a gun in his face, ordering him out of the car. But the car's clutch was dying, and the carjacker couldn't figure it out. He lurched a few feet and stalled. Another few feet and another stall. He stalled a third time before he even made it halfway through the intersection. He got out of the car, threw the keys into Tim's chest and ran off.

If the carjacker had succeeded and been in modern DC, he would have been a perfect candidate for Flock. Tim could have simply told a 911 dispatcher what his license plate was, and Flock's 120,000 cameras throughout the U.S. (except Alaska) would have instantly been on the lookout. Flock works with 7,000 police departments, or 40% of the total in the U.S., so once its AI spotted the license plate on the stolen car it would have been able to alert police and have a pretty good chance of recovering it. 

In theory, Flock will gradually prevent more and more thefts, because thieves will either be caught and incarcerated or will simply realize that car theft is a losing proposition.

I spend a lot of time with agency principals at very different stages of growth. Some are running lean operations in a single market. Others are managing multi-state books across several product lines. One thing I hear consistently, almost regardless of scale, is some version of this sentence: "We had our best enrollment year ever." My next question is always the same: What did that actually do for the health of your business?

The insurance distribution industry has spent years treating enrollment volume as the definitive signal of agency health. For most of the last decade, the underlying economics cooperated. A prolonged hard market in property and casualty, combined with rising premiums across many health and medical product lines, created an environment where agency revenue continued to grow, even when relatively few leaders stopped to examine what was actually driving it.

Most agencies have not had a reason to look too closely at those numbers, and the market has not forced the question. That is starting to change.

The Hidden Math Behind the Headlines

The 2025 Best Practices Study found top-performing agencies posted organic growth rates consistently between 10.3% and 10.7%. Those are historic numbers. On the surface, they make a compelling case that chasing enrollment is the right strategy.

The problem is what sits underneath. Financial benchmarking data shows that for the average insurance brokerage, roughly seven percentage points of that double-digit growth figure are attributable to premium rate increases and exposure base expansion rather than new accounts or deeper relationships with existing clients. Strip out the inflationary tailwinds, and actual organic growth for the average firm lands somewhere between 2.7% and 3.2%.

That is a very different business than the one most agencies believe they are running. The industry has benefited from macroeconomic tailwinds that made revenue growth look like strategic growth; premium inflation was confused with operational excellence. When the market softens and premium inflation cools, the distance between those two numbers will become impossible to ignore.

The Retention Gap Nobody Talks About

Industry data puts the average retention rate for a single-policy client at roughly 77%. An agency whose book is built primarily on single-policy households is losing close to a quarter of those clients every year. The agency is operating an expensive treadmill, constantly spending on client acquisition just to hold revenue flat.

When an agency deepens the relationship to five or more policies per client, that retention rate climbs to approximately 85%. That increase in retention sounds modest, but the compounding math over a five-year horizon is anything but. A multi-policy client is nearly twice as likely to still be in the book five years from now, representing roughly a 60% improvement in long-term customer value driven entirely by product depth rather than acquisition.

What makes this even more striking is how pervasive the single-policy problem is. Research indicates that approximately half of the average firm's customer base holds only one policy. That means half of the average independent agency's book of business is both an untapped revenue opportunity and a serious flight risk, simultaneously.

Why the Industry Stays Stuck

In most agencies, the barrier to deeper product relationships is operational, not strategic. The independent distribution ecosystem was designed to maximize transactions, not deepened relationships. Every new product can introduce another appointment, enrollment workflow, commission schedule, and other points of friction. Eventually, the economics favor moving on to the next prospect rather than deepening the relationship with the current one.

Consider what it takes to build a comprehensive household relationship spanning a major medical plan, a hospital indemnity policy, and a term life product. An agent navigates three separate carrier portals, re-enters the same client data three times, manages three different quoting engines with three different underwriting workflows, and then attempts to reconcile three fundamentally incompatible commission structures on the back end. That extra 30 to 45 minutes per client, multiplied across an open enrollment season, is simply not compatible with high-volume production targets.

The compliance landscape compounds the challenge, particularly in the senior market. CMS Marketing and Communications Guidelines generally prohibit agents from using a Medicare sales appointment to cross-sell non-health products without a documented Scope of Appointment secured at least 48 hours in advance. As a result, an agent who identifies a legitimate cross-selling opportunity often cannot pursue it in that moment. Instead, the opportunity requires a separate appointment, additional documentation, and another workflow. Each additional step introduces friction, reducing the likelihood that a valuable client need is ultimately addressed.

None of these barriers are insurmountable. They are real, though, and they explain why agencies that understand the economics of product depth still struggle to execute against it. The gap between what agencies know they should do and what their infrastructure allows them to do is where retention gets lost.

What Capital Actually Sees When It Looks at Your Book

The insurance brokerage M&A market is as active as it has ever been, driven by private equity consolidators, national aggregators, and large brokers competing for quality books. In an M&A context, the word "quality" has a very specific meaning that has nothing to do with enrollment volume.

Institutional buyers are not purchasing a snapshot of today's revenue—they are purchasing the predictability and durability of tomorrow's cash flows. A client retention rate at or above 90% is considered the standard for premium, platform-grade valuations. Agencies that consistently hit that threshold command meaningfully higher multiples, often one to two turns of EBITDA above lower-retention books. When a buyer identifies a book built primarily on single-policy clients with historical churn running below 80% to 85%, the response is earnout structures that shift up to 40% of the total purchase price into contingent payments tied to retention benchmarks the selling agency is unlikely to meet.

