The scarce asset is someone who can answer for it.
Insurable
AI
A brief wealth window opens for the humans who can carry the liability. Then the condition of coverage reverses.
The most consequential question about AI over the next decade may be the one an underwriter asks: What, exactly, are we willing to insure?
An AI may diagnose, recommend, negotiate, purchase, or operate. When its action causes harm, the injured person needs someone who can answer for it. In ordinary liability arrangements today, that means people and legal entities: a physician, a hospital, an operator, a developer, a company. The model itself does not supply the accountable balance sheet.
That gap creates an opportunity. In the early years, the human who is both willing and able to stand behind an AI can turn a technical capability into something customers and institutions can actually use. Their signature, judgment, assets, and access to coverage become scarce economic inputs.
My forecast is that this differentiation has a short life—roughly half a decade. I use 2026–2030 as the transition window and 2031 as a marker for the next phase. Those are forecasts, not actuarial findings or scheduled changes in law. The important prediction is the reversal: first, you must be insurable to use AI; later, in applications where the evidence warrants it, you may have to use AI to remain insurable.
The same gate. The opposite condition.
Change the condition.
Select a phase to see how the underwriting question changes. This is a conceptual forecast.
A shift in coverage conditions and the allocation of liability. Responsibility still belongs to people and legal entities.
01 / THE BOTTLENECK
Being capable is not enough.
A demonstration establishes that a system can do something. A deployable service must also establish who authorized it, what it may do, who can stop it, what happens when it fails, and who pays. My argument is that insurability will organize much of the distance between those two achievements.
Insurable, in this sense, means more than finding a policy with “AI” in its description. It means that the actual activity, actual operator, and actual kind of loss fit a financially credible arrangement. A performance warranty for an inaccurate forecast is different from cover for a patient’s injury. A hospital’s medical malpractice exposure is different from a vendor’s contractual promise.
Insurance also transfers specified financial consequences; it does not erase the underlying duty or make every action lawful. Regulatory permission, professional standards, insurance coverage, and reimbursement for a service are separate decisions. A system can satisfy one and fail another.
This makes insurability a powerful gate to deployment, alongside law, evidence, and patient acceptance. It is not a universal legal requirement for every AI use. It is a practical constraint wherever buyers, institutions, lenders, or regulators need a credible party to bear the loss.
02 / THE FIRST FIVE YEARS
The wealth belongs to the bridge.
Imagine an AI that can perform valuable work, but whose losses are too uncertain for a carrier to cover broadly on acceptable terms. Someone still has to decide whether to put that capability into service.
A qualified human or institution may make that possible by accepting a defined exposure, retaining some risk, obtaining limited coverage, and putting enforceable controls around the work. A buyer can then deal with someone who has obligations and something at stake.
That is the early wealth opportunity: the ability to make an otherwise unusable capability commercially accountable. It may belong to a physician group, a specialist operator, a well-capitalized firm, or an entrepreneur with deep domain knowledge and a credible route to insurance. “Willing” is only half the qualification. “Able” means the competence to manage the exposure and the resources to survive it.
A person lending a name to decisions they cannot evaluate does not create much protection. Someone who guarantees unlimited losses without adequate assets has not solved the problem either. The valuable bridge joins authority, judgment, operational control, and financial capacity.
The returns may come through ownership of the service, access to customers, contracts, and the ability to deploy while competitors remain stalled. But a serious loss can overwhelm those returns. Risk bearing earns a premium only when the risk is understood and bounded well enough to carry.
This is a thesis about a temporary scarcity. It is not a promise that taking on liability creates wealth.
03 / WHAT EXISTS ALREADY
The market is learning in pieces.
It would be inaccurate to say insurance companies cannot price any AI risk until 2031. Munich Re already offers aiSure coverage for specified AI-related exposures, including performance failures and associated losses. Armilla describes specialist underwriting based on technical review of individual systems, with coverage for named model and agent failure modes. Its public description makes clear that the issued policy wording controls.
