Download

AI Can Preserve Claims Adjuster Expertise

With fewer adjusters entering the field and claims growing more complex, AI offers capacity without replacing the human judgment insurers need most.

Insurance

I spent years at Meta helping disaster response organizations like the Red Cross put data to work to help others. After a hurricane or a wildfire, the hard question was never whether people needed help. The hard question was where or how much and how soon. I came to insurance because the property claim is where the hard questions get answered, one family at a time. And because the people who answer those questions are getting harder to find.

Fewer adjusters, harder claims

Every year about 21,600 claims adjuster and examiner jobs become available in the United States due to attrition, according to the U.S. Bureau of Labor Statistics. Meanwhile, colleges only graduate about 2,100 students a year in insurance and risk management, and that number is falling. That gap roughly translates to about 10 open claims roles for every graduate, assuming the unlikely reality that every graduate chooses claims.

BLS expects about 6% fewer of these claims jobs by 2035, partly due to the implementation of software and AI, but those tech-driven efficiencies do not solve the problem. There are still nowhere near enough people to fill the claims jobs that remain open.

Compounding the issue is the fact that the work is getting harder. Swiss Re Institute reports wildfires, storms, and floods drove a record 92% of global insured losses in 2025. Each of those events represents thousands of files a limited number of claims professionals must manage at once, and the more complex cases need an experienced hand. Nobody gives a new adjuster a complex loss in her first year. That judgment takes years to build, and it walks out the door when she retires.

Insurers can't hire their way out of this situation. Thus, the approach has to change by preserving the expertise that exists and pointing it where it matters most.

What AI is for

The strongest argument for AI in claims does not involve replacing adjusters. It is that AI gives them capacity.

Most of the AI conversation in claims is about paperwork: summarizing files, finding documents, flagging gaps. Useful, but small tasks. The larger opportunity is the estimate itself: measuring the room, identifying materials, scoping the damage, building the line items. The work is repetitive. It takes hours on every file, and it is harder than it looks because each line item has to match the carrier's guidelines. Repetitive, slow and rule-bound. These are exactly the types of tasks for which automation was built. Computer vision and AI can now produce a first draft of the estimate from a handful of interior photos, at the fidelity of a veteran adjuster, in hours versus weeks. Doing so lifts the cognitive load from the adjuster, thus freeing her for the human side of the claim while also leveraging her expertise and judgment.

What technology cannot do, and should not do, is make final decisions. Is the claim covered? Would this rule apply? What does this family need to hear from the insurer today? That has always been the adjuster's job and must remain so.

It also keeps knowledge in the building. A first-year adjuster who reviews a fully scoped complex loss learns how one is built, instead of creating their own, one file at a time.

Where the caution is right

The industry is right to be careful, and careful should mean a couple of things.

Accuracy comes first. In an estimate, the costliest mistake is rarely the measurement. It is a line item: the wrong one chosen, or the right one missing. Replace the flooring but forget to detach and reset the vanity, and the contractor sends a supplement. Include a line the carrier's guidelines don't allow, and the file gets flagged. One shortchanges the policyholder. The other leaks money. Both happen today in files no AI ever touched.

So, the right benchmark for AI tools should not be an imaginary process where people and legacy systems never make mistakes. It should be how claims run today, with controls that match the consequences of an error.

Second, every line item must be traceable, and every gap has to be visible. For each line, the adjuster should see the photo showing the damage behind it and the guideline it must satisfy. So should the auditor, the reinsurer and, if it comes to it, the judge. Claims executives ask me one question more than any other: will the number hold up in an audit, a complaint review, a dispute, or in front of a regulator? Traceability helps ensure the answer remains yes. An AI that cannot show its work does not belong in a claim file. Glass box, not black box. In claims, defensibility comes first, followed by speed.

Build around both

The choice on how to solve the issues facing claims is not between people or AI. There simply are not enough experienced adjusters to run claims the old way. At the same time, AI cannot carry the judgment, accountability or empathy a claim requires. Nor should it.

The model I believe in has four rules: 

One: AI does the repetitive, rule-bound work and drafts the estimate. 

Two: adjusters make every consequential decision, i.e., coverage, cause of loss, and what the policyholder is owed. 

Three: experienced adjusters train, test, and govern the AI. 

Four: success is measured in accuracy, resolution quality and adjuster capacity, not in the share of tasks automated.

Careful is not the same as slow. The talent gap is not waiting for perfect technology. Neither is the family standing in a flooded kitchen, nor the thousands like them in week six after a CAT, when the queue is still deep. The bellwethers will not be the carriers that automate the most. The real leaders will be the ones that empower adjusters to do their best work: look at the damage, decide what is fair and tell the policyholder so, quickly and accurately. The moment after loss should be the start of rebuilding, not the start of waiting.

AI Apocalypse? Don't Get Distracted

We're suddenly having a debate about whether AI is about to kill us all, and it obscures some pressing issues.

Image
Insurance

While we've suddenly landed in the middle of a debate about whether AI may be about to obliterate the human race, I hark back to a profile I did for the Wall Street Journal about a brilliant AI and robotics researcher from Carnegie Mellon named Hans Moravec.

The focus was his provocative idea that humans would be able to download their brains — their entire consciousness, their full personality, an exact replica of them — into computers, which could then teleport to any spot in the universe or spawn an infinite number of what Moravec called "mind children." 

The memorable headline was:

Good News: You

Can Live Forever;

Bad News: No Sex

I asked Moravec how long it would take for his vision to be realized. "Oh, a long time," he said. "Maybe 25 years."

That was 35 years ago.

So I'm not going to worry much for years about all the talk of impending doom. Timelines on sci-fi-like change tend to be way, way off. But, under the radar, there are plenty of AI issues that should be major concerns right now, including for insurers.

Let's have a look.

I'll start with Bill Gates's recent manifesto, which, among other potential dangers from AI, called out the prospect that AI will supercharge the work of malign actors, perhaps leading to bioterrorism, massive cyberattacks, and more. While we can discuss the potential long-term threats to humanity from AI, these are the kinds of threats I think we need to focus on today. These threats are already being pursued, whether by individuals looking to extort massive amounts of money or by nations looking for weapons in an increasingly belligerent world, and AI clearly provides exponentially more computing capability.

