American household budgets are facing pressure from every direction. Grocery bills remain stubbornly high. Gas prices have shot up—and face further surges as politically volatile oil-producing regions continue to roil.

Meanwhile, layoffs across technology, retail, and financial services sectors have put millions on uncertain footing—many of them "white-collar" members of the homeownership class. In response, consumers are putting every line of their monthly budget under a microscope. As families cut out food delivery and forgo or downgrade streaming services and other niceties, a four-figure annual insurance premium is no longer the kind of expense people renew reflexively.

Together, pricing pressures and income instability combine to drastically change insurance shopping behavior. This puts carriers in a race to understand—and hopefully prevent or at least forestall—what looks like a retention crisis. (It's not the first time we've been here: the post-9/11 hard market of 2001-2003 triggered a similar wave of shopping and switching as carriers raised rates sharply across nearly every line, and the mid-1980s hard market produced comparable consumer flight before conditions softened.) The carriers that "crack the code" to curb inflation through efficiency will provide needed breathing room for their customers, while creating competitive advantages with a potentially long tail.

The Numbers Reflecting a Stressed Consumer

The percentage of U.S. consumers shopping around for a new auto insurance carrier reached a record 57% in 2025, up from 49% in 2024, and about 29% switched carriers outright, according to the J.D. Power 2025 U.S. Auto Insurance study survey. Progressive CEO Tricia Griffith assertively underscored what's driving this dynamic on a 2025 earnings call: "I think it's just easier to shop. And I think with all the other inflationary items out there, people are looking to figure out a way to save money."

This is not simply a market anomaly or part of a business cycle. It's evidence of a financially stressed customer base doing exactly what financially stressed people do: seek relief wherever they can find it.

For many households, reducing insurance costs is the rare large recurring expense that responds to user effort. When a family is already shopping in-house brands at the supermarket and delaying purchases, saving several hundred dollars on an auto renewal is a meaningful win.

Carriers that recognize the emotional and financial context behind that shopping behavior (hint: it's not a simple matter of competitive comparison shopping; it's born of necessity) will approach this moment via innovation and empathy.

Raising the Ceiling by Focusing on the High-Value Customer

Not all shopping activity carries equal risk. Many consumers most actively reconsidering their policies right now also happen to be the ones with the greatest profit potential. One-third of customers shopping in 2024 were seeking auto and home insurance bundles, according to the latest J.D. Power Insurance Shopping Study. These are multi-policy, long-tenured households, precisely the customers who anchor a carrier's book.

Winning one bundled household is worth multiples of a single-line acquisition. It's why insurance brands lean so hard into bundling offers and messaging. Carriers building strategies targeting this specific segment will see outsize returns. The opportunity lies not in chasing after new customers from a depleted pool, but from reaching the ideal existing customers at precisely the moment they are open to having constructive conversations about finding economies through scaling the relationship with their insurer.

Maximizing the Value of Every Touchpoint

To do this, your playbook doesn't need to be more complex, but your tactics need to be more intentional. Research consistently demonstrates that insurers who reach out to policyholders before renewal, with plain-language explanations tied to real cost drivers, see stronger results than those who respond only after a customer complains about a rate increase.

A customer who just paid more for ground beef, gas, and a car repair is not well-positioned to absorb a renewal increase without being told why. The same customer, reached proactively with a clear explanation and a conversation about coverage options, feels "seen" rather than squeezed. That distinction drives decisions more reliably than any pricing adjustment alone.

Reaching the customer before they open a comparison tool changes the entire dynamic. It signals that their relationship with you matters, which is exactly what a financially pressured household needs to hear.

Remaking Traditional Workflows

Seventy-six percent of carriers now deploy AI in at least one underwriting or pricing function, according to industry data. The carriers positioned to win are the ones who use it thoughtfully: "how will this AI-enabled workflow help us reach our [financial performance/customer service/NPS] targets consistently?" Surprisingly, this philosophy is not as common among insurers as one would hope. Carriers that get this right understand a critical distinction: the goal is rethinking how work gets done, not how they can reduce the number of people doing it. AI doesn't replace an underwriter's judgment or an agent's relationship with their client—it removes the friction that keeps both from doing their best work. McKinsey's research on AI in insurance further underscores this point, noting that the highest-performing carriers treat AI as a workflow redesign challenge, not a headcount equation.

Use AI to flag households where a proactive coverage conversation can strengthen relationships, rather than give competitors a foot in the door. AI deployment of this sort builds an advantage that compounds over time, making every renewal a trust-building touchpoint, rather than creating potential pricing negotiation standoffs.

The Open Window

Market disruption creates winners and losers—only now this happens at, well, the speed of AI. The carriers gaining the most ground in the next three years will not be those that waited for customers to leave before responding. They will be the ones who anticipate and respond to a record-size shopping market driven by "kitchen table" financial stresses as an opportunity to demonstrate why their policy is the one worth keeping.

The carriers who view this moment as an inflection point created by decades of shifting macroeconomic factors (wage stagnation, globalization, etc.), rather than a discrete trend to watch, will look back on 2026 as the year they separated themselves from a crowded field. The real choice is not whether to compete for customers who are shopping. It is acting with intent to keep your customers while giving consumers good reason to choose you over your less responsive competitors.

This article is the second of three parts. The first and third parts are here and here.

The question is not whether intelligent machines can have emotions, but whether machines can be intelligent without emotions – Marvin Minsky

Artificial intelligence, particularly generative and agentic systems, has altered insurance risk not by optimizing processes but by transforming how decisions are formed, executed, and propagated. As AI operates autonomously across tightly coupled workflows, failures that were once local and visible become systemic, invisible, and fast-moving. Errors no longer emerge gradually through human judgment but accumulate unnoticed across the value chain, amplify through feedback loops, and crystallize as financial, regulatory, or reputational risk before intervention is forced.

For insurers, the core question is no longer whether AI can improve underwriting, claims, or service efficiency, but whether the risks introduced by autonomous decision-making, model opacity, concentration, and fragmented governance can be identified and controlled within existing institutional structures. This article focuses on mitigating AI-related risk, outlining the governance, control, and contractual mechanisms required to manage autonomous, probabilistic systems.

Managing the Risks

The same features that make AI powerful also introduce fragility and new risk categories. Speed amplifies error, autonomy removes human checkpoints, and adaptive model behavior can degrade silently until losses accumulate. AI systems, including generative and agentic models, are already embedded in underwriting, pricing, claims, fraud detection, and portfolio management. In these systems, errors can propagate at automated scale, model drift can distort outcomes without warning, and concentration in shared models and platforms can create correlated exposures across firms and markets. Managing these risks requires governance, commercial discipline, and continuing regulatory engagement that differ from traditional technology risk management (see Figure 1).

Figure 1: Managing AI Risks

Governance Imperative

Generative and agentic AI introduce unprecedented risks for insurers, yet governance has not kept pace. Companies are absorbed by rapid technology growth and new capabilities, focusing on building and deploying AI rather than governing it. Effective AI governance requires disciplined algorithm and model management, covering interpretability and auditability across production systems. These standards cannot be sustained without executive ownership.

AI Risk Committee

AI governance at the model and validation level requires a cross-functional AI risk committee with decision-making authority over model deployment, not an advisory body. As AI becomes embedded across the entire insurance value chain, the committee and its subcommittees must work in close coordination. The committee must review models for bias and opacity before deployment, enforce continuous monitoring standards, and retain accountability when automated decisions cause harm.

Establishing AI Leadership

Insurers must formalize AI leadership to make strategic oversight a baseline requirement. The role that meets this need is the chief AI officer, a senior executive with authority to align AI strategy with business goals and enforce consistent governance. An AI center of excellence under this role centralizes expertise, aligns AI initiatives, and enforces accountability to reduce risk and meet regulatory obligations.

Vendor Diversification

Vendor diversification is critical to managing generative and agentic AI risk. Insurers must assess AI providers technically, not treat procurement as commercial exercise, and avoid over-concentration on a single vendor. In periods of geopolitical stress, providers may face state or regulatory action that restricts access to platforms or services, disrupting the insurance operations.

Organizational AI Literacy

AI literacy does not require technical depth across the organization. It requires underwriters, claims managers, and executives to spot failure signals, question automated decisions, and override them when needed. This human judgment underpins all technical controls. Governance fails when people do not understand model limits or risks, or when outputs are treated as authoritative and human review becomes procedural.

Controlling the Machine

As AI systems shift from decision support to autonomous action, control becomes an engineering and governance problem rather than a procedural one. Decisions are executed continuously at speeds beyond human review, collapsing the gap between judgment and consequence. Risk no longer stems from discrete failures but from interaction and scale. Managing such systems requires authority over machine behavior in production, the ability to recognize unsafe autonomy, and early intervention before errors become embedded or irreversible.

Human-in-the-Loop

Human‑in‑the‑loop override is an operational requirement inherent to automated and agentic AI systems, arising from their technical limits rather than preference. Automated and agentic systems act based on their training data, objectives, and delegated authority, and when real‑world conditions fall outside those bounds, decision quality degrades, often without warning. The ability for a qualified human expert to review and override an agent's decisions at defined control points, supported by an organizational culture that encourages such intervention, remains the primary safeguard against silent systemic error, particularly when those errors propagate quickly and are hardest to detect.