Physical harm is not categorically outside the market either. HSB advertises AI liability coverage that can respond to certain bodily injury and property damage claims. These offerings establish that some AI risk is being insured today. They do not establish that a particular hospital, autonomous surgical system, or unlimited use case is covered.
The harder problem is making coverage broad, repeatable, and durable while the technology keeps changing. A 2026 Harvard Data Science Review paper on AI insurance identifies sparse loss history, changing model behavior, and correlated failures as central underwriting challenges. Its framework is a proposal for developing the market, not evidence that those problems are solved.
A benchmark score describes performance under test conditions. Underwriting must connect actual use to claims: how often harm occurs, how severe it is, whose conduct contributed, and how much of the loss falls within the contract. A model update may change that picture. A failure may become a claim years later.
The early window is therefore uneven. A narrow, stable application may already be insurable. A system allowed to improvise across many high-stakes tasks may remain difficult to cover long after 2031. Five years is my estimate of the broad commercial window, not a common clock for every risk.
04 / THE HOSPITAL TEST
First: who will answer for the AI?
Consider a hospital evaluating an AI diagnostic service. Its immediate question is not simply whether the model is more accurate on a published dataset. It is whether the service works for this patient population, with these devices, staff, procedures, and escalation routes—and who bears the consequences when it does not.
In the first phase, the hospital and its physicians need a defensible way to use the system while remaining within their coverage arrangements. A vendor may accept responsibility for a defined system failure. The hospital may remain responsible for implementation. A physician may remain responsible for clinical choices. The allocation depends on the facts, the contracts, the policy, and the jurisdiction.
The human bridge already has exceptions that point toward the later phase. Digital Diagnostics has publicly described accepting legal liability for its AI performance. That is the manufacturer taking responsibility for an autonomous diagnostic function, not the software becoming a legal person. A public statement is also not a substitute for inspecting the actual agreement.
The American Medical Association’s policy calls for liability to align with the parties best able to understand and mitigate the risks, and for developers of autonomous clinical systems to accept liability for system failures or misdiagnosis. That is a professional policy position, not a blanket rule that governs every lawsuit.
Surgery raises a different exposure. Assisting a surgeon, planning a procedure, and autonomously performing an operation are different activities with different opportunities for intervention. Evidence from diagnostic screening cannot establish the safety or insurability of autonomous surgery.
The FDA evaluates AI-enabled medical devices through device-specific regulatory pathways. That process does not, by itself, settle who will insure a hospital’s deployment or whether that deployment should become a required safeguard.
EVIDENCE CHECK / BENEFIT IS SPECIFIC
A promising trial is not a universal mandate.
The randomized MASAI mammography study provides a useful example of the evidence that can move this market. Its screening-performance analysis reported cancer detection of 6.4 per 1,000 in the AI-supported group versus 5.0 per 1,000 with standard double reading.
The 2026 interval-cancer analysis reported 1.55 versus 1.76 interval cancers per 1,000. The ratio was 0.88, with a 95% confidence interval of 0.65–1.18. It met the study’s non-inferiority criterion; it did not establish a statistically significant reduction in interval cancers. Radiologists remained involved.
Those distinctions matter. This supports a particular screening workflow. It does not prove reduced malpractice losses, reduced mortality, or the safety of every medical AI. The route to mandatory use requires evidence fitted to the specific harm an insurer is trying to prevent.
05 / THE REVERSAL
Later: who will answer for not using it?
Now move the hospital forward. Suppose repeated studies and local experience establish that a particular AI safeguard reliably prevents serious errors. Its limits are understood. Failures are monitored. The vendor stands behind its part of the service. The insurer can compare the outcomes of otherwise similar practices with and without it.
At that point, the underwriter’s question can reverse. A practice declining the safeguard may present the greater preventable risk. Favorable terms could become contingent on using it; eventually a carrier could decline the specific exposure without it.