This is the moment that tends to surprise founders who have spent years optimizing for enrollment. They assumed the number of clients in the book was what they were selling. Buyers see it differently. They are buying the quality of those relationships, and a book full of single-policy households is a book full of clients who are one price comparison away from leaving.

The Market is Moving Anyway

The argument for product depth has always been sound, but several structural shifts are making it urgent in a way it was not five years ago.

The traditional employer-sponsored group health model is fracturing. ICHRA adoption grew by 34% among large employers and up to 49% among mid-sized employers between 2024 and 2025, as businesses shifted from selecting group plans to providing employees with tax-free dollars to shop the individual market themselves. The expiration of enhanced ACA premium subsidies pushed average deductibles to nearly $3,800 in 2026, sending millions of individuals into the market with complex coverage needs and no institutional support to navigate them.

The senior market is moving in parallel. By 2030, all Baby Boomers will be age 65 or older, with roughly 10,000 Americans turning 65 every day. Original Medicare's well-documented gaps in dental, vision, hearing, and extended hospital stays create natural openings for the agent who takes the time to build a full household picture. Multi-product senior households are among the stickiest in the book. The relationship has real financial consequences for the client, which translates directly into retention.

Carriers are responding to these same market dynamics. The era of rewarding volume alone is steadily giving way to compensation structures that reward persistence and quality of business submitted. As acquisition costs rise and profitability comes under greater pressure, carriers are placing increasing value on agencies that consistently deliver durable books of business rather than simply higher application counts.

A Different Way to Measure Agency Health

The agencies I see creating the greatest long-term value are not necessarily the ones with the largest enrollment numbers. They are the ones that can answer two questions with confidence: how many of their clients hold more than one product and what their 12-month and 24-month retention rates look like by product line. Those metrics don't just describe yesterday's performance; they predict tomorrow's enterprise value.

The agencies leading this transition have recognized that relationship growth is an operational capability, not simply a sales initiative. They routinely identify single-policy households with unmet needs, but more importantly, they've invested in systems that make acting on those opportunities almost effortless. Quoting across product lines, managing compliance requirements, and reconciling commissions are integrated into a single workflow rather than a series of disconnected administrative tasks. When the operational friction is removed, producers stop abandoning cross-sells and the book composition changes.

The valuation gap between those two categories of agency will only widen as the hard market continues to ease and rate-driven premium inflation recedes. When that happens, growth has to come from somewhere real. The agencies that figured that out before they needed to will be in a fundamentally different position than the ones that did not.

Enrollment is no longer the end goal. It is the invitation, the beginning of the relationship. Agencies treating enrolled clients as a relationship will deepen those bonds rather than finish as a transaction. Years from now, they won't simply look back on this market as a period of strong enrollment. They'll recognize it as the period in which they built businesses that compound.

The insurance industry has spent the last three years asking one question: where can we apply AI?

It's time to start asking the harder one: where shouldn't we?

We've seen what AI can do in underwriting, claims triage, fraud detection, and customer servicing. The efficiency gains are real. The speed is real. But as we embed AI deeper into consequential decisions — decisions that determine whether a claim gets paid, whether a risk gets written, whether a customer gets flagged — we're walking into a compliance gap that most organizations haven't fully reckoned with yet.

The definition of compliance is changing underneath us.

Traditionally, compliance in insurance was about people and processes following rules. Regulators asked: did you follow the procedure? Was the policy applied correctly? Can you show me the file?

Those questions still exist. But they're being joined by a new set:

• Why did the AI make this decision?
• What data trained it, and was that data appropriate?
• Can the decision be explained to the customer who was denied?
• Can it be audited three years from now when a regulator comes asking?
• When something goes wrong, who is accountable — the model, the vendor, or the carrier?

These aren't hypothetical questions. They're already landing on legal and compliance desks across the industry. And most AI systems aren't built to answer them cleanly.

Here's the tension nobody talks about openly.

The more we automate decisions with AI, the more we need infrastructure that is not AI to govern those decisions.

Compliance requires consistency, predictability, and an evidence trail that holds up under scrutiny. In many cases, the right tool for that job is deterministic — rules engines, governance frameworks, workflow controls, immutable audit logs, human checkpoints at defined thresholds. These aren't legacy artifacts to be replaced. They're the architecture of accountability that AI, by its probabilistic nature, cannot fully provide on its own.

This isn't a limitation to be embarrassed about. It's an architectural reality to be designed around.

The Auditor Cannot Be the Accused.

You cannot use AI to audit AI.

Yet most organizations today are doing exactly that — monitoring AI models with more AI. If the original model has a flaw, the auditing model likely carries the same one. That's not oversight. That's a mirror.

We've already seen AI fabricate citations — confidently, cleanly, and completely wrong. Now imagine that happening in a claims denial or an underwriting decision, where the AI made the call and wrote the audit trail.

Regulators will ask for explainability reports that were produced independently — not one AI explaining another. That documentation needs to be deterministic, traceable, and human-readable. Built without AI. Full stop.

The next wave of RegTech won't be smarter AI. It will be AI governance.

While carriers and insurtechs race to deploy more models, there's an equally significant — and arguably less crowded — opportunity in building the layer that sits above those models. The technology that ensures every AI-driven decision is transparent, explainable, documented, and defensible. The systems that answer the regulator's question before the regulator asks it.

Some of this already exists in nascent form. Model risk management frameworks borrowed from banking are making their way into insurance. Explainability requirements are starting to appear in state regulations around algorithmic underwriting.

The organizations that get ahead of this won't be the ones with the most AI. They'll be the ones who built the governance layer early — and can demonstrate it when it matters.

A word to operators.