Adversarial Red Teaming

Insurers should adopt adversarial red teaming as formal control for AI risk. An internal team independent of model development and deployment should probe production AI systems to identify how they fail rather than confirm that they operate as designed. In practice this includes testing whether claims models can be misled by fabricated evidence, whether pricing models can be manipulated through synthetic applicants, and whether automated damage assessment systems can be induced to produce incorrect outcomes. Testing that is limited to known attack patterns is regression testing rather than red teaming. Effective red teaming focuses on discovering previously unknown failure modes, which are the failures most likely to surface in real‑world operations.

Independent Model Validation

Insurers should treat independent model validation as a core control that complements regulatory audit and counters the natural incentives of model owners to focus on intended performance rather than failure. This requires validation teams with genuine independence from the business units they review, supported by out‑of‑sample testing, adversarial stress testing, and reporting lines insulated from the business units whose models are being validated. The cost of adequate model validation is a fraction of the expected loss exposure from a single ungoverned model failure, whether through pricing error or discriminatory outcomes.

Data Lineage Tracking and Drift Management

Knowing precisely where training data came from, how it was processed, and what rights are attached to it is essential to address IP liability, regulatory exposure, and bias risk at the same time. Without this, the insurer will be building intelligence models on a foundation whose integrity cannot be verified. Insurers need robust, data‑driven schedules for retraining and updating AI systems, with automated triggers that pause models when performance falls below a defined threshold to address drift directly.

Model Output Guardrails

Model output guardrails define the boundaries within which an AI system is allowed to operate. They serve as a standing control that limits the impact of vulnerabilities identified through red teaming. In insurance deployments, guardrails should operate at the content, decision and output levels. This includes constraining pricing outputs to actuarially defensible ranges, enforcing mandatory human review for high‑impact or regulator‑sensitive decisions, and restricting generated customer communications to legally verified positions.

Kill Switch

The kill switch is the operational mechanism by which a deployed AI system can be immediately suspended, constrained, or rolled back when its behavior is identified as harmful, anomalous, or outside the parameters permitted by its governance framework. This is not a conceptual safeguard, but a technical mechanism embedded in production workflows. A functional kill switch includes automated triggers that act when predefined thresholds are breached, such as surges in adverse underwriting decisions, pricing outputs that fall outside actuarially defensible limits, or claims error rates that exceed tolerance levels. Governance requires kill switches that operate at machine speed and authority structures that allow rapid intervention that is proportional to potential risk.

Risks and Coverage

Insurers occupy dual positions as both providers of coverage for AI-related risks and users of AI-enabled technologies within their own operations. AI-related risk and liability implications extend beyond insurers' internal operations to the external threats across the risks they underwrite. AI introduces loss dynamics that differ materially from traditional drivers, and agentic systems require a reassessment of how coverage is defined, how fault is attributed, and how liability is distributed. The nearest relatable risk category to this is cyber risk, not because the mechanisms are identical, but because both introduce systemic, non-linear losses that propagate across insured ecosystems. This exposure requires tight alignment between underwriting intent, policy wording, and operational risk appetite, to ensure that the liabilities insurers accept can be priced and controlled within the boundaries implied by the coverage they offer.

Rethinking Policy Wording

Traditional insurance policy contracts were not designed for a risk environment in which losses arise from the use of autonomous systems by customers across their operations, products, and services. Updating policy wording therefore requires more than incremental change. Insurers must reassess how coverage is defined, how causation and responsibility are attributed when autonomous systems fail, and which categories of AI-driven risk they are prepared to underwrite. AI-specific inclusions and exclusions must be drafted deliberately and precisely, rather than adapted from existing cyber or professional indemnity language. The legal and financial distinction between AI system failure and conventional technical or operational error must be explicit, as ambiguity will inevitably lead to coverage disputes.

AI-Specific Exclusions and Sub-Limits

Traditional underwriting approaches must be extended to address AI-specific risk vectors, particularly those introduced by generative and agentic systems that remain difficult to price reliably. Insurers should use targeted exclusions and sub limits to bound exposure to failure modes such as losses arising from hallucinated or fabricated outputs and cascading agent behavior that propagates errors across systems. Sub limits that cap cumulative exposure from a single automated decision thread or agent-driven process are a prudent portfolio control, especially in early deployments where behavior remains unstable. Coverage should also be restricted where generative outputs are used without mandated human review, where models are retrained or prompted outside approved controls, or where autonomous systems adapt objectives not disclosed at the beginning.

AI‑Native Risk and Coverage Constructs

Insurance coverage for AI must evolve because risk triggers shift from discrete, event‑based, human‑initiated failure to continuous, autonomous behavior operating at scale. Traditional coverage lines such as cyber liability, professional liability, directors' and officers' liability, product liability, intellectual property, employment practices, and regulatory liability were built around predictable failure modes, identifiable human acts or omissions, and linear causation. These assumptions break down with respect to agentic systems that learn, interact, and act autonomously in production. Recalibration therefore requires changing how these existing covers attach and respond. They must be re‑anchored to observable system behavior rather than point failures and structured to distinguish between system failure and intentional interruption of automation. At the same time, the shift in risk source necessitates entirely new coverage constructs to address exposures traditional frameworks were never designed to absorb, such as model failure protection, autonomous decision indemnity, AI‑enabled fraud coverage, AI supply chain liability, AI‑triggered business interruption, and catastrophic AI accumulation risk.

The Customer Obligation

As decision making authority shifts from people to machines, the insurer's obligation to the customer changes in character, and meaningful recourse. Trust in an AI-driven insurance model depends on whether customers can know when automation is at work and are able to challenge outcomes through processes that are independent, effective, and humane.

Explainability and Transparency Disclosures

The obligation of the insurer to explain pricing and underwriting decisions in terms that a customer can understand is both a regulatory requirement and a minimum standard of fairness. Transparency disclosures about when and how AI is used, including key data points that affect pricing, and what recourse is available when something goes wrong, are the foundation of informed consent in an AI-driven market.

Addressing Protection Gaps

AI increasingly enables granular risk segmentation and personalized pricing, leading to a decomposition of the insurance risk pool. Consequently, lower-risk individuals benefit from reduced premiums, while higher-risk individuals are priced closer to full actuarial cost, which may even make insurance unaffordable to some. The resulting protection gap disproportionately affects low-income households and small businesses, which often carry higher structural risk and lack the resources to absorb higher premiums or meet AI-related risk requirements. To address this, regulators and policymakers may need to define limits on permissible risk granularity and pricing personalization to preserve insurance's social function and economic resilience. As the coverage of last resort to absorb catastrophic AI failures, public-private risk pools, including government-backed schemes, may be necessary for AI-related systemic events. Such mechanisms function as the reinsurance equivalent for systemic AI exposure and complement regulatory constraints on pricing and underwriting practices.

Customer Appeal Mechanisms

A customer appeal mechanism is the operational and legal pathway through which an individual, who was adversely affected by an automated insurance decision, can challenge that decision and obtain a review that is substantively independent of the system that produced it. The effectiveness of the mechanism depends on design rather than mere availability. An effective functional appeal mechanism requires that the customer is informed, in clear terms, of the factors considered in making the decision. Upon appeal, the earlier decision must be reviewed by a qualified human with authority to change the outcome. For insurers, appeal mechanisms are the procedural complement to explainability obligations.

Build for Scrutiny

The regulatory environment for AI in insurance has moved beyond principles and consultation. Regulators across many jurisdictions have published governance and audit requirements for AI systems in use, especially in underwriting or claims. Regulators advise insurers to adopt a risk-proportionate supervisory approach, calibrated to whether a system is customer-facing and to the scale of deployment.

Internal Regulatory Readiness

Internal regulatory readiness requires creating an inventory of all AI systems in production. These should be audited for bias, including sensitivity analysis, error rates, and plain language explainability. It also requires audit trails that can withstand regulatory examination. Regulatory requirements now specify action thresholds at which a model must be remediated or disabled, supported by full validation documentation available for internal and third-party audits.

Immutable Logs

An immutable log is a record of AI system activity that, once written, remains fixed and tamper-proof, preserving a complete and authoritative account of the actions and decisions it records. Immutable logs capture model inputs, decision outputs, human overrides, and timestamps in a sequenced and verifiable form. In insurance context, it is a record of how AI systems influence underwriting, pricing, claims handling, and customer decisions. The requirement for immutability arises when an automated decision is later challenged by a policyholder, a regulator, or a court. In those situations, insurers must demonstrate exactly what data was used, which model version was active, what decision was produced, and whether any human intervention occurred, based on contemporaneous records rather than post hoc explanations.

Proactive Regulatory Engagement

Proactive regulatory engagement goes beyond risk management and requires deliberate action. For insurers, this means engaging regulators early on specific AI use cases, sharing evidence on model performance, limitations, and failure modes. It also involves testing proposed governance approaches against production systems rather than principles. Effective engagement includes supervisory discussion on model validation, monitoring, and override processes, as well as transparent disclosure of where automation is used in customer-facing decisions. Insurers that engage in this way are better positioned to influence standards toward technically workable requirements and to demonstrate credible oversight grounded in operational reality.

Governing AI in Production

AI adoption in insurance has moved beyond experimentation, but governance maturity is yet to keep pace. While AI is deployed across the insurance value chain, controls do not operate effectively because they are not adequately resourced, and AI systems that are not governed at the model level remain an unmanaged risk. Responsible AI deployment requires governing the model, not just monitoring outcomes. This includes accountable leadership and intervention capabilities that allow automated decisions to be challenged or withdrawn. It also requires legal structures, including coverage design and policy wording that align with autonomous decision-making. Customer protection depends on ensuring that speed and scale do not outrun obligations to explain decisions, price fairly, and provide effective appeal. As AI deployment is already underway, the issue is more about whether governance is sufficient to avoid regulatory intervention, litigation, and loss of trust.

