The most common objection to underwriting AI is that there isn't enough claims history to support confident decisions, which is fair, but it risks treating claims data as the only useful source of information. Insurers that have already been underwriting AI companies have their own experience to draw from, and those with a longer history in frontier sectors are used to assessing risks before conventional loss data has had time to mature.
That is an important distinction because
The first question is how AI is being used
A year ago, much of the insurance conversation focused on whether a company was using AI at all, but that question is already becoming less useful because
There is a meaningful difference between AI assisting an employee and AI running a workflow with limited human involvement, just as there is a difference between a low-risk internal tool and a system making decisions that affect customers, counterparties or regulated activity. Those differences are much more useful to an underwriter than a broad label such as an AI company or AI user.
As systems become more agentic, that distinction becomes more important because the technology is taking more of the process on itself. An underwriter needs to understand where the limits sit, who owns the outcome, how quickly a person can intervene and whether the company can reconstruct what happened after something goes wrong.
Claims data is only one part of the picture
Claims experience matters, but it is only one input into an underwriting decision, particularly where the technology and the business models are changing quickly. Governance, accountability and the quality of the control environment can all reveal differences between risks long before a mature claims triangle exists.
That means asking practical questions about who is responsible for the model, how outputs are checked, what data the system can access, when human review is required and how incidents are recorded. It also means looking at whether the company can explain what happened after a disputed output, because the ability to reconstruct an incident says a great deal about how well the system is understood and managed.
Third-party dependencies matter as well, particularly as more companies build important processes around a relatively small number of AI providers. The underwriting question is not simply which platform a company uses, but why it uses it, how critical it is to the business, what happens if the underlying model changes and where responsibility sits if the failure originates outside the insured's own systems.
Liability becomes harder as autonomy increases
The technology also complicates some familiar liability questions. Foreseeability and causation can become harder to establish when a model behaves in ways that are difficult to predict, or where several different parties sit between the design of the system and the eventual harm.
That is why AI underwriting cannot rely on insurance expertise alone. Specialist legal and technical input matters because the liability chain can involve developers, model providers, deployers, operators and end users, with no single party controlling the whole process.
This has been described as the many hands problem, and it is directly relevant to underwriting because responsibility can become fragmented across the chain. Understanding that structure before a claim occurs is more useful than trying to untangle it for the first time afterwards.
Frontier risks are often underwritten before the data is complete
AI is not the first market where insurers have had to work with incomplete information. Digital assets, commercial space, fintech and other frontier sectors all developed before carriers had decades of reliable loss history, so underwriting had to place more weight on how the business operated and which controls materially changed the risk.
In digital assets, that can mean custody structures, key storage, signing thresholds and operational controls, while other sectors depend on different technical features. The principle is the same because the underwriter is trying to understand where loss can occur, how the insured manages that exposure and where the remaining uncertainty sits.
That experience matters because underwriting a new sector is not about pretending uncertainty does not exist. It is about identifying which uncertainty can be investigated, which can be controlled and which should still result in narrower terms, lower limits or a decision not to write the risk.
Waiting does not mean avoiding the exposure
There is also a practical problem with simply waiting for better claims data because AI exposure may already sit inside existing insurance books. Professional services firms, technology companies and other businesses are using AI inside activities covered by traditional policies, even where the wording was not designed with those exposures specifically in mind.
That makes inaction less neutral than it first appears. If AI exposure is already present across cyber, technology E&O, professional liability, casualty and other lines, then waiting for claims to develop without understanding where the exposure sits may simply mean learning about the risk after the loss has happened.
The better approach is to make the exposure explicit, understand how the insured is using the technology and decide what the carrier is prepared to cover based on the evidence available. In some cases that may mean narrower wording or lower limits, while in others the quality of the controls and the insurer's own experience may support a broader position.
Better underwriting matters more than a new product label
AI does not necessarily need an entirely separate insurance market. In many cases the exposure will sit within existing cyber, technology, casualty or financial lines, which means the bigger challenge is improving the way those risks are assessed rather than creating a new policy for every use of the technology.
The insurers that get better at this will be the ones that combine direct claims experience, knowledge from other frontier sectors, detailed scrutiny of controls and specialist technical or legal expertise where it is needed. None of those things removes uncertainty, but together they provide a much stronger basis for underwriting than waiting for a complete market-wide loss history.
Claims data will become more useful as the market matures, but the absence of a perfect claims triangle should affect how much risk an insurer takes, not whether it can distinguish one AI risk from another.