If you're running an insurance business and you've deployed AI in any customer-facing or claims-facing capacity, ask your team this week: if a regulator asked us to explain the last 1,000 decisions this model made, could we do it? And could we prove that explanation wasn't itself generated by another AI?

If the answer to either question is uncertain, that's your compliance gap — and it's growing faster than most people realize.

The competitive advantage in this next phase won't come from adding more AI. It will come from making the AI you already have trustworthy enough to defend.

Insurance enjoys the protection of a regulatory moat. Capital requirements, reserve rules, licensing, and market oversight protect policyholders—and, by extension, legacy carriers' market positions.

But the moat is built around the balance sheet, not the operating model. It was never designed to protect expense ratios, organizational layers, underwriting workflows, claims operations, or administrative overhead.

Private equity is finding the seams—the fee-generating, capital-light nodes adjacent to the regulated core—and systematically buying them, then using AI to compress cost inside while the actual risk gets parked elsewhere. Here's where the action is happening.

1. MGA and program administration

MGAs underwrite and distribute without carrying capital—and, in excess and surplus lines, without the rate-filing straitjacket that binds admitted business. In the E&S market, no state Department of Insurance reviews an actuarial memo before an MGA changes pricing. That's not a technicality; it's the ballgame.

Roll-up money has been pouring into program administration at a pace that should alarm anyone who thought "we're regulated, we're safe." AI-driven underwriting here doesn't need to win a fight with a state commissioner—it just needs to be better and faster than the human it's replacing.

2. Claims, fraud, and subrogation

Loss costs and loss adjustment expense consume roughly 60–70 cents of every premium dollar. That's the single largest pool of spending in the industry, and it's almost entirely a process problem—not a regulatory one. A state insurance commissioner has opinions about your rate filing, not about whether your computer vision model estimates hail damage better than an adjuster with a clipboard, or whether an natural language processing (NLP) model flags subrogation opportunities your staff missed while chasing cycle-time bonuses.

This is the cleanest PE trade there is: buy or build the platform, automate the workflow, collect the toll. No rate filing required, no market conduct exam in the way. Claims processing today is like bank credit card processing 30 years ago—bespoke and fragmented. Look for rapid consolidation. Think: Visa and Mastercard. Blackstone, Apollo, and KKR certainly are.

3. Fronting

Fronting carriers exist so an MGA or insurtech can write business without holding a balance sheet. They rent a fronting carrier's paper and lay the risk off to a reinsurer (often also private capital). It's elegant financial engineering, and it's grown explosively because it lets everyone upstream of the actual risk-bearing get paid without ever touching the risk. Though regulators have started naming fronting arrangements specifically in their review of PE ownership structures, real action may be years away.

4. Agency roll-ups

This isn't new, and it isn't really disruption any more so much as consolidation. Acrisure, Hub, BroadStreet, AssuredPartners—PE has been buying up the independent agency channel for 15 years, riding a demographic wave of retirement-age owners with no succession plan and no appetite to fight for a better multiple. It works. It will keep working.

AI helps at the margins—better cross-sell, better retention scoring—but the trade was never about the tech. It was about the math, and the math has gotten more expensive as everyone's figured it out.

5. Data and risk analytics

Everybody wants to be Verisk. Almost nobody gets to be, because Verisk, CoreLogic, and LexisNexis Risk Solutions already occupy the high ground, and moats in data businesses are real. The greenfield here is narrower than the hype suggests—you'll see PE money chasing climate-risk modeling and computer-vision property inspection at the edges, not a wholesale takeover of the analytics layer.

Note: If your strategy deck has a slide about "becoming the data platform for the industry," ask whether you're actually building a moat or just donating R&D spending to a market that's already been won.

6. Reinsurance and alternative capital

Cat bonds, ILS, sidecars—genuinely useful capital efficiency tools, and genuinely attractive to institutional and PE money looking for returns uncorrelated to public markets. But this is also the layer drawing the most direct regulatory heat right now, because it's structurally identical to the arrangement that has state regulators and the NAIC nervous about affiliated reinsurance in the life and annuity world: related-party transactions, opaque asset-liability matching, risk-based capital that may not be pricing the actual risk. These are known trades, under active review.

7. The balance sheet itself

And here, finally, is the piece PE mostly leaves alone: the actual risk-bearing carrier. Full statutory capital requirements, rate filings, market conduct exams, risk-based capital rules that don't care whose name is on the equity. This is the one link where regulatory protection still functions as advertised, and it's not an accident that PE has mostly declined to fight it head-on. Instead, private capital rents access to this layer—through fronting, through reinsurance, through MGA fee arrangements—rather than trying to own and run it directly. The fortress holds. Under siege are the lands that sustain it.

Eighty-four percent of large brokers with over $100 million in revenue now use generative AI. Claude, ChatGPT, and Copilot save producers hours every week on prospecting and reporting. But these tools are trained on public internet data, not insurance data.

That gap matters more in insurance than in almost any other industry. Regulation changes state by state. One wrong sentence in a submission or coverage recommendation can trigger an E&O event. And one in two agency records is missing critical fields, so generic AI often builds a confident answer on an incomplete picture.

The agencies pulling ahead are not the ones using AI first. They are the ones grounding AI in insurance-specific data. This whitepaper shows you what that gap costs and what to look for in a platform built to close it.

Come visit Zywave's website to learn:

• Why Insurance Needs Specialized AI Context
• The Hidden Cost of Incomplete Records
• What Generic AI Gets Wrong About Insurance
• The Difference Between Data and Intelligence
• Four Questions to Ask Before You Trust an AI Platform

See the difference>>

In 2012, the world avoided catastrophe by nine days. 