The MIT Technology Review goes into detail about how AI might produce a devastating bioterrorism attack: "A bioweapon might be a highly lethal virus that targets people according to their genes. It could be a fungus that wipes out a crop and causes food insecurity. Perhaps it would be a tasteless, odorless toxin that could be slipped into a region’s water supply, undetected....

"Today, AI bots can answer questions on topics spanning all realms of science. Anyone can use large language models trained on the knowledge and experience of 'almost every scientist who ever lived on this planet,' says Dunja Sabra, a biosecurity researcher at the University of Hamburg in Germany. Those models can provide instructions and video training on how to conduct experiments.

"Combine that with advances in biotech that have made gene editing and synthetic biology tools much more accessible (the “DIY biology” movement has already enabled many people to set up labs at home), and you’ve got a potentially very dangerous situation."

Wired, meanwhile, warns about all the vulnerabilities in software that AI bots are finding, and it's not hard to imagine how those weaknesses could be exploited. In July, hackers, thought to be based in Iran, disrupted 30 municipal water systems in Minnesota, and you can be sure Iran will ramp up attacks as fast as it can. North Korea, China, Russia, and other countries could stage similar, small attacks or could even try to shut down electric grids and stall commerce by using weapons of not-quite war.

Cyber attacks could easily lead to massive business interruptions of the sort insurers routinely cover and could increase geopolitical risks of every flavor.

Businesses and governments understand that the bots are making them vulnerable and are working as fast as they can to plug the holes, but they won't find all the holes, at least not right away. And you be sure that some hacker cartel or foreign government is storing up what are known as "zero day" vulnerabilities that can be unleashed on unsuspecting businesses and societies. 

There is also massive potential for operator error as AI is deployed more broadly. For instance, the plan to use AI in air traffic control, just now going live for DC-area airports, strikes me as an accident waiting to happen. I hope everything goes smoothly, but the potential for trouble is so great that insurers and everyone else should be wary. Air traffic control hasn't exactly acquitted itself well lately, and AI tends to amplify flaws by making everything happen faster. 

That's where I think the focus should be over at least the next couple of years, both for insurers and for society writ large: on the potential bio, crypto and other deliberate, organized attacks that AI makes possible from bad actors who would profit from those attacks, as well as on the potential for catastrophe as AI gets more involved in mission-critical efforts. 

Yes, there is always the possibility that the search for superintelligence could create an AI that will go rogue and do indescribable damage to the whole human race, for no apparent reason, so government officials should be erecting guardrails.

But business is already circumscribing what AI can do. Insurers certainly are. They've realized that justifying a decision with "the AI says so" won't fly, so they're requiring that every decision be explainable and are greatly limiting what actions an AI can take without explicit human permission.

The whole superintelligence debate has so many dimensions even beyond the technical ones — there are political elements, issues related to business models, massive public relations concerns, etc. Here, for instance, is a column in the WSJ that argues the whole apocalypse debate is an attempt by AI's leading developers to duck responsibility. 

The issue reminds me of Winston Churchill's description of the Soviet Union after it allied with Nazi Germany in 1939: "a riddle wrapped in a mystery inside an enigma." And I'm supposed to understand the flow of technology revolutions, having followed them for decades.

Fortunately, I think we can wait to puzzle out all the implications of this AI revolution — as long as we don't take our eye off the ball on the imminent threats it creates.

Cheers,

Paul

 

 

A Water Loss Prevention Program That Truly Works

LeakBot's 65,000 years of US underwriting data proves actuarially relevant loss mitigation: mature programs cut non-weather mains-related water claims ~60%, and the $5/month all-in model delivers positive ROI across market segments.

Insurance

The Problem

Non-flood damage from water costs homeowners billions a year, and insurers pay $15 billion in claims, just in the U.S. One in 60 U.S. homes suffers water damage each year, and the average claim is nearly $14,000—which doesn’t even include the deductible the policyholder pays, or the huge hassle they go through. Typical solutions involve expensive shutoff valves or a host of small sensors placed throughout a home, but providers have struggled to make a compelling case that their programs deliver an ROI.

The Solution

LeakBot takes a different approach, one that begins before a leak occurs and stretches to a solution. The approach starts with a LeakBot sensor that policyholders clip to the main water pipe in the house. The sensor detects any leaks anywhere in the house that exceed a teaspoon per minute. LeakBot’s app walks a policyholder through steps to find leaks, then, if necessary, sends one of its plumbers to visit the house to find any remaining ones, at no charge. The LeakBot plumber documents the repair, giving carriers visual validation of every claim saved.

The approach finds even tiny, hidden leaks weeks or months before they can cause significant damage. In 2025, LeakBot completed 7,000 leak repairs, 1,470 of which revealed water damage that had already begun.

 

More than two dozen carriers provide the LeakBot solution for free, paying LeakBot $5 a month per house, all in. LeakBot reduces the number of water loss claims by 60%, generating a significant ROI for its carrier partners.

LeakBot covers the full loop, detection through documented repair, which makes it the only true end-to-end platform in this space.

The Documentation

LeakBot has produced a white paper that clears the actuarial bar for detail and reliability. The paper draws on hundreds of thousands of device-exposure years. It uses cohort-level (not anecdotal) evidence. It relies on auditable repair records — rather than programs still running on pilot-stage promises.

-

Download the 2026 Leakbot White Paper

 

Sponsored by LeakBot


LeakBot

Profile picture for user LeakBot

LeakBot

LeakBot is the only end-to-end IoT solution protecting homes from water damage—one that begins with leak detection and can end with a free repair. Backed by more than 10 years in business and 29 patents, a single self-installed device clips onto the home's main water supply line, monitoring water usage and detecting micro-leaks as small as one teaspoon per minute. When a leak is detected, the homeowner can book an appointment for LeakBot's trained employee plumbers to visit, locate, and repair it using specialty equipment — at no additional cost to carrier or homeowner. Homeowners consistently recognize LeakBot's value, reflected in a Net Promoter Score of 82/100 and Customer Satisfaction Rating of 4.9/5. That impact is especially powerful when a hidden micro-leak is found and fixed before it becomes a claim—or a costly plumber bill. That's #PredictAndPrevent in action.

Please connect with us: 

 

Boards Struggle With Information Overload

Directors face mounting fiduciary pressure as board packs balloon while preparation time shrinks, creating a governance crisis few acknowledge.