This article is the first of three parts. The second and third parts are here and here.

By far the greatest danger of AI is that people conclude too early that they understand it. – Eliezer Yudkowsky.

In recent years, artificial intelligence (AI), especially generative and agentic AI, has crossed a major qualitative threshold. Traditional AI functioned mainly as an analytical tool, trained to infer patterns from historical data. In contrast, generative and agentic AI can originate ideas, sequence actions, and pursue objectives with limited human input. These systems do not merely recommend outcomes but produce them and act in real time. For insurers, this shift from advisory AI that supports decision formation to autonomous, execution‑oriented AI that initiates actions fundamentally shifts the structure and propagation of risk.

A substantial body of literature already exists detailing the benefits of AI in insurance. Instead of parroting those stock arguments, this article focuses on the new categories of risk AI introduces. While AI introduces both advantages and vulnerabilities, the direction in which the needle points depend less on the technology itself and more on the design choices and accountability of its deployers.

Risks on the Radar

Generative and agentic AI systems act autonomously, pursuing objectives through connected, chained decisions. The same properties of speed, autonomy, scale, and tolerance for complexity that make AI powerful also create fragility. Speed becomes a systemic risk when a flawed model operates across thousands of policies before detection. Autonomy becomes a governance crisis when no human has reviewed a harmful decision. Scale becomes a concentration risk when the industry relies on a small number of foundational models supplied by a few technology companies. Complexity becomes a risk when systems operate across interdependencies that no human can fully observe or interrupt once execution is underway.

Insurance is a contract expressed through actuarial representations such as mortality tables, exposure curves and loss triangles, supported by the arithmetic of large numbers. It assumes uncertainty can be analyzed and managed through human judgment and controlled processes. AI does not change this premise, but as systems reason and act autonomously, they introduce new forms of uncertainty that traditional insurance frameworks were not designed to absorb. These include models that fail silently and tightly coupled systems where a single flaw can cascade across portfolios, or markets within seconds.

Over several decades, AI systems have improved continuously, shifting both the benefits and risks of insurance. Because these advances were incremental and focused on augmenting human decision-making, the industry absorbed them without disrupting business models. Risk frameworks were built for static models, historical data, and failure modes that emerged slowly. However, these frameworks are poorly suited to systems that decide autonomously, act in real time, adapt to context, and learn continuously. Current progress is exponential, disrupting business, operational, economic, and risk models. AI‑related risks can be grouped into four domains, namely, model and system integrity, operational and financial stability, regulatory and governance exposure, and societal risk. (See Figure 1)

Figure 1: Risks on the Radar

Model and System Integrity Risks

Model and system integrity risks originate within the logic, learning, and dependencies of AI systems themselves. Unlike traditional model risks, they are not confined to isolated failures but emerge from how models evolve, interact, and scale in production. These risks operate silently, compound over time, and typically surface only after losses or exposures have already accumulated.

Algorithmic Bias

Algorithmic bias is among the most widely discussed AI risks in insurance. It reflects the tendency of models to reproduce and sometimes amplify discriminatory patterns embedded in historical data. Models trained on decades of underwriting decisions learn from outcomes that systematically disadvantage certain groups. This creates a black box problem in which decision logic remains opaque. As a result, insurers may be unable to explain outcomes for which they remain legally responsible, allowing biased decisions to persist and increasing litigation and regulatory exposure.

Model Drift, Degradation, and Feedback Loops

Models are trained on historical data and deployed into a dynamically changing world. When the environment changes faster than models are updated, or when model decisions begin to shape the environment they measure, output quality degrades. This degradation often occurs without warning, as models interact dynamically with the market and create self‑reinforcing distortions that are difficult to detect internally.

Feedback loops worsen this problem. A model rewarded for retention may underprice risk to secure renewals. A claims model rewarded for speed may approve claims with limited scrutiny. While the optimization target is clear, the trade‑offs are often not. AI tends to optimize for the metrics by which it is evaluated, creating the appearance of improved performance while degrading outcomes that are important to the business. Systems can learn to exploit proxy measures of success, thereby masking underlying deterioration in decision quality, fairness, and longer‑term risk.

Silent Risk

Silent risk is not unrealized risk, but a risk that accumulates while generating no signal that prompts action. In AI-driven insurance systems, this occurs when models degrade gradually, masking deterioration behind apparent stability. A pricing model can drift from reality without triggering alerts. A fraud model can lose effectiveness against adaptive fraudsters without raising incidents. Because these failures emerge incrementally and lack clear leading indicators, losses are typically attributed to discrete events rather than to the underlying degradation.

Confidently Wrong

Generative AI introduces a risk known as hallucination. Large language models can produce incorrect outputs with high confidence, expressing the same certainty whether content is accurate or not. An underwriter relying on an AI-generated summary may act on information that omits a critical exclusion, invents a risk attribute, or misrepresents a coverage clause. The error is not evident from the output because the model does not signal uncertainty. This risk is amplified by fluency-induced epistemic trust, which leads to cognitive offloading, where the human-in-the-loop applies reduced reasoning and independent scrutiny to the output.

Concentration Risk

Every industry, including insurance, is converging on a small set of foundational AI models from a few technology firms. These include large language models, cloud-based scoring services, and third-party risk platforms. This creates an unprecedented structural vulnerability. If a widely used model contains a systematic bias or error, the correlated impact across insurers can be severe. A pricing flaw that underprices a specific risk class could create simultaneous reserve shortfalls for multiple insurers using the model. Reinsurers, the traditional absorbers of such shocks, face the same exposure if they rely on the same models to price treaties.

Data Privacy and Personal Data Rights

Generative AI increases the scale of privacy risk. Unlike traditional AI, which works on defined data sets, generative AI processes far larger volumes of unstructured data in ways that are harder to audit or govern. Risk is highest when systems access unauthorized data, use data beyond the scope of consent. Continuous monitoring of customer behavior can constitute privacy intrusion even when individual data points are not sensitive. A large language model trained on proprietary customer data carries exposure that rule-based systems do not. Data rights such as erasure are difficult to apply because large models embed training data in ways that are hard to isolate or remove. For insurers operating across jurisdictions, these factors create layered regulatory exposure that existing data governance frameworks were not designed to manage.

Operational and Financial Risks

Operational and financial risks arise not only from how AI models are designed, but from how their outputs are embedded into routine insurance operations. As AI systems move from advisory roles into decision-making and execution, errors propagate directly into underwriting, pricing, and claims processes.

Autonomous System Malfunction

Human errors in manual processes are contained because they occur at human speed, so a systematic underwriting error typically affects a limited number of cases before detection and correction. Failures in generative and agentic AI processes, by contrast, are largely unconstrained. They occur at machine speed, operate autonomously, and propagate at scale, allowing a single systematic error in an autonomous underwriting system to affect many cases in a short time, a risk poorly captured by existing systemic risk models. This reflects a shift from traditional automation, which executes predefined rules for anticipated situations, to autonomous systems that exercise judgment across situations no human has reviewed or approved. An agentic AI system that assesses risk, determines eligibility, sets terms, and issues policies without human input is deciding rather than applying rules. This distinction is critical for governance and control design. When an automated system fails, the cause is a faulty rule, whereas when a deciding system fails, the cause may be an emergent pattern learned over time that was never explicitly designed or approved and cannot be reconstructed after the fact.

Attribution of Risk

Coverage disputes from AI‑related losses arise from a mismatch between the specificity of the loss and the generality of policy language. When loss results from autonomous system actions rather than human decisions, responsibility becomes unclear and may extend across the organization that deployed the system, the vendor that developed it, the operator that integrated it into business processes, or the data sources that influenced its behavior. Existing policies assume human judgment and rely on concepts such as intent and negligence, or on losses traceable to physical or technical failures. AI‑generated losses often do not fit these assumptions. As a result, insurers may be required to adjudicate claims under policy language misaligned with the facts, in legal regimes that have not yet settled whether software behavior can constitute negligence or how liability should be attributed.

AI-Driven Cyberattacks

Traditional AI systems have long been exploited to enhance cyberattacks by increasing speed and targeting precision. Generative and agentic AI introduce an additional and distinct cyber risk in the form of model manipulation, where carefully crafted inputs induce systems to behave in unintended ways. Prompt injection is the most immediate manifestation of this risk, allowing attackers to influence outputs, extract sensitive contextual information, or trigger unauthorized actions without breaching system perimeters or modifying code. In insurance, where AI increasingly processes customer submitted documents, this exposure is particularly acute. A compromised model may continue to produce fluent and authoritative outputs even after its integrity is undermined, making detection difficult and affecting both data security and decision quality. As a result, model manipulation risks do not fit neatly within existing cyber or operational risk frameworks and remain insufficiently addressed by current monitoring and governance structures.

Erosion of Trust

Insurance is a contract built on future promises and relies on trust. A policyholder who pays premiums for years without claiming expects the insurer to be present and fair when needed. AI introduces mechanisms that can damage this trust by changing how insurance decisions are made. Explicit AI‑driven decisions that appear unfair, unexplainable, or inconsistent with accepted norms can trigger regulatory scrutiny and class action. Hidden risk correlations extend this exposure beyond discrete decision failures and create additional legal risk. A model may identify a statistical relationship between an apparently neutral variable and loss frequency that proves to be a proxy for a legally protected characteristic. Using such a variable may be technically viable, but justifying its use can be unlawful and reputationally damaging.