A solar storm emitted a plasma cloud of gigantic proportions that summer. Had it hit the Earth, it would have done trillions of dollars in damage, wiping out communications and GPS satellites while causing so many transformers to explode that it would have taken down much of the electric grid for weeks, months or even a year. But the plasma cloud passed through the spot in the Earth's orbit where it had been nine days earlier and did almost no damage. Nine days....

The Earth wasn't so fortunate in 1859, when what's now known as a Carrington Event — named after one of the astronomers who saw the solar eruption — wiped out the world's telegraph system. And scientists say such a massive flare will surely hit the Earth again, with its now exponentially more wired and, thus, vulnerable electrical and electronic infrastructure. The only questions are: when, and how devastating will the flare be?

Models suggest that the likelihood of a Carrington Event is only about 1% in any given year, but the effects would be so devastating that we need to at least be running scenarios to see how we can help clients be more resilient while also making sure insurers would survive such a catastrophe.

Let's have a look.

The Carrington Event itself isn't what a reporter would call new news. While the near miss in 2012 received remarkably little coverage, we ran an article in 2014 on the potential danger that holds up quite well, and have been periodically raising the issue since then. This article in 2021 goes deep into the implications for insurers, partly by looking at a solar storm in 1989 that, while far less severe than a Carrington Event, showed in miniature the sort of damage one might cause. A solar storm in 2024 — again, far less severe than a Carrington Event — prompted me to revisit the issue, and, for those looking for a truly deep dive into the history of solar storms and their implications, our colleagues at the International Insurance Society published this piece that same year. As a solar cycle peaked last year, making intense storms more likely, we ran another piece on the potential implications.

What is new is a recent study that found, according to the subhead of a Wired article: "Scientists have long assumed that there’s an upper limit to the intensity of the solar winds that reach Earth. New research suggests there’s not—and the implications are alarming."

The article says:

"For years, scientists have believed that there is a natural limit to the intensity with which Earth responds to the most extreme solar storms. According to that idea, when the solar wind reaches very high values, Earth's magnetic field stops reacting proportionally, and its response enters a kind of 'saturation.' However, this new study suggests that perhaps that limit never existed. What appeared to be a physical phenomenon could, in reality, be an illusion caused by the way the measurements are analyzed."

In other words, our magnetic field may not offer the protection we long thought it did. 

A Carrington Event is still seen as a truly rare occurrence — some models say one will hit Earth every 100 to 250 years, while I've seen models that say such an event is even rarer, coming just once a millennium. But the threat can't be dismissed as a black swan. When one hits, it won't be a total surprise. We know another Carrington Event — or even worse — will occur. So the threat is more of a grey swan, and there's no excuse for not planning at least at some level for when it hits. 

The effects of a massive solar event would be beyond anything we've seen in the modern day. This wouldn't just be an extended blackout. The shutdown of the grid could be so total that we'd lose access to the financial system, meaning no ATMs or credit cards. We would lose much of our communication network because of damage to satellites as well as to the cell towers and WiFi that need electricity. Water purification systems would be down. Farmers would lose access to many needed supplies, endangering our food supply. 

So insurers must help clients map out how they would keep their businesses operating or at least recover quickly — in the U.S., parts of the Atlantic Coast and upper Midwest seem to be most vulnerable, so backup capabilities located elsewhere might be wise. Generators, batteries (though many will be overwhelmed by a Carrington Event's electric surge), water supplies and so on should also be part of a plan, keeping in mind that something that may be routinely available now will be in excruciating demand if the electric grid goes down even for weeks, let alone months.

I think there are technology opportunities here, too. Solar power and batteries are improving so fast and dropping in price so quickly that they are providing unprecedented, decentralized power sources that could allow for considerable resilience even with the grid down for an extended stretch. I don't know just what would be involved in hardening them enough to withstand a massive electric shock, but doing so would greatly diminish the length of the damage from a Carrington-level event. The good news about such a solar storm is that we get about 18 hours of warning — the time between when astronomers spot a solar flare and when the storm hits the Earth — so there is time to switch power sources into bunker mode. 

Insurers obviously also have to protect themselves so they can reliably pay claims and help the world get back on its feet. Insurers should cover as much risk as possible but have to stay in business. So they need to be careful as they underwrite risks that could be beyond measure when the next Carrington Event hits.

I'd suggest also working with government to develop a recovery plan, though I'm not optimistic. I suspect that even a White House that believed in FEMA, as the current administration clearly does not, would be vague and noncommittal. So I suspect that coordination between insurers and governments will largely have to be arranged after the fact — and that the finger-pointing and legal action will be mind-bending.

I hope I don't come across as a boy crying "wolf," having repeatedly raised an alarm for a dozen years that has yet to be borne out. But remember: The fable ends with a wolf appearing and eating all the sheep.

Cheers,

Paul

"Should AI stand in the customer's shoes?"

That idea struck me as fresh the first time I heard it. But the more I sat with it, the less it seemed to mean — if it only means presenting results through language and interfaces users are already familiar with, the payoff is thin.

Looking back at the AI projects I've been part of, I've come to a different conclusion: a model can answer a question correctly and still have no idea what task the person is actually trying to complete — whether the right move now is to keep answering, ask a follow-up, stop, or hand off to a human. For AI to take on real work, not just answer questions, companies need to translate what experienced professionals know about context, risk, timing, and consequence into something a system can actually recognize: signals, task states, response protocols, and validation criteria.