Insurance

Trust in business leaders is eroding. The 2025 Edelman Trust Barometer reports that 70% of respondents, an all-time high, believe that business leaders purposely mislead people. 

A governance crisis is quietly unfolding in boardrooms globally, yet almost nobody is talking about it.

Board packs are getting bigger. Regulatory expectations are rising. The sheer volume of data directors are expected to review and absorb before going into a meeting and holding the executive team to account is rapidly increasing. And the time between the board pack landing and the board meeting is shrinking. For non-executive and independent directors, who are contracted to work, in many cases, for two days per month, this creates unnecessary pressures.

For executive directors, board meeting prep is one item in a diary that is already full. They usually have a team around them: someone to prepare the data, someone to prepare the board pack, someone to circulate. For non-executive and independent directors, the pack arrives with these individuals a few steps removed from the day to day operations, often late at night, with little time to truly scrutinize.

The expectation from both sides of the table, and the regulator, is that every director arrives prepared: knowing the numbers, assessing the risks, identifying signals for consumer harm.

That expectation is a legal one. A regulatory one.

The Fiduciary Duty Problem

Directors owe a duty of care. Courts and regulators have consistently held that this duty requires directors to be informed, which means to read the papers, understand the material, and bring genuine challenge and oversight. In the UK, the FRC and the FCA set clear expectations for financial services firms. Regulators, quite rightly, look for evidence of challenge, of oversight, of holding the executive to account. They look through the lens: if it's not in the minutes, it didn't happen.

But there is a real tension at play, which few people talk about, save for a few grumblings when the board arrives at 22:35, three days before the board meeting. Board packs follow a structure, a format, and have grown in size and complexity. They are designed to give a complete view of the business at 30,000 feet. They are designed to ensure the executive team has shared what the board needs to see and know, regardless of the density or complexity. Or the timeline.

The result? Board packs regularly run into hundreds of pages, with several appendices in different formats. Executive summaries that are anything but a summary of that section. Complex regulatory updates that require a legal degree to interpret. Risk registers built by someone who never uses them, let alone has to interpret them.

Directors are expected to read all of this, understand this in detail, then challenge the executive during a three-hour meeting. A meeting with 15 agenda items giving 12 minutes per item. It's unrealistic for this expectation to be met. Not because directors are incapable or negligent, but because the entire board process makes this difficult.

What Directors Actually Say

During March this year, I surveyed 35 non-executive and executive directors across the UK, Gibraltar, Malta, the Netherlands, and Germany. The feedback was strikingly similar:

  • 67% of directors said board packs had increased significantly during the last five years
  • 74% of directors said they don't feel they have adequate time to review board materials fully before meetings.

For the insurance industry whose regulators insist that firms have a majority of non-executive directors on their boards, this should be a wake-up call.

What I found more interesting is the gap between the use of AI in directors' professional roles and its near-complete absence from risk or governance. 84% of directors I spoke to are already using commercially available AI tools in their day-to-day work. This tech isn't foreign to them; it's part of their daily routine now.

Yet in the background, a much more concerning number is at play. 95% of directors confirmed data security is a critical factor. And yet many directors are uploading market-sensitive financial results, M&A discussions, regulatory correspondence, and people matters, adding these data points to a commercial AI model, with the data stored on a server outside of their control.

There is also a real concern on the accuracy of AI. 68% of directors said that accuracy is critical, yet several were unaware of the phrase "AI hallucinations," let alone what this means.

What was clear during my research is this. Directors are not opposed to using AI in the boardroom, the opposite in fact. What they are concerned about is data sovereignty and accuracy.

The key question is: Can I trust AI with sensitive materials and data?

What Would Actually Help

The interviews surfaced two critical points that will answer the question:

  1. Data sovereignty is key. Any AI solution must protect sensitive, often legally privileged information.
  2. AI has to get things right. A wrong answer is a non-negotiable.

The insurance sector has yet to answer these seriously.

The Practical Case for Change

Tech must allow board directors to become better equipped to enter the boardroom knowing the risks, knowing the data, knowing where the next crisis is unfolding. Tech can solve this.

The real risk is that directors arrive in the boardroom less prepared and more exposed. The regulatory and fiduciary duty for a director who has missed a conduct risk, who hasn't spotted complaints data going in the wrong direction, or who has failed to see a change in the business that requires regulatory approval is real. A governance risk is accumulating at each board meeting.

AI is here to stay. Used properly, it means directors walk into a boardroom prepared. Not by giving a generalized AI summary like most board portals do, but something more useful: allowing directors to interrogate the board pack. To ask questions with answers grounded in the pack itself. To identify pattern changes across historic board packs and meetings. All within a secure environment where sensitive data is safe, and not sitting on a third-party server forever.

This isn't about replacing directors. It's about equipping directors to fulfill their fiduciary responsibilities.

The Question Boards Should Be Asking

The key question that should be put on the table at the next board meeting is this:

Are we confident that our current process for board preparation enables the directors to exercise their fiduciary duties?

If the answer is no, the next question should be:

What tools exist today that can change this?

These tools do exist. Boards that explore and adopt these are the ones that enable directors to do what they are best at: holding the executive to account and ensuring good outcomes for customers.


Andy Wright

Profile picture for user AndyWrightConseleo

Andy Wright

Andy Wright is the co-founder and CEO of Conseleo, a Board tech and board intelligence platform built for regulated industries, including insurance. He previously served as a senior manager within Tesla's European entity, with approval from the MFSA and FCA, and most recently as managing director of Zego Insurance Ltd. Following regulatory approval, he rejoined the Zego board as a non-executive director. In addition to Conseleo, he continues to run his insurance consulting firm, Resnova, specializing in insurance strategy and capacity placement for insurtechs and MGAs.

Insurance Claims Require Verifiable AI Trails

AI-generated claim summaries gain investigator trust only when findings trace back to original evidence and survive expert scrutiny.

Insurance

Only 16% of claims professionals say they have medium or high trust in AI-generated outputs used without expert review. That figure rose to 60% when an expert checked the output. For an insurance investigation, confidence depends on whether each finding leads back to the original evidence.

Complex claims are rarely contained in one neat document. An investigator may need to connect a photograph with a medical record or compare documents written months apart. A retrieval tool can pull out the right passages and still miss the connection between them. That can produce a convincing summary that the investigator can't even check against the original file.

