Editor’s Note
This week, several developments pointed toward the same transition.
AXA has attached a recurring earnings target to its AI strategy. Aon is watching AI-related claims migrate into familiar directors and officers liability theories. State regulators are debating how broadly their AI examination framework should reach. And Cowbell is using AI not simply to price cyber risk, but to tell insureds what to do about it.
AI is becoming less separable from the businesses using it.
That changes the questions.
How much value does it create? Who is accountable when it fails? Which systems deserve additional scrutiny? And can the technology become part of the mechanism used to control the risk it creates?
Those are insurance questions now.
– James W. Moore, Editor-in-Chief
AXA Puts a Number on AI
AXA’s new 2027–2029 strategic plan includes something still relatively uncommon in insurance AI announcements: a financial target.
The company says artificial intelligence could generate an additional €500 million to €700 million in recurring pre-tax earnings annually by 2029. AXA describes the figure as net of investments. (AXA Investor Day)
That matters because it begins converting AI from an innovation program into something management can evaluate using the same language as the rest of the business.
A recent analysis from AI Risk takes the argument farther. It notes that AXA’s plan combines automation, better prediction, operational efficiency and growth, while asking whether those advantages remain durable once competitors gain access to similar models and capabilities. (AI Risk)
The article builds that critique around its own “Intelligence Capital” framework. The framework is less interesting to me than the underlying question.
Why it matters: Once insurers start publishing expected AI earnings, boards and investors can start asking what produces them.
Is the return coming from fewer employees? More business written with the same employees? Better pricing? Lower fraud? Higher retention? Lower claims leakage?
And, eventually, how much of the advantage survives when everyone has comparable technology?
The first stage of enterprise AI was proving that the technology could do useful work. The next may be proving that the economics survive after the technology becomes ordinary.
AXA has at least given the market a number against which that can eventually be measured.
AI Liability Is Starting to Look Like Ordinary D&O Liability
Aon published an analysis this week examining the expanding potential for AI-related directors and officers claims.
The important point is that the litigation theories do not necessarily require a completely new category of AI law.
Aon points to securities litigation involving AI-related disclosures, derivative actions involving board oversight, and claims alleging companies overstated AI capabilities or failed to manage associated risks appropriately. It also notes that traditional D&O coverage generally remains available for many of these claims. (Aon)
That creates an important distinction.
AI may be technologically unusual without producing legally unusual allegations.
A shareholder does not necessarily need to prove that a board misunderstood the architecture of a foundation model. The allegation could be much more familiar: management made a material representation, failed to disclose a risk, or did not provide adequate oversight.
Why it matters: AI governance is beginning to acquire a second audience.
The first audience was internal: technology, compliance, legal, risk management and the board.
The second is external: plaintiffs, regulators and insurers looking backward after something went wrong.
Policies, committees and principles matter. Evidence that those mechanisms actually operated may matter more.
Who received the warning? What information was available? Who had authority to act? What was escalated? What decision was made?
Those questions existed long before generative AI.
AI is simply creating a new set of facts to which they can be applied.
The NAIC Is Still Deciding How Wide the AI Examination Net Should Be
The NAIC’s Big Data and Artificial Intelligence Working Group met Thursday to continue discussions on its AI Risk Evaluation Supplement, following the September 29 deadline for comments on Version 5.0.
The supplement is being developed as a common instrument regulators can use when examining an insurer’s AI systems, governance practices, risk controls, model inventories and data. The NAIC has been piloting the broader effort with participating states and has been working toward consideration at the 2026 Fall National Meeting. (NAIC Working Group)
The latest comment process exposes a basic problem regulators will have to resolve: what belongs inside the definition of AI risk worth examining?
Industry groups have pushed back on including established statistical tools such as generalized linear models and generalized additive models. Consumer representatives have raised a different concern, questioning whether insurers should effectively determine for themselves which models qualify as high risk. Other comments address proxy variables, unfair discrimination and what constitutes meaningful human review. (NAIC Version 5.0 Comments)
Another public discussion is already scheduled for November 2. (NAIC)
Why it matters: The difficult regulatory question may not be how to examine AI.
It may be where AI begins.
A modern insurer can contain traditional statistical models, machine learning, third-party scoring systems, rules engines, generative AI, agentic systems and combinations of all of them.
Draw the boundary too broadly and nearly every consequential analytical system becomes an AI governance problem.
Draw it too narrowly and the regulatory framework may miss systems capable of producing exactly the consumer outcomes regulators care about.
The eventual boundary will say something important about whether insurance supervision treats AI as a special technology or increasingly treats automated decision systems according to what they do.
Cowbell Moves AI From Underwriting Into Loss Control
Cyber insurer Cowbell launched an AI-powered Risk Advisor this week that uses information about individual policyholders to produce prioritized recommendations for reducing cyber risk.
According to Cowbell, the system combines its existing risk indicators with security findings, dark-web exposure, connected security tools, policy information and AI-specific measures including observability, autonomy and governance. The resulting recommendations are presented to policyholders with supporting evidence. (Cowbell)
The vendor language deserves the usual caution. A product announcement does not establish that the recommendations reduce losses.
But the operating model is worth watching.
Why it matters: Insurance has never been limited to paying claims.
Commercial insurers inspect buildings. Boiler and machinery carriers employ engineers. Workers’ compensation insurers provide safety services. Cyber insurers scan external infrastructure and recommend controls.
AI can make that loss-control relationship substantially more continuous.
Instead of periodically reviewing a risk and waiting for renewal, an insurer can potentially observe changing conditions, identify a problem and recommend an intervention while the policy is still in force.
That creates an interesting feedback loop.
Better information improves underwriting. Underwriting identifies the controls that matter. Continuous monitoring identifies when those controls change. Loss experience eventually tests whether the intervention worked.
The insurer becomes more than the organization pricing the risk.
It becomes part of the system trying to change it.
The question is whether the loss results eventually prove that the additional intelligence actually made the risk better.
Sources
- AXA, September 2026 Investor Day and “Growing Forward” 2027–2029 Strategic Plan
- AI Risk, “AXA Has the Best AI Plan in Insurance. But It’s Missing the Part That Lasts.”
- Aon, “AI-Related D&O Claims and Emerging Litigation Risks”
- National Association of Insurance Commissioners, Big Data and Artificial Intelligence Working Group, AI Risk Evaluation Supplement Version 5.0 and Public Comments
- Cowbell, AI Risk Advisor Announcement
AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.