This piece starts from a scenario I know well — model testing — and walks through how that translation happens, and why it forms two decisive thresholds standing between AI and real deployment. AI companies that understand and clear these thresholds are better positioned to build durable capabilities within a vertical domain, turning what they learn from a single deployment into delivery capabilities that are repeatable, verifiable, and scalable.

Beyond model testing: AI still must understand the task

On most AI projects, technical teams build a battery of test prompts and run the model against them repeatedly. Did it identify intent correctly? Did it retrieve and apply the right knowledge? Was the answer stable, and free of hallucination, policy violations, or logical errors? That's the bulk of testing and debugging.

I've watched this play out on my own company's projects many times. In a field like insurance sales, though, the testing team often also includes experienced domain experts sitting alongside the engineers. They read the same conversations and flag the same kinds of issues — but they're looking through a different lens entirely.

The technical team asks first whether the model got the answer right. The domain expert asks something else: does this answer fit where the conversation is right now? Where does it take the customer? And what should happen next — keep answering, probe further, pull back, or bring in a human?

That gap between the two lenses is exactly what most AI projects run into the moment they move from model testing into a live business setting.

The first threshold: recognizing when the task has changed

Say a critical illness policy includes a 90-day waiting period — a clause under which coverage for specified illnesses begins only after a set period, subject to the policy terms. A customer, mid-conversation with a sales assistant, asks: "How long is the waiting period on this critical illness policy?"

The AI checks the policy and answers: "The waiting period is 90 days. A covered illness first diagnosed after that point is eligible for a claim under the policy terms."

Judged purely on product knowledge and wording, nothing is wrong here. The model understood the question, cited the correct clause, and made no promise beyond what the policy covers. By ordinary QA standards, this counts as a clean, successful response.

But an experienced insurance agent may notice something a conventional QA review would miss: why is this customer asking about the waiting period right now?

Maybe they're simply comparing products. Or maybe they just had a checkup, noticed a symptom, or are quietly worried about a health issue that hasn't been diagnosed yet. The same question, asked against a different backdrop, can send the rest of the conversation in an entirely different direction.

What this threshold really tests is whether the AI can register that the task itself has shifted. The system first must notice what's happening: is this still a routine product inquiry, or has the task already shifted toward clarifying health risks, disclosure obligations, and coverage expectations?

If the customer is simply comparing products, stating the waiting period is enough. But if they've already mentioned something like an abnormal test result or a health concern, that same product inquiry has likely moved into different territory.

If the AI, still optimizing for conversion, nudges the customer to buy now so the waiting period starts ticking sooner — every individual line in that exchange might still look defensible on its own, and the short-term conversion numbers might even look good. But measured against the actual job — a sound, compliant sale — the direction has already gone wrong. That drift is hard to catch precisely because each individual output holds up fine under its own scrutiny; the problem only shows up once you look at where the pieces are heading together.

From a human-AI interaction standpoint, the customer opened with a product question, but the task that actually needed handling may have quietly turned into something else mid-conversation. The model answered correctly. It just never noticed the nature of the job had changed.

The second threshold: understanding what the task is actually for

Recognizing that a task has shifted only answers "what's happening right now." The AI still needs to understand what the task is meant to accomplish, and what the next move might set in motion.

This kind of failure is hard to catch through accuracy checks alone. Sometimes the problem is missing sufficient probing before the answer. Sometimes it's an inappropriate action taken after answering.

A customer asking a service rep about refund policy may have already been through several failed attempts and actually need an escalation, not a repeat of the rules. A sales manager asking AI to explain a revenue decline might be heading toward a decision about which customer segments to cut, which frontline process to change, or how to reallocate resources. A customer asking about a specific policy clause might simply have a preference — or might be quietly weighing a risk they haven't said out loud.

Language and knowledge let AI produce a better answer. But taking on real work requires understanding the task at hand: its current state, its intended goal, and the consequences of the next action.

This threshold tests whether AI understands the context and goal behind the task — not just that the task has changed, but where it has moved to, what information is still missing, and what should happen next.

Only with that in place does "AI standing in the customer's shoes" mean anything concrete. A natural tone, a clean interface, a flexible dashboard — these all improve how information gets seen and used. Understanding the task itself is a different matter altogether: it reaches into timing, limits, action, and who's accountable for what.

How professional judgment becomes a system capability

The gap between the engineers and the domain experts isn't really about who knows the technology and who knows the business. The real difference is in the basic unit each side uses to judge whether an output is any good. By "unit," I mean: when you're deciding whether one output is acceptable, what's the first thing you look at? The technical team checks the model against a set of evaluation dimensions — intent recognition, retrieval accuracy, answer quality, safety constraints, output stability. The domain expert puts that same answer back into its full context: when did it show up, where is the customer in the process, and what could it lead to?

So the engineer's first move is to check the answer itself. The domain expert cares more about where that answer sits within the whole job: does it fit the moment, could it be misread, and how does it affect trust, compliance, and the eventual business outcome? That difference in unit of judgment has real consequences for how well human and AI work together. A model can look great in a test report and still keep producing the wrong response once it's live — and the problem often isn't the content of the answer at all. It's that the system never noticed the task had already changed.

A good domain expert can often flag a problem instantly: "That's not something you say here." For an AI team, that sentence is just a starting point. It raises questions that still need answers: what signal did the expert notice? What outcome were they worried about? Which situations are fine to keep going, and which need a follow-up question or a handoff? And where does that knowledge belong — the system prompt, the context, a rule, a workflow, or a hard permission boundary in the system itself?

No matter how sharp someone's professional judgment is, if it stays locked inside their own intuition, it can never become a system capability that scales. So the real question — the one that determines whether a team clears both thresholds — is how to turn expert instinct into structures the system can recognize, execute, and validate.