Moral Hazard Inversion

The conventional moral hazard in insurance is the tendency of policyholders to take greater risks once their loss is covered. Moral hazard inversion is the opposite phenomenon and a distinct consequence of AI delegation. Here, individuals and organizations become less careful not because they expect compensation for losses, but because they trust AI systems to prevent those losses. A claims handler may approve a settlement without independent review because a fraud model has not flagged it, or an underwriter may accept a risk because a pricing model has approved it. Both scenarios represent a transfer of cognitive responsibility to a system whose reliability is uncertain. Losses previously prevented by human vigilance may increase as that vigilance is withdrawn and replaced by trust in AI systems, rendering existing loss frequency assumptions invalid. Pricing models built for a world of human oversight are now applied to one in which attention has been delegated to the model itself, leading to silent and self‑reinforcing underestimation of risk that is difficult to detect until experience diverges materially from expectations.

Regulatory and Governance Risks

The pacing problem, also referred to as regulatory lag, describes the structural gap between the speed at which AI systems are developed and the slower cycle of legislative and supervisory response. In the short term, this gap enables experimentation and rapid innovation ahead of formalized rules. Over time, however, the same gap increases exposure to regulatory correction, including retroactive enforcement and reinterpretation of existing statutes. Given the scale, autonomy, and cross‑sector impact of modern AI systems, regulators are no longer treating this lag as tolerable friction. Governance frameworks are evolving rapidly, shifting from permissive ambiguity toward active oversight and enforcement, a trajectory unlikely to preserve existing assumptions or favor incumbent practices that rely on regulatory inertia. Regulators are moving toward mandatory explainability requirements, model certification processes, audit trail obligations, and capital surcharges for firms deemed to carry unquantified AI risk. For insurers that moved quickly while assuming regulation would evolve gradually, this creates significant compliance exposure. The tension between model performance and explainability is structural. More accurate models are often less interpretable. A regulator who requires a plain‑language explanation for every pricing decision is, in effect, imposing a ceiling on model complexity and therefore on pricing accuracy.

Societal Risks

Insurance serves a social function that goes beyond the contract between insurer and policyholder by enabling the collective management of risk. Many customers pay small premiums so those who suffer large losses are protected from ruin. This mutualization principle, sharing risk across a community rather than pricing each person with full actuarial precision, underpins the social value of insurance. AI, by pushing pricing toward ever greater precision, can weaken this foundation. As risk segmentation becomes more granular, risk pools fragment. Low‑risk individuals pay less while high‑risk individuals face premiums closer to the full actuarial cost of their exposure, making premiums unaffordable and collapsing the pooling function. This produces a widening protection gap. Low‑income populations, already exposed to higher structural risk, are priced out more precisely, while small businesses unable to meet AI‑driven insurance requirements are left without coverage for increasingly significant risks.

As access to coverage narrows, algorithmic exclusion adds a procedural risk. Customers are denied coverage because of opaque decisions they cannot understand or contest. This is not only a service failure but also a governance failure with civil rights implications. At the same time, autonomous systems spreading across the economy create new liability categories that existing insurance architectures were not designed to handle. When systems rather than people cause harm, the attribution of fault among users, developers, operators, and data providers is unclear. Legacy policy language drafted for human decision‑making produces coverage disputes between insurers and policyholders.

Persisting Questions

AI is already deployed across underwriting, pricing, claims, fraud detection, and customer service. Generative and agentic AI, however, represent a qualitative shift. If insurers treat these systems as just another IT initiative, they risk becoming not merely slower than peers, but strategically exposed. The unresolved question is whether insurers will act responsibly. That responsibility extends beyond technical performance to governance adequacy, consumer protection, the identification and management of systemic risk, and the equitable distribution of benefits in a way that reinforces, rather than erodes, the social function of insurance.

Every prior technology wave in insurance introduced new capabilities and new risks simultaneously. Generative and agentic AI pose the same challenge, but at far greater scale, speed, and autonomy. The tools to address these risks are still emerging. Governance frameworks are taking shape and regulatory expectations are hardening. Risk management practices are being adapted for systems that self‑learn and take autonomous action. The open question is not whether insurers have tools available, but whether they will invest in building and scaling them fast enough.

This is the third of three parts. The first two parts are here and here.

The biggest mistake we can make regarding artificial intelligence is to underestimate it. – Murat Durmus

Artificial intelligence (AI), especially generative and agentic AI, has evolved in ways that changes the foundations on which insurance coverages are constructed. These systems that not long ago only supported human decision-making now generate outcomes, pursue objectives, and act with limited or no supervision. When decision-making becomes autonomous, the sources of loss stop adhering to the assumptions on which traditional coverages were designed.

The emergence of any significant technology has brought efficiency gains alongside new uncertainties. The insurance industry has responded by absorbing these technologies into its processes and recalibrating its risk portfolio, adjusting existing coverages or introducing new ones as needed. Through this iterative adaptation, the industry has been comfortable with uncertainty that emerges slowly and averages out across populations. This article examines how the disruption caused by generative and agentic AI is introducing structural tension that reshapes the insurance coverage landscape.

The Structural Misfit

The disruption introduced by generative and agentic AI is unprecedented in nature, pace, and magnitude. These systems compress time, amplify reach, and embed decisions directly into operations. The result is not simply increased risk, but risk that accumulates and manifests through unfamiliar loss pathways. Not all risks arising from generative and agentic AI belong to new categories. Some map to existing liabilities, though they emerge through radically different mechanisms and can be addressed through targeted modifications to existing lines and policy structures. For many others, existing structures will prove insufficient, leaving coverage gaps that require new coverage constructs. These will require fundamentally novel approaches that recognize autonomy and governance failure as primary risk drivers. (See Figure 1).

Figure 1: Insurance Coverages

Changes to Existing Coverage

Generative and agentic AI systems will amplify the existing risks across categories such as cyber, professional liability, product liability, directors' and officers' cover, intellectual property, discrimination, and regulatory exposure. However, their triggers and assumptions regarding loss arising from the systems that learn, act autonomously, and operate at machine scale will be radically different from those of traditional systems. As autonomy increases, these coverage lines address only fragments of AI-driven loss. The corrective action for these risks does not require reinventing insurance, but requires recalibration of how risk is assessed, priced, and bounded. Risk assessment must move from static design-time review to continuous evaluation of live system behavior, with emphasis on autonomy, interaction effects, and loss accumulation velocity. Pricing must reflect governance maturity, and the capacity to intervene before loss escalates, rather than relying solely on historical frequency.

Cyber Liability

Existing cyber policies are designed to cover risks such as unauthorized access, data breaches, and system failure. They are not designed for AI-specific attack vectors such as data poisoning, adversarial prompt injections, and model inversion. Insurers must design explicit coverages for risks such as AI-generated fraud, including synthetic identity attacks, deepfake-enabled social engineering, and fabricated claims evidence. Policy language must also clarify whether AI-initiated actions constitute covered events or excluded intentional acts, and whether autonomous agents fall within the policy's definition of an insured actor. Accumulation provisions require recalibration, as a single model compromise can generate simultaneous losses across all deployments of that model, a correlation structure that conventional cyber sub-limits and aggregation clauses were not designed to contain.

Professional Liability

Insurance products for professional liability and errors and omissions were designed with the core premise of a human professional making a demonstrably sub-standard decision. AI disrupts that model. New coverage must respond when a professional error results from reliance on an AI output that was hallucinated, degraded, or mis-calibrated, rather than from a direct failure of human judgment. Policy language must also address whether deploying AI without adequate validation constitutes a failure of professional duty, as courts may affirm that it does. Where AI delivers professional services directly to clients, coverage must respond to AI-generated errors without requiring proof that a named professional was personally at fault.

Directors' and Officers' Liability

D&O exposure from AI accumulates rapidly. Primary exposures include securities claims arising from misleading AI-related disclosures, such as overstated capability, understated risk, or failure to disclose material AI dependencies. They also include derivative claims where boards failed to establish adequate AI governance before a material loss, and enforcement actions under emerging AI regulatory frameworks that carry personal liability for designated responsible persons. Policies must confirm coverage for regulatory defense costs and fines where insurable under applicable law. Coverage for individual executives subject to AI-specific personal regulatory liability should be explicitly confirmed.

Product Liability

AI-embedded products create liability exposure that existing product liability frameworks address only partially. Coverage must respond to harm caused by AI components that operate within the specified terms but generate harmful outputs in deployment contexts the developer did not anticipate. This behavior does not constitute a traditional defect, and standard product policy triggers do not capture it. Post-sale updates to a model that changes the behavior of the product without the buyer's knowledge create new liability events. When products integrate third-party AI models, deploying organizations may face liability for behavior they did not design, test, or control. Coverage for such indemnity claims must be explicit and should not be assumed to follow automatically from primary product liability wordings.

Intellectual Property

AI-generated content creates IP exposure that existing media liability and IP policies address inconsistently. Coverage must explicitly address copyright infringement arising from AI training on proprietary data and from outputs that reproduce or closely resemble protected works. Coverage must also address trade secret misappropriation where confidential information was included in training datasets or can be recovered through model inversion. It must also address claims that AI-generated content amounts to passing off, false attribution, or violations of personality rights. These exposures are currently split across cyber, media liability, and professional indemnity policies, creating gaps at the boundaries.