A claim file is a chain of evidence

Claim files are assembled by different people for different purposes. A clinician records treatment, while an adjuster documents decisions made months later. Because neither record was created to explain the other, the investigator has to reconstruct the missing narrative and identify where the accounts diverge.

Some of the most dangerous failures in an AI-assisted claims workflow are invisible because the application can remain fast and responsive even when a scanned form has been misread, or a table has been separated from the note that explains it. By the time the information reaches the language model, the context has already been lost, yet the model will still produce a polished answer from the partial record it received.

Insurers need to examine how an answer was assembled before relying on it. A claim chronology should preserve which record supplied each event and make conflicts visible. If two documents give different dates, the system should surface that disagreement and leave the resolution to the investigator.

A finding has to survive scrutiny

The risks increase once an AI-generated finding prompts an escalation or enters a legal proceeding. In recent California insurance litigation, lawyers representing State Farm acknowledged that AI-assisted filings contained seven nonexistent case citations. The incident concerned legal research, but it exposed the same weakness that can affect a claim review by showing how fluent language makes unsupported information appear settled.

A usable finding needs an evidence trail that another professional can follow. The investigator should be able to open the cited page, read the surrounding material, and see whether the finding comes directly from the record or has been inferred across documents. Counsel can then examine the same evidence if the claim progresses.

A citation can still point to a page the system has misunderstood. Stronger systems therefore corroborate important findings in more than one way. A language model may identify an inconsistency, while a structured chronology checks whether the dates support it. When those views disagree, the investigator has a clear reason to look closer instead of receiving a tidier but less reliable answer.

Human oversight only works when the reviewer can challenge the output. The professional needs the original evidence and enough context to decide what deserves further investigation; otherwise, approval becomes little more than a rubber stamp.

Judge AI by the investigation it supports

Claim files can include medical histories and other information that carries legal and ethical obligations long after the review. HHS guidance says a cloud provider that stores or processes electronic protected health information can be a HIPAA business associate even when it cannot view encrypted data. Leaders need a clear account of where claim information is processed and who can access it before a file enters an AI workflow. The deployment decision is part of the risk assessment.

Testing a system against completed claims that experienced investigators already understand shows whether it finds the evidence that changed the case and whether its citations lead to the right place. The misses deserve close attention because overlooking an administrative detail is very different from missing a medical record that challenges the chronology.

The people who will use the system should be part of that test because they can judge whether the output helps them form a line of inquiry or simply gives them another summary to check. Their involvement also shows whether the tool fits the way investigations move between an SIU and counsel.

Speed remains useful, but it belongs in that wider picture. The meaningful gain is the time saved between opening a large file and reaching a decision that can be explained and defended. When every important finding remains connected to its evidence, AI can shorten the search without weakening the investigation.

Insurers Should Rethink Outsourcing of Data Work

Insurers outsourced data integration, but managed services often trade speed and control for convenience that quietly becomes dependency.

Insurance

When insurers first adopted fully managed data integration, the pitch was simple: hand over the technical complexity, and focus on the business. For years, that trade made sense. Integration work was specialized and hard to staff for, and outsourcing looked like a clean efficiency gain.

But convenience has a way of turning into dependency over time, and dependency tends to stay invisible right up until the moment a business actually needs to move fast.

Take something as ordinary as onboarding a new employer group, or updating a data flow to reflect a new product rule. In a fully managed model, this rarely stays routine for long. It becomes a ticket, then a statement of work, then a wait measured in weeks or months. One major group insurance carrier recently found that executing a single employer group statement of work took an average of 56 days before onboarding of that employer group even began. Not because the technical change itself was complex, but because the business rules governing it lived in a vendor's queue rather than in the carrier's own hands.

That's the cost worth examining closely: not what shows up on the invoice, but what the invoice doesn't show – how much of a company's own operational agility has quietly moved along with the technical work it outsourced.

When Speed Belongs to Someone Else

Every managed integration contract answers a question most companies never ask directly: who gets to move at the speed of the business, and who has to move at the speed of a vendor's backlog?

In a fully managed model, the logic that actually runs the business, the mappings, the transformation rules, the exceptions built up over years of institutional learning, typically lives inside the vendor's systems, in the vendor's tools, understood mainly by the vendor's staff. That works fine when nothing needs to change quickly. It stops working the moment something does: a new regulatory requirement, a partner who needs to be live before a deadline unrelated to anyone's ticket queue.

"Fully managed" often ends up meaning "fully dependent." And dependency isn't really a technology problem. It's a business one.

Knowledge That Lives in Someone Else's Code

There's a second risk that tends to surface only when a relationship changes, whether a vendor gets acquired, a contract comes up for renewal, or a company decides it's time to modernize.

Over years of operation, a great deal of institutional knowledge accumulates inside integration logic: how a specific partner's data should be read, which exceptions matter, or why a rule was written the way it was in the first place. When that knowledge lives entirely within a vendor's proprietary configuration, it isn't really the company's knowledge anymore; it's on loan.

This isn't an abstract concern about vendor lock-in. It's a concrete question: if a company needed to walk away from a managed integration relationship tomorrow, could it actually reconstruct the rules governing its own data? Or would that knowledge have to be rebuilt from scratch, at real cost, because it was never fully owned to begin with?

A Simple Test

Here's a question worth asking honestly, regardless of how an organization currently handles integration: if a business stakeholder needed to onboard a partner or change a rule this week, could the internal team actually do it, or would it have to file a ticket and wait?

For many organizations, the honest answer is the second one. That isn't automatically a crisis. But it's a signal worth sitting with because it says something about whether integration is functioning as a strategic capability or as an outsourced dependency that still runs smoothly enough not to notice.

Before the Next Renewal

A few questions tend to surface the real state of operational control, and they're worth asking before signing the next managed-services renewal or kicking off a modernization effort. Does the internal team have real visibility into the business rules governing its own data, or only a vendor's summary of them? If the relationship ended tomorrow, what would it take to rebuild that logic independently? Is the current turnaround time for a routine change driven by genuine complexity, or by contractual process? And who within the organization actually understands, end-to-end, how its data moves today?

None of these questions have one right answer. Some organizations will look at their scale and resources and conclude that a managed model still makes sense, with trade-offs included. Others will realize that the convenience they signed up for years ago has quietly become a limit on how fast they can move today.