Bringing domain expertise into an AI system tends to follow a fairly consistent path:

It starts with finding real cases where the model answered correctly but the task still went off course. From there, the team breaks down exactly what signal the expert picked up on, and uses that to define the task state that signal points to. Only then does the team frame a concrete hypothesis about what intervention might help — and only after that does any of it get written into the system, tested against real outcomes, and refined according to the results.

That process has to start from a concrete case. Ask an expert in the abstract what to watch for during a sale, and you'll mostly get principles. Put an actual — or simulated — conversation in front of them and ask where the problem first started, and the instincts they've built up over years, the ones they've never quite put into words, finally have something to attach to.

Take the waiting-period case. A customer mentioning a recent checkup, a physical symptom, asking "is it too late to buy now," or treating the waiting period as some kind of guarantee of a future payout — these are all signals worth flagging. But a signal is just an observation. Plenty of people who ask about a waiting period have no health issue at all, and the model shouldn't jump to conclusions off a single word. The team still must work out under what conditions a given signal indicates the task has shifted.

Once that shift is defined — a routine product inquiry moving into a health-risk and disclosure conversation — the system can change how it responds: instead of pushing forward, it clarifies the risk, and hands off to a human if needed.

Through this process, an expert's instinct gradually turns into signals and task states the system can recognize. Once a task state is defined, the team still has to answer something more fundamental:

For a customer showing a possible health concern, does one extra clarifying question — instead of answering directly and pushing the sale forward — improve the completeness of health disclosure and reduce downstream disputes?

That question still lives at the level of a business hypothesis and an intervention design — it hasn't yet become a prompt or a system rule. It defines who it applies to, what action to test, what to compare it against, and what outcome it's meant to improve. At this point, the team has moved past "what should AI say" and into "does changing what AI does actually change the business result?"

From there, the technical team has to engineer the hypothesis itself. First: which signals should trigger the risk-clarification path. Then: what the system should do once triggered. For example — when a customer asks about the waiting period and also mentions a recent checkup, a symptom, or timing around the purchase, the AI should first understand why they're asking, flag the importance of honest disclosure and underwriting review, avoid implying a claim is guaranteed, and escalate to a human when necessary.

Some of this belongs in the system prompt. Other parts require context, annotation, rules, state management, and workflow support to work together. Where compliance lines, commitment authority, or high-stakes decisions are involved, tighter system-level guardrails are usually required too.

Bringing expert judgment into a system, in other words, can't be done with a single block of prompt text. The team must work out, piece by piece, which layer each bit of knowledge, experience, and constraint belongs in — that is what reliably constrains model behavior and improves the consistency of both outputs and actions.

Once the system design changes, the testing standard has to change with it. In a routine product comparison, the AI can simply state the waiting period. When a customer mentions an abnormal test result and asks whether it's too late to buy, the system needs to recognize a different task state and adjust its handling accordingly. If the customer then asks the AI to help conceal a health issue, it should stop, explain the disclosure requirement, and route the conversation to a qualified human.

And in the end, it all comes back to real outcomes. Did the extra question improve disclosure completeness and cut down on disputes? Did it create friction, interrupting a large share of routine inquiries that never needed it? Did the human handoff genuinely improve customer understanding and underwriting quality?

These questions only get answered against real data over time. It's only once expert judgment enters that loop — test, deploy, measure, refine — that it has a chance to settle into something stable and repeatable.

Closing: human-AI collaboration starts with a shared understanding of the task

What a given project needs to deliver shapes what human-AI collaboration ends up looking like. If a partner cares most about the final sales result, UI/UX may not be the primary focus of the delivery. But where customer acquisition, operational analytics, or direct decision-making by the client's own staff are involved, a dashboard that can flexibly switch segments, adjust workflows, and surface results often matters a great deal.

This piece started from a version of the same question: should AI present its output in language and formats the customer already knows well? That question deserves real weight — information nobody can use, however accurate, never makes it into real work. But push on it a little further, and a different picture emerges. Interface and presentation shape how people understand and use what AI produces. The deeper problem in human-AI collaboration is whether AI understands what task it's actually part of.

The same question can be routine, or it can mark a turn in the task. The same answer, delivered at a different moment or against a different backdrop, can lead to entirely different consequences. AI needs to recognize when the task state has shifted, understand the goal it's currently serving, and know how to act on what comes next.

That kind of understanding — of a vertical domain and its specific operational contexts — does not emerge from the model on its own. It comes from technical teams and domain experts repeatedly deconstructing real cases together, turning the signals, context, responses, and risks scattered across professional experience into something a system can recognize, act on, and verify.

Right now, most AI projects ask domain experts to verify whether answers are correct — whether a particular line should have been said or whether a number is accurate. That understates the value they bring. What they really contribute is task understanding and professional judgment: how a task shifts, when intervention is needed, what form that intervention should take, and which outcomes cannot be assessed through short-term metrics alone. The technical team must then decompose that expertise and encode it at the appropriate layers of the system — the system prompt, context, rules, workflows, permission boundaries, and the broader mechanisms governing human–AI interaction.

Human-AI collaboration, then, is a task-centered partnership. People define the goals, constraints, and accountability, and ultimately own the outcome; AI participates in analysis, inquiry, judgment, and action. Organizations then refine both the division of labor and the system design in response to real-world results.

For an AI company, this is also the real barrier to going deep in any vertical. General-purpose models and technical tools are becoming widely available, but the domain expertise built on a single deployment doesn't turn into scalable capability on its own. That experience becomes scalable only when an AI company can continually translate domain expertise into task states, system mechanisms, and a functioning validation loop. Only then can a single success be reliably reproduced across more clients and more scenarios.