That is the strong version of the forecast: for a defined activity, you can only get insurance if you use the qualifying AI. It is a proposed destination, not a claim that insurers impose a universal requirement today. The condition would concern a validated system in a specified workflow, not a general instruction to “use AI.”
The clinical standard of care could evolve in parallel. A 2026 npj Digital Medicine legal analysis discusses the possibility that failure to consult an established AI tool could eventually constitute malpractice. That is an emerging legal argument. A court’s finding of negligence and a carrier’s decision about coverage remain different things.
Using AI would not require blindly following every recommendation. A defensible system still needs exceptions, clinical judgment, outage procedures, and a record of justified overrides. The relevant safety measure may be the combined performance of the professional and the AI.
06 / THE EXPIRING ADVANTAGE
The bridge loses its scarcity.
As carriers develop reliable ways to evaluate the exposure, more organizations can obtain backing. Vendors can offer defined warranties and indemnities. Buyers can compare documented controls and coverage. The human who once made deployment possible faces competitors who can now secure similar arrangements.
This is why the wealth window closes. The person’s ability may still be valuable, but their mere willingness to carry the uncertainty no longer separates them. What was an unusual personal qualification becomes part of the expected operating system of the business.
The founders who do well through this transition will have converted the early advantage into something more durable: customer relationships, effective operations, lawful access to outcome data, domain expertise, capital, or ownership of the service. An early willingness to sign is a weak long-term moat.
The differentiation may disappear within half a decade even while the underlying business grows. Those are compatible outcomes. The bridge earns its greatest scarcity value before others learn how to cross.
07 / BEYOND THE HOSPITAL
Coverage shapes the product.
Self-driving vehicles offer a concrete example of insurance acting as a deployment gate. The UK Department for Transport’s self-driving pilot guidance requires evidence of suitable insurance for the use case, including specified third-party and automated-vehicle liabilities. This is a bounded transport framework, not a medical precedent. But it shows that an autonomy proposal must include a credible loss-bearing arrangement.
My inference is that similar economics will favor AI services that can define and document their actions. An agent that can stop at an authorization limit, preserve a decision record, and escalate uncertainty is easier to present as a bounded exposure than one granted unlimited discretion.
Insurance can therefore influence how AI evolves: which tasks become products, how much autonomy they receive, what controls are built in, and which organizations can offer them. A technically superior model may lose commercially to a service whose risks can be carried.
There is a counterforce. Sharing one model across thousands of businesses can concentrate failures. EIOPA’s September 2026 analysis warns about external-provider dependence, concentration, interconnectedness, and amplified operational risks. Improving an individual system does not automatically make an entire portfolio easier to insure.
The market may consequently require tighter limits rather than broader autonomy. Some losses may remain excluded. Some organizations may retain risk themselves. Insurability is neither a certificate of perfect safety nor a guarantee of compensation for every loss.
HOW TO TEST THE THESIS
Watch the conditions.
- Who stands behind the action?Does responsibility move from an exposed professional toward a vendor, an operator, or an explicitly backed service?
- What evidence changes the decision?Do real outcomes and claims make a defined workflow easier to cover, across more than one institution?
- What becomes compulsory?Does an insurer move from permitting a safeguard to requiring it for the covered activity?
- What stops being scarce?Can ordinary competitors obtain the backing that once distinguished a few people?
The forecast weakens if reliable outcome evidence fails to accumulate, losses remain too correlated, or AI creates more harm than it prevents. A calendar alone cannot produce the reversal.
The opportunity has an expiry date.
First, insure the human.
Then, require the AI.
The first phase rewards people who can carry consequences that the market cannot yet comfortably distribute. The next phase reduces that advantage by making those consequences measurable, transferable, and increasingly routine.
Eventually, in the activities where AI proves to be the safer arrangement, refusing it can become the exposure no one wants to carry.
Follow what is insurable. It will tell you where AI can become a business—and when the humans who opened that market are about to lose their differentiation.