Either way, it's a decision worth making deliberately, as a question of operational control rather than simply cost or convenience. The companies that move fastest over the next few years probably won't be the ones with the most sophisticated integration technology. They'll be the ones who know, with certainty, whether they actually own the rules that run their business, or whether someone else is running the show.

The Next Frontier in Healthcare Risk

Digital musculoskeletal care has proven its market value, but the next frontier is using AI to measure functional decline before it becomes disability.

Insurance

Over the past several years, digital musculoskeletal (MSK) care has evolved from an emerging concept into a validated healthcare category. Companies such as Sword Health and Hinge Health have demonstrated that technology-enabled MSK solutions can attract significant investment, expand access to care, and create new models for managing one of healthcare's largest cost areas. Sword Health has reached multibillion-dollar private valuations, while Hinge Health successfully shifted from a highly valued private company to the public markets, demonstrating strong investor confidence in digital MSK care models.

Their success has proven something important: employers, healthcare systems, and insurers are ready for technology-enabled approaches to musculoskeletal health.

The need for innovation has never been greater. Musculoskeletal disorders represent one of the largest drivers of pain, disability, healthcare usage, and lost productivity. According to research published in the Journal of Medical Internet Research, annual healthcare spending associated with MSK conditions in the United States is estimated to be approximately $300 billion. The impact extends beyond direct treatment costs, as MSK conditions are often associated with other chronic health challenges, including obesity, diabetes, cardiovascular disease, arthritis, frailty, and falls. The true healthcare and societal burden is therefore significantly broader than the cost of treatment alone.

But the next question is even more fundamental:

How do we move from delivering care more efficiently to understanding changes in human function earlier?

The next evolution of healthcare risk management may not simply be better treatment after a condition develops. It may be the ability to objectively measure functional changes before they progress into disability.

This shift is increasingly reflected in healthcare research priorities, including efforts focused on aging, musculoskeletal health, digital health technologies, and artificial intelligence. The National Institutes of Health and the National Institute on Aging are emphasizing the importance of technologies capable of objectively measuring functional changes, understanding age-related decline, and supporting approaches that preserve mobility and independence.

Maintaining mobility and physical function is a fundamental component of healthy aging. Changes in strength, muscle activation, movement quality, balance, and physical performance can contribute to increased risk of falls, frailty, disability, loss of independence, and reduced quality of life.

Yet functional decline is often gradual. A person does not suddenly become frail. Changes may occur over months or years before they become apparent through traditional healthcare encounters.

Today, functional status is primarily evaluated through episodic clinical assessments, performance-based measures, and subjective reporting. These approaches provide valuable information, but they offer only limited snapshots of an individual's health and may not capture subtle changes in physiological and biomechanical function over time.

Healthcare has become increasingly sophisticated at measuring disease. The next opportunity is developing the ability to measure changes in human function. Wearable technology has already transformed personal health monitoring. Individuals can track steps, heart rate, sleep patterns, and activity levels. However, the next generation of digital health technology must move beyond activity tracking toward functional intelligence.

The question is no longer only:

"How much did someone move?"

The more important question is:

"How did someone move?"

Mobility is influenced by complex interactions among muscle activation, coordination, biomechanics, range of motion, and movement quality. Technologies that integrate physiological sensing, biomechanical measurement, and artificial intelligence have the potential to provide deeper insight into individualized functional patterns.

An AI-enabled platform integrating technologies such as electromyography (EMG), inertial measurement units (IMUs), and advanced movement analytics could establish a foundation for objective digital measures of mobility, functional aging, and early changes that may affect independence and quality of life.

The challenge in healthcare is no longer simply collecting data. The challenge is transforming increasingly complex streams of information into meaningful insights that can support better decisions.

Artificial intelligence offers the opportunity to analyze large amounts of physiological and movement information and identify individualized patterns that may not be visible through traditional assessments. The goal is not to replace clinicians, but to provide clinicians, insurers, employers, and individuals with better information to support earlier and more personalized decisions.

AI-enabled healthcare has the potential to move the system away from simply reacting to decline and toward identifying meaningful changes earlier, when there may still be an opportunity to preserve function.

For insurers, this represents a potential transformation in risk management. The traditional healthcare model has largely followed a reactive path: a condition develops, treatment begins, and a claim occurs.

The future opportunity is a more proactive approach in which functional changes are identified earlier, personalized strategies are developed, and risks associated with decline may be reduced before they progress into costly healthcare events.

Objective measures of functional health could support earlier identification of risk, more personalized care pathways, improved management of chronic conditions, support for aging populations, and value-based healthcare models focused on maintaining function rather than simply treating decline.

This shift is especially important as populations age. Aging does not automatically mean disability. Many individuals remain active, engaged, and independent when changes in mobility, strength, and function are recognized early and appropriately addressed.

Preserving function is not only a healthcare goal; it is becoming an essential component of managing healthcare costs, improving outcomes, and supporting a more sustainable healthcare system.

The future of healthcare risk management will require a fundamental shift in how we define prevention. Prevention can no longer focus only on identifying disease after it develops. It must also include understanding changes in human function before they become disability.

The most valuable healthcare technologies of the future may not simply tell us what happened after a problem occurs. They may help us understand what is changing — enabling healthcare systems, insurers, employers, and individuals to act earlier to preserve mobility, independence, and quality of life.

The future of healthcare risk management will not only be defined by how effectively we respond to disease and disability, but by how well we can identify change earlier, preserve function, and help individuals maintain healthier lives.

How to Solve the IT Backlog (and How Not to)

P&C carriers lose millions quarterly when underwriting changes languish in IT backlogs, but uncontrolled self-service creates chaos.

Insurance

In most mid-to-large property & casualty (P&C) carriers, changing a routine underwriting limit or pricing parameter requires riding the exact same deployment pipeline as core application code. The change must navigate an IT ticket, a sprint, a test cycle, a change advisory board, and a release window. By the time a single line of configuration ships, weeks or months have passed.

The result is a business operating at analog speed in a digital market. While executive leadership talks about digital transformation, front-line business teams remain tied to IT backlog ticket systems. Shortening sprint cadences or pushing for faster CI/CD pipelines fails to resolve this friction because the bottleneck is architectural, not process-driven.