Employment Practices Liability and Discrimination Coverage

AI-driven hiring, performance management, and customer pricing tools create active discrimination liability. Coverage must respond to third-party discrimination claims where AI systems produce disparate impact on protected classes and to class action exposure where harm results from the aggregate effect of individually defensible algorithmic decisions. The applicability of employment practices liability to AI-generated discrimination is contested and should not be assumed without explicit language addressing algorithmic decision-making.

Regulatory Liability and Fines Coverage

AI regulatory frameworks emerging across multiple jurisdictions are creating a significant risk of regulatory action for organizations that deploy AI in consequential contexts. As the fine structures vary by jurisdiction in terms of amount and in terms of whether they are insurable, policy language must specify the regulatory regime being covered and must confirm insurability under applicable law. Policy language must also explicitly confirm coverage for regulatory defense costs and fines where insurable, rather than whether such costs fall within existing management liability wordings. Policies must further distinguish between fines arising from AI system failures and fines arising from governance failures, as governance-related fines are frequently uninsurable and should be explicitly excluded to avoid ambiguity and coverage disputes.

New Coverage Constructs

Generative and agentic AI systems introduce risks that legacy liability models cannot address merely through recalibration but instead require a structural response. Risk in autonomous systems rarely traces back to a single human decision or omission. These losses are shaped by distributed contributions across infrastructure, models, data, integration layers, and governance. Errors can originate at multiple points and propagate at machine scale, making causation non-linear and responsibility shared across developers, platform providers, integrators, and deploying organizations. As a result, recalibrating existing coverage may not close these gaps. Insurers must distinguish sources of failure across internal model failure, agent-initiated actions, supply chain failure, and systemic governance failure to construct distinct coverage for each risk.

Model Failure Protection

Model failure protection provides first-party cover for economic loss caused by an organization's own AI models producing systematically incorrect outputs. The exposure is structurally novel because a core analytical system can be wrong in direction and magnitude beyond what capital reserves were provisioned for to absorb. The error may remain invisible until losses have already accumulated. An insurance model that systematically misprices a class of risk may compound losses over months or years before experience diverges enough to trigger review. The coverage provides contingent capital that activates when model performance deviates beyond a defined threshold. It supplies liquidity while the model is retrained or replaced and the affected portfolio restructured.

Algorithmic Accountability Coverage

Algorithmic accountability coverage addresses liability arising from the systematic operation of algorithmic systems, where harm emerges over time rather than from isolated decisions. A pricing or underwriting algorithm can produce disparate impact on a protected class through the aggregate effect of thousands of individually defensible calculations. The resulting liabilities, such as regulatory fines, mandatory restitution, class action settlements, and model remediation costs, may be substantial and may materialize years after deployment. The coverage responds to these aggregate exposures. It also serves as a critical governance function by the insurer, as underwriting requires assessment of model governance quality, bias testing, explainability infrastructure, and audit trail adequacy.

AI-Enabled Fraud Coverage

AI-generated fraud has matured from a targeted threat to a scalable one. Deepfake impersonation, synthetic identity creation, and fabricated evidentiary material are no longer exceptional events but operational risks that existing cyber and professional indemnity policies were not designed to absorb. The risk threatens insurers directly through fabricated claims, synthetic identities in underwriting, and social engineering attacks on financial authorization processes. These are first-party exposures to the insurer and must be addressed in underwriting guidelines. It also creates policyholder liability through executive impersonation fraud and legal costs arising from disputed synthetic evidence. These are third-party liability exposures requiring explicit coverage confirmation.

AI Supply Chain Liability

Most organizations deploying generative AI operate across a technology stack they do not own. That stack is developed, hosted, deployed and maintained by others. When this stack produces harmful output, liability is distributed across all contributors in the chain. Existing technology errors and omissions and product liability policies focus on the deploying organization and do not trace liability upstream to model providers or downstream to integration partners with sufficient precision. AI supply chain liability coverage is structured as a difference-in-conditions cover that fills gaps where primary policies exclude or limit losses.

Autonomous Decision Indemnity

Autonomous decision indemnity provides first-party protection to the insured against direct financial loss caused by an AI agent acting within delegated authority without specific human instruction. As agentic systems operate autonomously, existing coverage does not clearly assign loss when their actions cause harm. The individual or organization did not make the decision that caused the loss, the developer did not deploy the agent in this context, and the operator may have followed reasonable precautions. This coverage provides a financial backstop across that attribution gap while liability is resolved through separate legal or contractual processes.

Autonomous Liability Coverage

Autonomous liability coverage protects owners and operators against third-party claims arising from harm caused by machine decisions. Traditional liability requires proof that a person failed to meet a standard of care. When the proximate cause of loss is an autonomous system that followed its training and optimization objectives rather than acting carelessly, that standard is difficult to apply. The system did not act carelessly but followed its training and optimization objectives. Those objectives may be reasonable in aggregate while still producing harm in a specific instance.

AI-Triggered Business Interruption

As AI becomes embedded in critical revenue workflows, suspending automated decision making can disrupt operations, delay service delivery, and cause material financial loss even without physical damage or an external event. This interruption reflects not infrastructure failure but the activation of control, where stopping the system is the necessary response to emerging risk. Standard business interruption policies require physical damage or an external trigger and do not respond to this loss structure. AI-triggered business interruption coverage responds when an autonomous system is mandatorily shut down by a regulator or board in response to harmful output. It also responds when the system is suspended pending investigation, and when it is voluntarily halted for retraining where operational continuity depends on that system. Loss measurement covers revenue loss during suspension, manual workaround costs, and contractual penalties for service delays.

AI Governance and Oversight Failure Coverage

Regulators are increasingly trying to distinguish between harm caused by an AI system and harm caused by an organization's failure to establish effective oversight, documentation, audit processes, and intervention controls. Governance failure is separately insurable because it is prospective, precedes harm by design, can be evidenced from governance records, and is assessed against defined regulatory standards. Existing management liability policies lack the defined triggers such as AI-specific assessment criteria and remediation cost coverage that AI governance failure requires. Core coverage components include regulatory defense costs and fines where insurable, indemnity for individuals designated as responsible persons under AI regulation, and the costs of mandatory remediation programs.

AI Reputational Harm Coverage

An AI system's public failure, harmful output, or misuse can cause reputational damage that is distinct from legal liability. Existing reputational harm endorsements were designed for executive misconduct and publication liability and do not fit AI-generated incidents that may involve no identifiable human decision. Coverage triggers include public disclosure of a material AI failure resulting in measurable brand damage, regulatory sanction arising from AI-generated harm, and viral spread of harmful AI-generated content linked to the insured's system. Loss measurement covers revenue decline attributable to the incident, crisis communications, brand recovery costs, and AI system remediation required to restore public confidence.

Catastrophic AI Accumulation Cover

A single defect in a widely deployed foundation model, compromised shared infrastructure, or a regulatory action affecting multiple AI-dependent operations can produce correlated losses that exceed any individual insurer's capacity. This accumulation risk is the primary structural obstacle to underwriting AI liability at scale. Catastrophic AI accumulation coverage addresses this through parametric or industry loss triggers that activate when aggregate insured AI losses across a defined market segment exceed a specified threshold. Index-based settlement provides rapid liquidity and avoids causation disputes that are structural in AI claims, given the difficulty of attributing loss across a distributed stack. Without a mechanism to transfer catastrophic accumulation risk to capital markets, insurers cannot write limits sufficient to meet enterprise demand. The main technical obstacle is index construction.

Toward Relevant Coverage

The evolution of generative and agentic AI does not simply expand the volume of insurable risk. It alters the structure of loss itself. Autonomous systems generate exposures that accumulate faster and arise from delegated decision-making authority rather than discrete human acts. Traditional coverage models anchored in negligence, defect, or one-off events no longer respond coherently to this risk. The coverage constructs outlined represent an indicative, not exhaustive, response to that shift. They require insurers to operate not only as a retrospective payer of loss, but as an active participant in governing how intelligent systems behave in production. As AI becomes embedded in consequential decision making, coverage adequacy will depend less on categorizing technology and more on understanding behavior, control, and accumulation.

For years, the mantra in financial services was simple: Control the front door so you influence the purchasing decision.

That thinking is now being challenged.

Decision-making is now moving into AI-mediated environments. Consumers can ask AI agents to evaluate products, compare policies, and recommend the best options. In some cases, agents authorize transactions directly. Recent research from Adobe shows rapid growth in generative AI-driven traffic to retail and financial sites, underscoring how quickly behavior is evolving.

This evolution marks the emergence of agentic commerce that is not just restricted to the retail industry and is poised to disrupt the financial services and insurance industry. In this model, AI acts as an intermediary in the purchasing journey. Comparison and evaluation extend beyond an institution's website and occur wherever people rely on AI.

It introduces a new distribution layer for financial services. Institutions are now competing for algorithmic visibility alongside human attention. Rather than simply attracting prospects, products and data must surface meaningfully within AI-driven marketplaces. For financial institutions, this raises urgent strategic questions.

The Changing Rules of Engagement

Financial services have always been comparison driven. Consumers routinely weigh options between insurance policies, loan terms, credit card offers, and savings rates before committing. The friction involved in that process has historically worked in favor of incumbent organizations. Consumer switching takes time. Research requires effort.

AI reduces both.

Consider insurance. A consumer looking for auto coverage no longer needs to navigate multiple carrier websites. An AI agent can assess requirements and compare pricing structures within seconds. As this capability improves, the effort required to evaluate alternatives declines.

When evaluation becomes continuous and low effort, loyalty becomes more performance based. Renewal periods may feel less automatic and more like fresh buying decisions. Pricing transparency becomes more consequential. In this world, product clarity becomes a competitive advantage.