Getting the answer right only proves the model understood the question in front of it. Catching a shift in the task the moment it happens, understanding what comes next, and knowing when to answer, ask, stop, or hand the task back to a person — that's when AI starts doing the job.

Artificial intelligence (AI) has rapidly become one of the most important global business topics and risks. This is clearly reflected in the Allianz Risk Barometer, where AI rose from 10th place worldwide in 2025 to 2nd place in 2026 – the largest increase of any risk category. Across regions, AI now ranks among the most relevant risks, highlighting the speed at which it has moved up the corporate risk agenda.

A gap between perceived risk and current losses

From a claims perspective, however, the current loss situation appears relatively calm. The number of clearly confirmed AI-related losses remains limited, and if assessed purely on today's claims data, AI would not yet rank among the most significant loss drivers. This apparent gap between perceived risk and actual losses requires closer attention.

A large "grey area" exists in which AI involvement is difficult, if not impossible, to prove. In many incidents – whether cyber-attacks, business decisions or operational failures – AI is likely to play a role, even though it cannot be conclusively identified. As a result, current claims data may underestimate the true level of underlying risk.

AI risk emerging across multiple lines of business

What is already evident is that AI-related exposures are emerging across multiple lines of business simultaneously, but in very different forms. In cyber, AI is increasingly assumed to be used in areas such as phishing or malware development, even if its involvement cannot always be verified. At the same time, intellectual property disputes related to the training of large language models are gaining prominence, alongside data protection issues where AI is deployed without sufficient safeguards.

In liability insurance, risks are particularly relevant in connection with autonomous systems and AI-supported decision-making. One of the most visible areas currently is directors & officers insurance, where so-called "AI washing" – the overstatement or misrepresentation of AI capabilities – is already leading to claims. In addition, AI is significantly enhancing the effectiveness of fraud schemes, from business email compromise to advanced social engineering and voice-based attacks.

AI is not a risk that is limited solely to certain business areas and applications. It is emerging across many insurance products simultaneously – often not as the direct cause, but as an aggravating factor.

Why losses are expected to increase

Looking ahead, there are strong indications that AI-related losses will increase substantially. AI acts as a powerful risk accelerator, with the number of incidents and related litigation rising globally. External observations point to a clear upward trend.

What sets AI apart from many traditional risks is its probabilistic nature. Errors are not exceptions but an inherent feature of the technology, even when systems are properly designed and implemented.

When such systems are deployed at scale and embedded into everyday processes, even a small error rate can translate into a significant number of incidents within a short period of time. Combined with the speed of deployment, evolving governance frameworks and the widespread use of AI tools across organizations, this creates a risk environment in which losses can escalate rapidly once they begin to materialize.

AI fundamentally changes how risk behaves. Unexpected output and errors are not an exception; they are inevitable. When not addressed and managed properly, even small deviations can lead to significant losses quickly.

Understanding the roots of AI risk

At its core, AI risk operates on several levels. One lies within the technology itself. AI systems often generate outputs with a high degree of apparent confidence while remaining sensitive to small changes in input and prone to various forms of bias. This can have significant implications in areas such as decision-making, risk assessment or negotiation.

A second level arises from the way AI is implemented within organizations. While experience with AI continues to grow, many companies are still developing the necessary governance, controls and expertise, often under considerable pressure to adopt the technology quickly. Even under responsible use, the probabilistic nature of AI means that errors cannot be fully eliminated.

A third level relates to indirect effects. In many cases, AI is not the direct cause of damage, but rather a force multiplier that enhances existing risks by making attacks more convincing and fraud more effective.

Implications for risk management

Against this backdrop, it becomes clear that AI risk is not purely a technological issue. It is a business, governance and leadership challenge that requires a holistic approach. The fact that losses remain comparatively low today should not lead to complacency. Historically, losses tend to lag behind adoption – but once they emerge, they can develop quickly and at scale.

Effective risk management therefore starts with a clear understanding of the fundamental characteristics of AI and their translation into concrete business risks. Organizations need to make deliberate decisions about which risks can be mitigated, which can be accepted, and which should be transferred. At the same time, the opportunities offered by AI remain significant. Realizing these opportunities sustainably depends on addressing risks in a transparent and proactive manner.

AI offers enormous opportunities, but those opportunities can only be realized sustainably if risks are understood, addressed early and managed proactively.

A risk that is only just beginning to materialize

As AI continues to be embedded more deeply into business operations, its effect on the risk landscape is expected to grow further. What is observable today is still only an early stage of a broader transformation – one in which the true scale of AI-related risk will become more visible over time.

My son Riley burned to death pinned inside his Honda Civic in a side underride crash in 2015. His death was part of a documented crash pattern that the trucking industry and federal regulators have studied, debated, and ignored since the 1960s.

Side underride crashes happen when a passenger vehicle, cyclist, pedestrian, or motorcyclist goes under the side of a semitrailer. The consequences are often devastating because the trailer frame bypasses the vehicle's safety systems and strikes the occupant space directly. Riley's collision occurred at a shallow angle and a relative speed of about 14.5 mph, lower than speeds at which tested side-guard designs had prevented underride.

The litigation that followed showed how a fatal truck crash can affect the entire transportation chain. The truck driver's insurer and motor carrier settled early. A motor carrier that had served as an unlicensed broker settled after a few months of discovery. The lawsuit against Utility Trailer Manufacturing, a major refrigerated-trailer manufacturer, went to a jury.