The High Financial Cost of "Locked-Configuration Syndrome"

Historically, embedding underwriting guidelines, pricing limits, claims triage, and compliance checks directly into core source code or static configuration files was standard practice. That approach worked when market shifts happened over multi-year cycles, but today it locks basic operational decisions behind full software release schedules.

That friction carries a quantifiable financial penalty. For a carrier writing $2 billion in Gross Written Premium (GWP), delaying a necessary rate or eligibility adjustment on a deteriorating loss segment by a single quarter can easily affect the combined ratio by $15 million to $25 million. That is before accounting for the hidden opportunity costs of delayed product launches and missed competitive moves.

The Pitfalls of Uncontrolled Self-Service

To bypass this bottleneck, carriers often try giving business teams direct self-service capabilities. But granting autonomy without governance just trades slow delivery for production instability. When carriers attempt to bypass IT pipelines using direct database updates or rudimentary rule tools, they consistently run into three major failure modes:

  • Silent Misconfiguration: A business analyst updates a rating factor in a live environment. The change is syntactically valid, but without contextual validation, it unexpectedly alters logic in downstream claims systems. Thousands of policies are mispriced over 48 hours, requiring emergency rollbacks, data patches, and reputational damage control.
  • Audit Vacuum: During a regulatory examination, regulators request the 24-month change history for a specific underwriting rule. Because changes were made informally by multiple users across various tickets, pre-change values, timestamps, business justifications, and formal sign-offs cannot be produced. The result is regulatory penalties and corrective action plans.
  • Environmental Drift: Development teams spend six months testing a major software release in Non-Production Environments (NPE). Upon deployment, critical rules fail because business users made unrecorded configuration changes directly in production over preceding months. Testing in NPE becomes an illusion because test environments no longer reflect reality.
Configuration as a Service: A Governed Control Plane

Fixing this doesn't mean choosing between rigid IT stability and unchecked business speed. The path forward is establishing a dedicated Configuration as a Service (CaaS) layer, a purpose-built "Configuration Center" positioned between business users and core administration systems.

Functioning as a centralized control plane, this layer routes parameter updates through an automated, policy-enforced pipeline instead of relying on IT release backlogs or risky database edits. Core application platforms simply query the engine at runtime to retrieve current values and execute logic safely.

This framework rests on three core pillars:

  1. Risk-Tiered Autonomy

    Granting self-service rights based on the potential financial and operational impact of a change rather than user convenience.

  2. Integrated Governance Controls

    Built-in pre-flight simulation, automated impact analysis, mandatory sign-offs, and immutable audit trails.

  3. Continuous Environment Synchronization

    Ensuring configuration changes made in production automatically propagate back to non-production environments to eliminate environment drift.

Deciding What Deserves Self-Service: The Three Tiers

The most critical architectural decision is not choosing software tools, but classifying configuration assets by risk. Not every parameter should be editable by business teams, and items that are self-serviceable must follow controls proportional to their potential impact. 

HIGH RISK / ~10% OF CHANGES

Tier 1: Tech-Managed

Covers core data models, integration schemas, rating calculation engines, and strict regulatory policy logic. Errors here directly risk financial loss, compliance breaches, or platform instability. These assets stay firmly in IT's domain, following standard SDLC processes and formal CAB review gates.

MEDIUM-HIGH RISK / ~30% OF CHANGES

Tier 2: Governed Self-Service

Covers underwriting eligibility matrices, core rating rules, and workflow decision trees. Mistakes can affect underwriting performance, but risks remain manageable if caught before release. Business teams author and update these rules, but deployment demands pre-flight simulation and dual sign-off from both the business owner and an IT/risk lead.

LOW-MEDIUM RISK / ~60% OF CHANGES

Tier 3: Routine Self-Service

Covers bounded operational toggles, customer-facing advisory messaging, and routine threshold updates. Financial exposure is minimal, with changes mostly driving day-to-day operational efficiency. Business leads manage these end-to-end using single-approver workflows and automated fast-track promotion pipelines.

Accelerating Safety With Artificial Intelligence

Integrating targeted AI capabilities into the configuration pipeline further reduces human error while accelerating throughput:

  • Change Impact Analysis: Large Language Models (LLMs) trained on historical rule changes analyze new proposals, flagging potential downstream system impacts, edge-case scenarios, and historical outcomes before sign-off.
  • Natural Language Authoring: Analysts describe desired business logic in natural language, which AI translates into formal Decision Model and Notation (DMN) tables for validation.
  • Anomaly Detection: Machine learning models continuously evaluate configuration change streams to detect statistical anomalies in timing, parameter ranges, or change volume.
  • Automated Simulation: AI automatically generates and executes targeted test cases against proposed configuration changes in a sandbox environment, attaching results to the approval queue.
Business Outcomes and Key Metrics

Carriers implementing a governed configuration control plane transition from slow, reactive cycles to proactive market responsiveness. Key operational benchmarks demonstrate the impact:

Compressed Cycle Times: Low-risk operational changes move from multi-month IT backlogs to sub-4-hour automated deployment windows.

Incident Reduction: Production incidents stemming from configuration errors drop by 80% to 90%.

Complete Audit Compliance: Full traceability satisfies stringent market conduct standards, recording who changed what, why, and who approved it.

Resource Efficiency: Engineering hours spent managing routine configuration tickets decrease by up to 90%, freeing technical teams to focus on core platform innovation.

The tension between business agility and production stability is real, but it is not irreconcilable. By replacing one-size-fits-all IT release management with a governed, risk-tiered configuration architecture, P&C carriers can deliver the autonomy business teams require without compromising stability or regulatory compliance.

For a deeper architectural breakdown, including failure mode analyses, environment synchronization patterns, and implementation road maps, read the full white paper: Configuration Center: A Blueprint for Core System Agility.

We're Kidding Ourselves on Business Continuity

Insurers face growing exposure as the gap widens between documented business continuity plans and actual operational resilience under stress.

Insurance

The continuity plan existed. It had been exercised, and it still described a recovery the company could not have delivered. 

Insurers understand this failure mode better than most and still struggle to quantify it. A continuity plan, like an actuarial assumption or a reserve estimate, can be technically sound but still drift from the operating reality it describes. Nobody notices until an event forces the question. At that point, the cost of not noticing has compounded past the point where it would have been cheap to fix.