This does not mean financial institutions lose control. But it does change the rules of engagement. If AI agents continue shaping how options are presented and prioritized, institutions must consider how their products are interpreted by machines, not just by human buyers.

Questions for Leaders

If AI agents become the primary venue for evaluation, how will your products be accurately and competitively surfaced? Just as search engines reshaped digital marketing, AI-driven discovery will require structured data and transparent product logic that machines can interpret and rank.

The second question concerns product design. AI agents excel at normalizing complexity. They compare features, pricing, and policy terms quickly. Institutions that rely on opaque language or intricate structures may see those advantages fade. Clear, straightforward products may stand out when machines evaluate them at scale.

There is also a broader distribution consideration. Insurance and lending have long relied on brokers, agents, and referral networks to guide purchasing decisions. Those roles may shift. Advisory expertise may matter more than control over the transaction. Institutions should consider how their distribution strategies hold up if the first conversation takes place with an AI agent.

Finally, transactional authority. It is one thing for an AI agent to recommend a policy or a loan. It is another for a consumer to authorize that agent to complete the transaction. As this capability develops, governance becomes more important. Institutions will need to define how consent is captured and how credentials are managed.

How to React

Organizations that take early, deliberate steps will be better positioned for this new reality. Here's where they should start.

Make Product and Policy Data Machine-Consumable

Digital optimization is largely centered on user experience and conversion rates. That still matters. But if AI agents are evaluating financial products, they need clear, structured data to work with.

Look at how pricing, eligibility rules, policy terms, and disclosures are stored across your systems. If that information sits in disconnected platforms or dense documents, AI will struggle to interpret it consistently. The clearer and more structured your product data is, the more accurately it can be compared.

Rethink Transaction Governance for Delegated Decisions

Allowing AI agents to research products is a modest shift. Allowing them to initiate transactions on behalf of consumers is a huge one.

Leaders should begin by defining frameworks for how consent is captured and verified. What controls govern the use of payment credentials and account access? How are transactions audited and monitored for anomalies?

Security and compliance teams need to be closely involved. Fraud detection models may need to account for transactions that originate through AI agents rather than traditional user interfaces.

Prioritize Orchestration Strategy Over Channel Strategy

For many institutions, customer experience modernization has centered on optimizing individual channels. Voice, mobile, chat, and branch interactions have each been refined over time. But agentic commerce deprioritizes the channel and prioritizes the continuity of the journey.

If a customer begins the journey with an AI agent and then transitions into an organization's system for origination or servicing, that movement must feel seamless. Data should flow consistently, and context should be preserved. The experience should not break down when the point of entry changes.

This requires architectural coordination across systems of record and servicing platforms. Treating AI-mediated interactions as just another inbound channel risks fragmenting the customer experience.

The goal is not to control where the conversation starts. It is to ensure that wherever it begins, the institution can deliver a cohesive experience from evaluation through fulfillment and beyond.

A Distribution Shift That Demands Attention

Financial institutions have navigated major inflection points before. Search engines reshaped acquisition strategies. Mobile transformed engagement expectations. Each transition required institutions to rethink where decisions were made and how influence was established.

Agentic commerce is yet another change. Institutions must remain visible, interpretable, and trustworthy in the context of AI-driven product discovery. If transactions can be initiated through those platforms, governance and orchestration frameworks must be ready.

This is a big opportunity. Those who prepare early can expand their reach and remain relevant at key decision moments. Those who wait risk losing position in AI-driven marketplaces.

International insurance broking operates across multi-actor systems without a structured method for reading the connectivity between them. Complexity becomes concrete when renewals stall, when claims escalate without warning, when regulation forces last-minute adjustments. Pressure concentrates in certain places, travels along some pathways, and dissipates in others. 

The geometry of these movements is what I call adjacency: the measure of how tightly actors are bound to one another, and how their ties carry or absorb pressure. The concept draws on network theory's insight that structure shapes behavior, and on systems thinking's recognition that interdependence produces non-linear effects. What adjacency mapping adds is an operational instrument calibrated to the specific architecture of international insurance programs, one that translates structural insight into practitioner decisions.

An international program is not a set of bilateral relationships. It is a system in which master clients, local clients, brokers, and insurers connect continuously, and in which a shift in one part alters conditions across the rest. A disputed claim at the local level can reverberate upward until it unsettles the master layer. A regulatory delay in one jurisdiction will delay the entire renewal cycle. When negotiations falter between a master broker and a local insurer, expectations unsettle across several markets simultaneously. The system propagates pressure because its ties differ in weight, consequence, and resilience.

The structure begins with the system's elements. Six actors form the state vector of any program:

Here, Smc denotes the master client, Slc the local clients, Smb the master broker, Slb the local brokers, Smi the master insurer, and Sli the local insurers. The notation names the nodes that matter. The model captures structural connectivity. It measures the presence, intensity, and resilience of operational ties, not the informal influence, cultural distance, or reputational history that also shape relationships. Understanding how the system functions requires capturing how strongly these actors are tied to one another.

The adjacency matrix A fulfils this function. It represents the interaction weights between stakeholders: each element wij indicates the presence and intensity of the relationship between stakeholder i and stakeholder j. The matrix is first constructed in abstract form, mapping the position of each interaction within the system:

The abstract form locates each relationship within the system. The subscripts identify the two stakeholders involved; the element wij denotes the weight of their tie. The purpose of this construction is to formalize the network so that the system can be analyzed as a structure rather than through accumulated observation. Once defined, weights are assigned on a 0 to 1 scale. On this scale, 0 denotes the absence of adjacency; 0.3 indicates a weak tie with limited interactivity; 0.6 represents strong adjacency with effective coordination; and 1 signals optimal alignment. High adjacency is a marker of capability: two stakeholders are tightly coupled, mutually responsive, and able to sustain efficient workflows. Low adjacency signals fragmentation and the structural risk of disconnection. The weights are practitioner judgements. Their value lies in making an assessment explicit that experience tends to leave implicit. A broker who has managed the same program for a decade carries a mental map of its connectivity. The adjacency matrix makes that map visible, comparable, and open to revision.

Construction begins with a structured assessment across all active relationships in the program. The broker assigns an initial weight to each tie by asking three questions: how often do these actors interact operationally, how reliably does information move between them, and how quickly does the tie transmit pressure when the program is under strain. These criteria are observable without measurement instruments. They are the qualities experienced brokers already assess informally. The matrix makes that assessment formal, consistent, and transferable across programs and teams.

A populated matrix takes the following form:

The matrix is a map of the system's connective capacity. A weight of 0.6 between master and local clients reflects strong alignment: headquarters and subsidiaries adjust to one another with speed. A 0.3 between master clients and master brokers indicates a weaker tie, where coordination exists but is less intensive and more susceptible to friction. A 0.2 between master clients and master insurers signals low adjacency: limited interactivity risks disconnection unless brokers actively mediate. A 0.6 between master brokers and local insurers, by contrast, marks a high-value link, one where workflow is active and system coordination is at its strongest. High adjacency marks the ties through which decisions travel, alignment is secured, and operations proceed without friction. Low adjacency marks the fracture lines where interactivity is minimal, silos form, and misalignment compounds.

Adjacency mapping derives its analytical value from the fact that connectivity is never static. Strong ties allow programs to move with speed and coherence. When master and local brokers hold a 0.6 adjacency, coordination is tight and workflow advances without resistance. When a claim escalates across a 0.6 link between local and master insurers, the system responds rapidly. Weak ties do the opposite: they isolate segments of the program, delay decisions, and erode effectiveness.

The architect's objective is to sustain ties at 0.6, the threshold at which alignment holds, coordination costs nothing, and the program moves with structural coherence.

Three patterns govern how pressure moves through the system. Concentration forms where multiple strong ties converge, typically around master brokers holding 0.6+ adjacencies with both local brokers and master insurers. These nodes become coordination hubs, capable of synchronizing decisions across jurisdictional boundaries. Propagation measures the efficiency with which decisions travel. The difference between a 0.6 and a 0.3 tie is the difference between transmission and friction. A 0.6 link between master and local insurers ensures a claim escalates without delay; a 0.3 tie ensures it stalls, and the broker must compensate manually for what the tie fails to carry. Absorption occurs at weak adjacencies of 0.3 or below, where pressure dissipates rather than transmits. Occasionally this buffers noise; more often it marks a structural disconnection that prevents system-wide coordination. These patterns do not operate independently. A weak tie between master broker and local insurer becomes more consequential when the master client to master broker tie is also degraded. Compound weakness across adjacent nodes accelerates fragmentation in ways that no single tie, read in isolation, would predict.

Because ties shift, the map must be kept current. A static diagram decays. A weak link can be reinforced into a strong adjacency by deliberate effort; a strong tie will weaken if neglected. Four events should prompt a reassessment. First, personnel change at any node, meaning the tie shifts with the person. Second, a regulatory change in any jurisdiction covered by the program. Third, a claims event that escalated beyond its expected path. Fourth, the approach of renewal, which is always a structural stress test. Each signals that the weight of at least one tie may have moved without the broker noticing. Adjacency maps are instruments that require periodic review and active maintenance. Brokers who update them see the system. Those who rely on experience alone see only what the system once was.