Riley's crash was one of roughly 300 deaths and 400 serious injuries that side underride causes each year, with an estimated societal cost of $7.4 billion. In 2023, 5,472 people were killed in crashes involving large trucks, about 30% more than a decade earlier. Seventy percent of those killed were occupants of other vehicles; another 12% were pedestrians, cyclists, or other nonoccupants.

Mistakes happen. Side underride guards, like seat belts and airbags, reduce the severity of consequences regardless of fault.

The cost objection

NHTSA has denied four formal petitions seeking side guards on semitrailers since 1979. The 1979 denial (FR 44:55077-55078) closed with: "If the evidence gathered by the agency indicates that side underride rulemaking could contribute significantly to safety, the agency will commence rulemaking."

The trucking industry's cost objection over whether to require side underride guards on semitrailers has a long history but a thin evidentiary base. In 1969, the industry's trade association told federal regulators that side underride "is not a hazard that warrants some kind of regulatory remedial action."

In 1991, NHTSA declared without evidence that side underride guards were not cost-effective. A 2021 Freedom of Information Act request confirmed the agency held no data or analysis to support that assertion. The industry clings to the same claim in regulatory comments and litigation: guards would be too expensive, add weight, and increase fuel use. Juries are unconvinced. 

I tested those assertions.

In my 2026 peer-reviewed cost-benefit analysis, side underride guards produced positive net economic benefits under every modeled scenario, ranging from roughly $137 million to $2.8 billion.

The break-even threshold in my analysis is that modest: guards need to prevent as few as 24 deaths and serious injuries per year to be economically justified. The analysis also excluded litigation, settlements, defense costs, and insurance losses. A side guard with an aerodynamic skirt costs approximately $2,500 per trailer.

For insurers and fleets that bear the litigation costs, the actual break-even is lower still.

What the record shows

Discovery in Riley's case showed that Utility Trailer had known for decades that side underride was a serious hazard but had not designed, built, or tested its own side guard before the lawsuit. The company criticized the Angelwing, a side guard that an independent inventor had already designed, built, and crash-tested without the engineering resources of a major manufacturer, and made no attempt to develop its own solution. Two years before the verdict, IIHS independently crash-tested a side guard at 40 mph and confirmed it prevented underride. Wabash National successfully crash-tested its own design the same year. The jury found Utility negligent and returned an $18.9 million verdict. After the verdict, Utility publicly maintained that it had not negligently designed its semitrailers and argued that side guards would make them more dangerous.

Then, after arguing at trial that side underride guards were impractical, Utility developed and patented one. The company now offers a Side Impact Guard, describing it on its website as evidence of a "long-standing commitment to the safety of every motorist sharing the roadway" and claiming its innovations "set the gold standard for performance and reliability." As of 2026, Utility has delivered nearly 70 semitrailers factory-equipped with the guard.

The technology was never the obstacle.

Nearly every major semitrailer manufacturer now holds a side guard patent. The impracticality argument was an industry position, one the industry's own trade association funded specifically to develop defense strategies against side underride lawsuits, with manufacturers suppressing information from the public under a joint defense agreement.

The costs of inaction

The Utility verdict was not a one-off. Vehicle side underride cases have produced $32.5 million in Georgia for a fatality, $26.1 million in Iowa for a debilitating injury, and $13 million in a second Georgia case. Pedestrian and cyclist cases have produced an $81 million verdict in Utah, $11.5 million in Virginia for a bicyclist, and $8.5 million in New Hampshire for a bicyclist fatality, among others. Many more cases resolve confidentially. These are the hidden costs the industry never mentions when it argues against requiring guards.

The pattern is consistent. When side underride is not addressed before a crash, the costs appear afterward in verdicts, settlements, defense expenses, and rising insurance premiums. The math does not favor waiting.

Those verdicts trace back to a specific, identified crash mechanism with a documented fix. Side underride is not an anomaly. It follows from a design decision, the absence of a guard, that is identifiable, verifiable, and correctable. Insurers are not pricing it as one. Commercial auto insurance has posted 14 consecutive years of underwriting losses. Raising rates does not fix that. Preventing fatal and serious injury from the collisions that generate the claims does.

The insurance decision

Fleets already get underwriting scrutiny for driver safety programs, telematics, dash cameras, maintenance practices, and loss history. Side underride protection is a documented liability exposure that belongs in that same conversation.

Insurers are not encouraging side guard adoption. They should be, through underwriting questions, premium credits, or better retention terms.

Testifying before the Senate Commerce Committee, the American Trucking Associations asked Congress in June 2026 to preempt state tort liability for vehicles that meet federal safety standards, while opposing the side guard mandate that would create the standard preemption requires. The industry wants the litigation exposure eliminated, but opposes the only mechanism that would justify the change. In the meantime, fleets and their insurers absorb the verdicts.

None of this requires waiting for an NHTSA side guard rule. Regulation is not a prerequisite for action.

The Supreme Court recently underscored a related point in a freight-broker liability case, Montgomery v. Caribe Transport II (2026). The unanimous decision recognized that common-law duties and standards of care are part of state safety authority. Justice Kavanaugh wrote that "truck safety is a matter of life and death," and that "Not all truck accidents can be prevented. But some can." He also acknowledged that litigation and insurance costs can be significant even when brokers prevail. The lesson is broader: truck safety decisions become liability decisions.

The question for fleets, insurers, brokers, shippers, and manufacturers is not whether federal regulations will eventually require side guards. It is why the industry keeps defending inaction in court instead of addressing a known hazard. The verdicts have been answering that for years, and the amounts keep adding up.