That compounding is where a lot of underwriting and claims exposure lives, and the data most carriers work from doesn't capture it well. IBM's 2026 Cost of a Data Breach Report put the mean time to identify and contain a breach at 247 days (about eight months), reversing five consecutive years of improvement, while the global average cost of a breach reached a record of $5 million. The same report found that organizations making extensive use of security AI and automation shortened their breach lifecycles by 65 days and lowered average costs by $1.93 million. Two months of response time was worth close to $2 million per breach.

The Regulators Moved First

Insurers don't need an outside argument for how seriously this is being treated. The Digital Operational Resilience Act became binding across the European Union on Jan. 17, 2025, and applies directly to banks, insurers, reinsurers and investment firms. The U.K.'s parallel framework for critical third parties, set out in the joint FCA, PRA and Bank of England policy statement, took effect on Jan. 1, 2025. Both moved the compliance bar from "here is our plan" to proving under testing that critical services stay within defined recovery tolerances during a severe event. A completed annual tabletop used to satisfy an examiner. What supervisors want now is continuing evidence that a plan describes how the business runs today.

Both Sides of the Desk

The move from documented plans to demonstrated recovery reaches insurers twice, as underwriters and as payers of claims, and it doesn't appear to be fully quantified on either side yet.

On underwriting, a policyholder's stated continuity of posture and its working operational resilience are increasingly two different things, and a standard questionnaire won't reliably tell an underwriter which one is in front of them. At one multinational insurance company, 48 internal processes all depended on a single third-party vendor, a concentration that had stayed invisible through every tabletop exercise the company had run. Those controls were in place and functioning, but they had never been tested against the way the business connects to its vendors day to day. That same concentration risk is just as likely sitting inside the insurer's own operations, not only the policyholders. The question is worth turning inward before assuming it applies only to what's being underwritten.

On claims, severity depends heavily on how fast an insured understood its exposure and started making the right recovery calls. The CrowdStrike outage of July 2024 shows the range. Parametrix estimated that Fortune 500 companies absorbed roughly $5.4 billion in direct losses, with healthcare and banking hit hardest and airlines carrying the highest per-company costs, largely because recovery required manual intervention across systems that offered no fast way to trace what connected to what.

The AI Assumption

There's a version of this conversation happening now around AI, and the industry risks drawing the wrong lesson from it. The common assumption is that AI makes an organization more resilient because it can produce a continuity plan or a vendor questionnaire in minutes. That holds a statement about speed, but it says nothing about accuracy. An AI tool can generate a polished recovery plan from the same outdated dependency data a person would have used, and it can do that faster. Whether the plan describes the business any better is a separate question, and it's the one that determines what happens when something breaks.

IBM's 2026 findings suggest that the governance side is lagging behind the adoption. Close to seven in 10 breached organizations had no policy for managing AI or for spotting unapproved use of it, and breaches involving AI ran roughly $1 million above the average. That gap in governance sits squarely with operations and technology leadership, not with underwriting or claims. Deciding how AI touches resilience data is a CIO and CTO call before it's anyone else's. The easier it becomes to generate documentation, the less that documentation tells anyone about whether an organization can recover quickly. What matters more is whether someone can trace, while an event is unfolding, what a failure touches, what it cascades into next, what it's costing by the hour and what should be fixed first when the people available to fix it are limited.

None of this asks the industry to rebuild how it quantifies risk. It asks for decision speed to be treated as its own category of exposure, layered on top of whatever continuity documentation an insured already has on file. An underwriter can ask how quickly a policyholder can produce a current dependency map for a named business service and how recently that map was reconciled against the systems it claims to describe. Those two answers say more about recovery than a completed questionnaire does.

What makes a gap like the one at that multinational insurance company worth examining is how ordinary it was. The controls were in place. The plan had passed every review it was ever given, and the concentration risk only surfaced when someone tested it against how the business actually connects to its vendors, rather than whether a plan existed on paper. That distinction is the whole point. A plan on file is not the same as a plan that works, and most industries, including this one, still price the two as if they were identical. The question worth asking, of a policyholder's plan or an insurer's own, isn't whether it exists. It's when it was last checked against what the business does now.


Mike Campbell

Profile picture for user MikeCampbell

Mike Campbell

Mike Campbell is CEO of Fusion Risk Management.

In 1989, he founded Campbell Software, which was acquired by SAP. He went on to serve in senior roles at SAP, FICO (EVP and COO), and International Decision Systems (CEO).

He holds a bachelor of science degree from Purdue University and was honored as a distinguished engineering alumnus and outstanding aerospace engineer.

AI Governance Needs Continuous Monitoring

AI systems in insurance evolve constantly after approval, requiring governance that monitors configuration changes, not just initial deployment.

Insurance

AI governance cannot end at deployment. In life insurance, changes to models, prompts, data, policy rules and workflow integrations can materially alter an AI system long after its initial approval. However, most AI governance programs in insurance companies revolve around this single point in time, at deployment, and if everything passes muster, the system goes live.

Then the system starts changing.

The model got upgraded to a foundation model. The prompt got rewritten. A source for retrieval was added. The policy form got changed. A new rider was included. An underwriting data source was replaced. The threshold changed. A service workflow got linked to a new API.

None of these events would look like a brand-new AI system.

A governance question arises that insurers need to ask more often: is the AI running today still the system that was approved?

Approval is a snapshot, not a permanent state

The key takeaway here is that AI approval is not a permanent attribute of an application.

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers takes a lifecycle view of the governance process, taking controls that range from development to implementation, and use all the way through monitoring, updating, and retirement. This is a very critical point because the risk profiles of an AI-enabled workflow can change even after it has been launched with the same use-case name.

For example, suppose an AI-assisted workflow for policy servicing was originally built and tested on a specific set of life insurance products and policy forms. After six months, the underlying AI model has been updated, a new source for retrieval has been added, a rider has been updated, and the workflow is calling a different downstream service.

The approval can still be present in a governance inventory. But production has a different configuration now.

Approval belongs to a governed configuration, not to a label.

That list should capture everything that the model was tested with, such as the actual prompts, retrieval sources, policy content, business rules, key data sources, thresholds, connected tools, and even the design of human review. If any of those elements change in a material way, then the question of whether the model is ready for deployment should be considered new.

The system can change without the model changing

The traditional approach to change management always begins with the same question: Is the code new?

A better question to ask is this: Did the behavior, context, or consequence change?