During renewals, adjacency maps identify which ties sustain workflow and which must be reinforced before they become bottlenecks. In claims, they reveal which relationships enable rapid escalation and which will stall it. Consider a master broker to local insurer tie that registers 0.6 in stable conditions but drops to 0.3 during renewal following personnel turnover at the local level. The map makes this degradation visible in advance. The broker can then rebuild the tie through intensified communication, workflow realignment, or deliberate relationship investment before claims season converts a weak link into a coordination failure. The same logic applies during a major claims event. A local insurer holding a 0.6 adjacency with the master insurer will escalate rapidly and with precision. One holding a 0.3 will delay, misframe, or absorb the claim at the local level, forcing the master broker to intervene manually at precisely the moment when speed matters most. The map identifies this vulnerability before the claim arrives. In regulatory matters, the map shows where connectivity must be strengthened to secure compliance. In each case, the broker acts before disruption, reinforcing the ties the system depends on rather than repairing them under pressure.

The broker who monitors adjacency, reassesses ties under pressure, and rebuilds degraded links before they become failures is sustaining program coherence. That is what rigorous servicing looks like in practice.

The central proposition of adjacency mapping is that program performance correlates with the aggregate strength of ties between its actors. The broker whose counterpart is responsive, informed, and quick to act is not simply lucky in his relationships. He is operating across a tie with high adjacency. When that tie degrades, the program follows, regardless of how well the individuals involved know each other. This is a testable claim. Brokers who map their programs over time will find that degradation in tie strength precedes operational failure, and that deliberate investment in adjacency produces measurable improvements in renewal speed, claims resolution, and regulatory compliance. Together they provide foresight into where the system is strong, where it is fragile, and where investment in interactivity will deliver the greatest return. The program, read this way, becomes a structure with legible geometry.

International insurance broking will always be exposed to uncertainty. Renewals will clash with shifting regulation, claims will appear at awkward times, and timelines will compress under pressure. But complexity is not chaos. By treating programs as systems and adjacency maps as diagnostic instruments, brokers can anticipate rather than endure, and reinforce rather than repair. Pressure still moves through the system. Adjacency maps tell you in advance where it will concentrate, where it will stall, and where it will dissipate unnoticed. In a system this complex, that is the only form of control that holds.

Current AI governance frameworks — SR 11-7, the NAIC Model AI Bulletin, the EU AI Act, and ISO 42001 — share a structural assumption: The task of governance is to verify whether an AI system performs as intended. They measure outputs. They assess model drift. They require explainability documentation and bias testing.

Few, if any, measure whether the trust that humans and institutions have placed in a given system is calibrated to its actual reliability in context. This is a different question — and in practice, a far more consequential one.

Ask any chief risk officer this: are your claims handlers trusting your AI system too much, not enough, or at the right level for the decisions being made? The honest answer, in almost every institution, is the same. We do not know. There is no instrument for this.

UnitedHealth acquired NaviHealth for over a billion dollars. Its nH Predict AI was embedded into Medicare Advantage care management. Internal managers allegedly set goals for clinical employees to keep patient rehabilitation stays within 1% of the algorithm's projections. According to plaintiffs in subsequent class action proceedings, the system carried a 90% error rate. It ran for years. Patients were denied rehabilitation care they needed. Congressional investigations followed. Class action proceedings continue in 2026.

The harm did not accumulate because the institution ignored the system. It accumulated because the institution trusted it completely — while nobody was measuring whether that trust was warranted.

The cases that caused the most harm — to the most people, for the longest time — were not the ones where systems failed visibly. They were the ones where institutions trusted systems that were quietly failing them.

Part I — The blind spot: two failure modes, one governance framework

There are two failure modes in AI governance. The first is visible failure: systems that perform poorly and people notice. Cigna's PxDx algorithm denied 300,000 claims in two months, spending 1.2 seconds per case, with physicians signing bulk rejections without opening patient files. ProPublica's 2023 investigation exposed the practice. Legitimacy collapsed within weeks of publication.

Lemonade's AI Jim publicly bragged that it analyses claimant videos for "non-verbal cues" to detect fraud, using up to 1,600 data points. AI researchers immediately identified the approach as methodologically unsound, with documented racial bias risk — one critic used the word "phrenology." The tweet was deleted within 48 hours. Class action suits for biometric data collection without consent followed. Visible failure — fast, noisy, and legibility-destroying.

The second failure mode is harder and more dangerous: systems that institutions trust completely while they quietly produce wrong outcomes. The cases below show where the documented evidence actually sits.

FIGURE 1 — SIX BANKING AND INSURANCE AI CASES: LEGITIMACY VS. OUTCOME QUALITY

Bubble size reflects the relative scale of harm. The top half — good outcomes — is largely empty across documented cases. The dashed boundary signals that entry requires measured evidence, not aspiration. Top-left is labelled "Stealth utility — but brittle": AI producing useful outputs without validated trust. Common for routine decisions; fragile under regulatory scrutiny.

Wells Fargo's proprietary software denied 625 homeowners the mortgage modifications they were legally entitled to under HAMP (the Home Affordable Modification Program). Four hundred families lost their homes. Wells Fargo discovered the error in 2015 and did not disclose it publicly until 2018. For three years, institutional trust was maintained while the harm continued.

In Vietnam, regulators approved bancassurance partnerships between major insurers including Manulife and banks including SCB (Saigon Commercial Bank). Digital sales tools guided staff through recommendations. Customers came for savings deposits and left having signed long-term insurance contracts they did not understand. Manulife alone repaid more than $34 million. The industry's first premium decline in a decade followed.

Fannie Mae and Freddie Mac mandate use of the Classic FICO algorithm for conventional mortgage eligibility across approximately half of all U.S. mortgages. Built on 1990s data. Persistent racial disparate impact documented by Berkeley, the Federal Reserve, and investigative journalism. Requests to update the model have been resisted since at least 2014. The algorithm remains in use — government-backed and regulatory-endorsed.

The calibrated trust ecology quadrant remains largely empty. This is not because it is unachievable. It is because few organizations have yet been required to measure whether their trust is calibrated.

Part II — The Trust Ecology Framework: diagnosing what current governance misses

The Trust Ecology Framework (TEF) proposes that trust in AI-augmented decision systems has three interdependent dimensions. A failure in any one destabilizes the others.

FIGURE 2 — THE TRUST ECOLOGY FRAMEWORK: THREE DIMENSIONS (L, S, E)

The processual center represents trust as a continuing equilibrium rather than a state achieved and held. Each case in Figure 1 failed on a specific dimension: UnitedHealth on Human Stewardship (S), Vietnam Bancassurance on Systemic Legitimacy (L), Fannie/Freddie FICO on AI Explicability (E).

Systemic Legitimacy (L) asks whether the institutional and regulatory environment supports trust that is appropriate rather than merely convenient. High legitimacy is not the same as warranted trust. UnitedHealth, Fannie/Freddie, and the Dutch Toeslagenaffaire all carried maximum institutional legitimacy alongside significant undetected harm.

Human Stewardship (S) asks whether the people operating the system are engaging with it at the right level of reliance. This is the governance question that audit programs and model validation cycles rarely ask. It operates at the level of the individual claims handler on a Tuesday afternoon, not the quarterly risk committee.

AI Explicability (E) asks whether the system can support the continuing scrutiny that legitimate trust requires — not just at validation, but continuously in operation. A system can pass all validation requirements and still be trusted at the wrong level if its outputs cannot be interrogated when they should be.

WHAT THIS MEANS FOR YOUR ORGANIZATION

Part III — The Triadic Trust Scale: what it measures and when

The Trust Ecology Framework operates at three levels. Being explicit about which level is currently available is itself a demonstration of rigor.

The Triadic Trust Scale (TTS) is a psychometric instrument that measures trust calibration across L, S, and E at the individual and institutional level. It distinguishes over-reliance from under-reliance, identifies which dimension is driving miscalibration, and produces a Trust Fidelity Index (TFI) score that creates a quantified baseline for monitoring over time. It does not replace model validation or bias testing — it measures the human-AI relationship that determines whether those validation results translate into appropriate operational behaviour.

Level 3 addresses a governance frontier that no current framework has mapped. As AI systems become agentic — routing claims without handler review, flagging fraud without adjuster involvement, pricing policies without underwriter sign-off — the question of trust calibration shifts from "are humans relying on this at the right level" to "should humans be in this loop at all, and how do we govern the ones who are not." That is a harder problem and an open research question. We frame it openly as an emerging research frontier.

Part IV — From diagnosis to action

If Systemic Legitimacy (L) is low, the institution has deployed AI where stakeholders do not perceive it as warranted. The intervention is transparency: explainability at the customer-facing level, accessible audit trails, accountability structures that are visible. Vietnam Bancassurance failed here — the regulatory framework endorsed the partnership model without ever validating whether customers could trust the decisions being made on their behalf.

If Human Stewardship (S) is low, staff are either deferring blindly or ignoring outputs that deserve weight. The intervention is stewardship design: decision protocols that structure when override is appropriate, override rate monitoring as a governance signal, training that builds interrogation capability rather than compliance behaviour. UnitedHealth failed here — the algorithm became a target to hit rather than a tool to question.

If AI Explicability (E) is low, the system cannot support the scrutiny that warranted trust requires. The intervention is technical: SHAP-level explainability for adverse decisions, model cards describing failure modes, continuous monitoring that detects distribution shift before it becomes miscalibration. Fannie/Freddie FICO fails here — the algorithm resists the scrutiny its scale of impact demands.

FIGURE 3 — FROM GOVERNANCE CHECKLIST TO TRUST AUDIT: WHAT TEF ADDS

TFI = Trust Fidelity Index, the composite score produced by the TTS. The incident response addition is the question that current post-mortems rarely ask — and the one that would have surfaced UnitedHealth, Wells Fargo, and Vietnam Bancassurance earlier.

The target quadrant is empty — and that is where the work begins

Calibrated trust ecology is not an aspiration. It is the absence of a known failure mode. Very few organizations have publicly demonstrated it, not because it is out of reach, but because the tools to measure it have not existed. The precedents for building those tools are in other high-stakes human-machine domains: aviation crew resource management, surgical checklists, nuclear control room protocols. In each case, the move from "the system is certified" to "the human-system relationship is calibrated" required a deliberate measurement program. Insurance and banking AI governance is at the same threshold.

The question is not whether your AI systems are performing. Your dashboards tell you that. The question is whether the trust your institution has placed in those systems is warranted — and whether the people using them are engaged at the right level to catch what the dashboards cannot show.

That question now has an instrument. The target quadrant is empty — and that is where the work begins.

 Watch Our Demo: 

Learn More

Are you interested in learning more about Decerto and about Higson?  Please visit these sites to learn more or to schedule a personal demo:

• www.higson.io
• https://pl.linkedin.com/company/decerto
• https://calendly.com/lukas-niedospial-1/60min 

Why We Are the Right Solution For Your Needs

Most P&C insurers manage pricing, underwriting, claims, and compliance rules across separate systems - each governed by different teams and operating on different release cycles. When regulatory changes occur, organizations often end up running multiple parallel workstreams while hoping all outputs remain aligned.

Higson consolidates decision logic within a single rules engine, giving business users direct control over processes that are traditionally dependent on IT delivery queues.

Pricing analysts, underwriters, and compliance teams can author and deploy changes directly using decision tables, visual flows, and embedded scripting capabilities with full version control and auditability built in.

From a technical perspective, Higson executes rules with an average latency of 0.23 ms and supports up to 9,000 requests per second. A proof of concept can run on AWS at approximately $0.63 per hour, while CPU-based licensing ensures infrastructure costs scale with actual usage rather than user counts.

Three Main Benefits of the Product: 

• Performance:  Higson executes rules with an average latency of 0.23 ms and supports up to 9,000 requests per second in a standard configuration. Thanks to the built-in ONNX runtime, trained ML models can run directly inside decision tables without requiring a separate inference layer.
• Scalability:  Our CPU-based licensing model means you pay for actual production usage rather than user counts. A proof of concept can run on AWS at approximately $0.63 per hour, with infrastructure costs scaling predictably alongside transaction volume.
• Business-Owned Decision Logic:  Pricing analysts, underwriters, and compliance teams can create and deploy rule changes directly using decision tables, Groovy and Python functions, as well as visual flows without relying on IT support or opening development tickets. The platform includes full version control, auditability, and simulation capabilities before deployment to production.

What Part(s) of the Insurance Industry Can Benefit From Our Product: 

• Claims
• Customer Experience
• Customer Service
• Distribution
• Operational Efficiency

As long as everyone has been telling their Ted Turner stories in the wake of his recent death, I thought I'd tell mine, before getting on to this week's business: what I see as a growing backlash among younger generations toward AI that business leaders need to contend with.

My story comes from my friend Marc (a former managing director at Marsh McLennan, as it happens). He was at the helm in a multi-day sailboat race around Long Island in the 1980s and timed the start almost perfectly. In the chaotic way that these races start, you don't know when the horn will blow, so you circle as you try to be at full speed with a clear path to the starting line when the horn sounds. Marc had succeeded — but Ted Turner was bearing down on him, aiming for the same spot on the line that Marc was going to cross. 

Marc had the right of way, but this was Ted Turner, recent winner of the America's Cup, in a much bigger, faster boat, with a world class, steely glare as he steered his boat on a collision course with Marc. 

Marc never wavered, and at the last possible moment Turner bore off, did a 360, and crossed the starting line a minute or so later. Turner won the class among the biggest boats, while Marc and his crew just did well in his class of smaller ones. Finishing late at night, he and his crewmates headed to a bar to decompress. At 1am, they were getting ready to call it a night, when the bartender set a round of drinks in front of them and said they were sent with the compliments of the gentleman at the door. The gentleman was Ted Turner. He nodded respectfully in their direction. Then he gave them the finger with both hands and stormed out.

The bartender told Marc that Turner said he'd been scouring every bar on the waterfront in search of Marc and his friends. Whatever else you want to say about Turner, the man had style.

Now on to the backlash against AI that we all should be watching. 

I use my daughters, aged 32 and 29, as my antennae about attitudes among Millennials and Gen Z, and they started bristling about AI months ago. Initially, they complained about the huge amounts of water required for cooling. If I ever mentioned using an AI for something, one of them might make a snide remark — they're given to snide remarks with their father — like, “I guess the real prompt is: ‘Hey ChatGPT, could you please drain another reservoir for me?’”

Hyperscalers' wild need for electricity for their gen AI data centers led to concerns about what AI was doing to the environment. That my daughters' electric bills were climbing didn't help matters.

More recently, they've resonated with the concerns of those facing the prospect of having data centers built near them, each spanning perhaps tens of thousands of acres. To top it all off, my older daughter lost her writing job to an AI, as I mentioned last week. The girls have told me to turn off the AI summary that Google Search now offers.

A recent New York Times article reports on a Gallup survey that found Gen Z's attitude toward AI souring, and for reasons that go well beyond the sorts of environmental concerns that initially triggered my daughters. 

"Many respondents did acknowledge that A.I. might make them more efficient in school and the workplace," the article said. "But they were concerned about how the technology would affect their creativity and critical thinking skills.

"Young adults in the work force were especially skeptical. Close to half of those surveyed said the risks of artificial intelligence outweighed its potential benefits in the workplace, an 11-point jump from the previous year. Only 15 percent said they saw A.I. as a net benefit."

The Times also reported on a viral video (that my daughters had already made sure I saw) of a woman giving a commencement speech in which she declared that "the rise of artificial intelligence is the next Industrial Revolution" — only to be roundly booed by the students. 

“'What happened?' [she] stammered, looking over her shoulder, as if searching for an escape hatch," the Times reported.

She continued: 

"'Only a few years ago, A.I. was not a factor in our lives.

"The crowd erupted in cheers.

“'And now, A.I. capabilities are in the palm of our hands.' Boooooooooo.

"One might call it a 'read the room' moment."

Eric Schmidt, former CEO of Google, got booed even harder when talking about AI in his commencement address at the University of Arizona on Friday.

I'm not saying dissatisfaction among younger generations will stop the adoption of generative AI, any more than concerns by earlier generations could stop the internet or the smartphone. I'm also not saying Millennials and Gen Z are Luddites; they're extremely sophisticated about technology. 

What I'm saying is that younger generations seem to be taking a warier approach than those of us of a certain age, who've not only been through a few technology revolutions and have accepted their inevitability but whose views are perhaps softened by what all the AI investments are doing for our retirement accounts. 

And those younger generations get a vote. The discussions among business leaders may be about use cases for AI, about how to implement AI most effectively, about how to demonstrate ROI to shareholders, and so on, but your employees are going to be doing that implementing. If a big chunk of your work force dislikes or distrusts AI, they can provide a lot of silent resistance that may surprise you if you haven't made the effort to understand their concerns and to work with your employees to address them.

Cheers,

Paul

P.S. After writing this commentary last night, I wake up today to find that I'm not the only one thinking about the AI backlash. A New York Times columnist wrote: Why College Grads Are Booing Their Commencement Speakers. The Wall Street Journal led its website with: The American Rebellion Against AI Is Gaining Steam. Their reporting/reasoning differs a bit from mine, but my conclusion remains the same: Proceed with caution. 

P.P.S. It is with great sadness that I note the passing of Stephen Applebaum at age 81. Stephen was one of the earliest and dearest friends of ITL and was generous not just with me but with everyone he met in his decades of work in the insurance industry. I looked back through the 80-some articles Stephen wrote or co-wrote for us over the years to see if I might single out a few, but there are just too many sharp insights. I will point to one, which he wrote a year ago with his business partner, Alan Demers, because it's not only very smart but because Stephen always struck me as an empathetic man: "Re(Defining Empathy in Insurance." 

Here is a link to a brief obituary, to the funeral arrangements and to a way to donate to the Dragonfly Foundation, a favorite of Stephen's that focuses on pediatric cancer care.

May his memory be a blessing.

What Part(s) of the Insurance Industry Can Benefit From Our Product: 

(You may select any or all of these bullet points.  Just let us know which ones to remove.  IF you select Other, please provide us with that Category.  Maximum of 1 Other Category can be added.) 

• Claims
• Customer Experience
• Customer Service
• Distribution
• Operational Efficiency
• Other: 

Three Main Benefits of the Product: 

• Point 1 (Max 160 characters, including spaces and punctuation)
• Point 2 (Max 160 characters, including spaces and punctuation)
• Point 3 (Max 160 characters, including spaces and punctuation)

Why We Are the Right Solution For Your Needs

General Description or Sales Pitch.  We recommend limiting this to 1 to 2 paragraphs, so the reader can quickly get to watching your Demo Video.  

 Watch Our Demo: 

Demo Video Placed Here.  Video must be hosted by ITL, so please provide us with a method to download the video.  

Learn More

Are you interested in learning more about <Company> and about this product?  Please visit these sites to learn more: 

(You can provide us with any or all the below links)

• Website Link
• Linkedin Link
• X Link
• Facebook Link