Take policy servicing. If a new rider changes how a benefit operates and the retrieval corpus is not updated correctly, the model and prompt may be untouched, but the AI can remain technically "healthy" while giving an incorrect answer based on outdated contractual context.

The model did not change at all. What changed was the insurance product that was built around the model.

The same issue appears in underwriting. The NAIC's accelerated-underwriting guidance for life insurers focuses on data quality, governance, transparency, and the use of external data and analytics. Even if models aren't changed, a new source of data, an evidence rule or a referral threshold can alter how cases move through the pipeline.

Let's suppose the application is connected to a different servicing integration. For instance, the replacement API may support a different set of transactions, or the workflow may feature different steps. The application may make the same recommendation, but the effect of that recommendation may differ.

And this is why it has to look beyond the model artifact, and actually govern the decision system around the model.

Life insurance makes change governance harder

Life insurance expands this predicament on the grounds that this business context itself endures a lifetime.

A single policy can stay in force for decades. Products, riders, regulations, servicing processes, data sources, and technology platforms will undoubtedly change over that length of time. The artificial intelligence system that supports a policy today will likely exist in a very different environment than when the use case was initially authorized.

The Society of Actuaries Research Institute's 2026 work on AI and life underwriting similarly emphasizes data readiness, workflow design, professional judgment and governance as insurers expand AI-enabled underwriting.

That suggests a broader principle:

The validity of an AI approval depends not only on whether the model changed but also on whether the insurance environment around it changed.

A new policy form can matter. A revised rider can matter. A new underwriting data source can matter. A different state requirement or servicing procedure can matter.

The governance process needs a way to recognize those changes before they quietly become production behavior.

Build a continuous AI governance layer

Insurers need a mechanism between change and production.

I think of this as a continuous AI governance layer.

This is not simply another committee. It does not replace legal, compliance, actuarial, security, or model-risk review. This is the operating mechanism that detects changes to a governed AI use case, determines if they are material, and triggers the right level of revalidation.

The first requirement is a governed baseline for each material production use case that will determine which models, prompts, knowledge sources, policy forms, data sources, rules, thresholds, tools, and workflow controls constitute the approved configuration for that case.

Second, change detection. If someone updates your model, revises your prompt, adds a retrieval source, changes your data schema, updates a policy form, builds a new tool, or changes a threshold, they are causing a governance event and should not expect a diligent colleague to notice it days later.

The third is materiality: Is the edit important enough to warrant a months-long review?

Every routine refresh of a retrieval index must have automated regression tests in place; every tweak to a prompt that could change the interpretation of a policy must undergo targeted scenario testing; every new foundation model dropped into an underwriting workflow must prompt a revalidation of the entire workflow; and every new integration that changes the consequences of an AI-assisted action must trigger a fresh risk review before it can go out.

The goal is risk-proportionate governance, not change paralysis.

Govern the delta, not every edit

This allows for continuing governance by focusing on the delta between approved configurations and proposed configurations.

It should also involve questioning things such as what specifically changed, which parts of the existing workflow are affected by the changes, and what evidence can be produced to show that the system still operates within the approved risk boundary.

This makes versioning a governance control instead of a mere release metadata.

You'll need a strong, continuous layer of AI governance. This layer should route changes into different responses: automation for routine, low-risk changes; human review for changes that will affect behavior; and full revalidation when the risk of the use case or customer consequence has significantly changed.

Making these decisions at an enterprise level speeds up change by simplifying individual release planning. Instead of each team debating governance from scratch, they know what counts as material and what sort of evidence is required at each level.

Monitoring should be able to reopen governance

Change is not always announced through a release ticket. Sometimes the first sign is behavior.

NIST's 2026 report on the challenges of monitoring deployed AI systems emphasizes the importance of real-world monitoring after deployment. The NIST AI Risk Management Framework Playbook likewise treats post-deployment monitoring, testing, change management, and risk response as continuing activities.

Finally, it is essential to equip runtime monitoring with the ability to trigger a governance review even in cases where no planned technical change has occurred within the insurer.

If you notice an increase in escalations, overrides, or retrieval failures in your policy servicing functions, or if you get inconsistent answers for the same type of policy form, your environment is likely to change. If you notice unusual changes in underwriting referrals or exception patterns or changes in the availability of underwriting data, the same condition applies.

The monitoring system does not need to recognize the cause of issues automatically.

A pre-deployment review asks, "Is the system ready to go live?" Continuous governance asks, "Are the assumptions that led to approval still true?"

Six questions insurance leaders should ask now

Before scaling AI further into underwriting and policy servicing, insurance leaders should be able to answer six questions:

  • Can we identify the exact model, prompt, data, policy content, rule, and tool configuration approved for each production use case?
  • Do changes to those components automatically create a governance event?
  • Have we defined materiality levels and the revalidation required for low-, medium-, and high-impact changes?
  • Can a policy-form, rider, regulatory or workflow change trigger retesting even when the AI model has not changed?
  • Can production monitoring reopen governance when behavior shifts outside the expected baseline?
  • Can we return quickly to a known, previously validated configuration when a material change does not perform as expected?

If those answers are unclear, the governance gap is no longer whether the AI passed its initial review. It is whether the organization knows when that approval has become stale.

From approved AI to continuously governed AI

These AI systems, once deployed, don't sit in stasis. Their models change, their prompts change, their data changes, the policy language changes, the integrations change, and the business processes change.

Insurance governance should assume that change is normal.

Change detection, materiality assessment, proportionate revalidation, and production monitoring are all parts of the next maturity step: a continuous AI governance layer.

The aim is not to have an everlasting review of everything, but to know when the changes are enough to challenge the assumptions that went with the approval of whatever is in front of the reviewer.

Approval is conditional on the configuration and context that were approved.

Once that configuration or context changes materially, governance should not wait for an incident, complaint, or audit to discover the difference.

Governance has to govern change, not just systems.


Bhargavi Vepuri

Profile picture for user BhargaviVepuri

Bhargavi Vepuri

Bhargavi Vepuri is a director in the insurtech and fintech field and an independent researcher with more than 10 years of experience in enterprise technology, AI-driven systems and cloud architecture. 

She holds a master’s degree in computer science from the University of Missouri-Kansas City and is currently pursuing a second master’s degree in artificial intelligence and an executive MBA at the University of Texas at Dallas. 

She is also an author, speaker, peer reviewer and technology community contributor.