AI Liability: Lagging in Time, Leading in Importance

AI’s next wave of value will come from the industry that insures and governs it

I spent the first part of my career in law and the second in consulting. For a long time, I thought of them as two separate jobs. Looking back, they represented views of the same economy.

In consulting, I was able to see the innovation and business side of things. I saw the guardrails that follow in law, whether through enforcing rules and governance structures, or deciding who is responsible when something goes wrong. While business in general moves first and builds fast, you need a lagging support system that builds in safety mechanisms and accountability. 

At one point, after having done some personal research in the intersection of law and business, I realised that every technological cycle can be broken down into the same sequence of steps. Value is the initial step, and rules usually are the final. The cycles all go through three waves, in my view. First, the technology itself creates value. Second, at almost the same time, the “picks and shovels” grow alongside demand for the technology: the suppliers that the technology cannot function without, like the merchants who sold tools to miners in the California Gold Rush. Third, and always last, come the frameworks: regulation, policy, and governance. The industry that booms in that third wave is insurance.

Today, AI is the leading technology, and we are standing at the precipice of its third wave. I believe we have a duty to look closely at what is happening right now. AI insurance and AI policy are only beginning to form.

A clear pattern for how new technologies create value

Ambition drives technology into existence. Carlota Perez describes the early phase of each technological revolution as investment capital allied with young entrepreneurs: people with the belief and the incentive to build before the market exists. As the market comes into being, industries that adopt and use its technology grow, built by people driven more by financial incentives rather than pure ambition. The mature industry never exists on the day the technology is “ready.”

Railroads are the clearest example. Alfred Chandler, a Harvard business historian, described them as the first modern business enterprises. The technology of railroad tracks, engines, and attached cars came first. Professional management, modern accounting, and new ways of raising capital followed, because running a railroad demanded these tools. The railroad, then, ultimately created the template for the modern company.

Electricity is another example; it shows how long it can take before a technology’s value is actualized. In 1990, the economist Paul David showed that factories first simply replaced their steam engines with dynamos—and gained almost nothing. Productivity gains came only when owners redesigned the factory around electric power. US manufacturing productivity finally leapt in the 1920s, about four decades after electricity itself was commercialized.

The internet repeated a strikingly similar pattern. The technology (a scaled network of connected computers, first pioneered by US defense agencies), once again, came first; then, e-commerce, fintech, and online dating were built on top of its infrastructure. E-commerce did not exist at scale in 1994, but in the second quarter of 2026, it accounted for $340.2 billion, or 17.1%, of all US retail sales.

Three waves of value creation 

As the examples above show, the first wave is the technology itself: the railway, steam power, electricity, the car, the internet, AI. Technology creates value directly and disrupts the status quo, often across every sector at once. 

The second wave arrives almost at the same time as the first. Its contribution consists of picks and shovels, the suppliers of which the technology cannot run without. For AI, that is chips, data centers and energy. Their value rises at the same time as the technology’s.

The third wave always comes last. I call it a trailing edge, after the back edge of an aircraft wing: it trails behind, but it is where the flaps, the flight control surfaces, are located—the parts that modulate aerodynamics to steer the plane and land it safely. Regulation and governance lag innovation, as nobody can write definitive rules for a technology before its widespread adoption makes its consequences clear. Every new technology forces lawmakers to decide who carries its risks, which often evolves into a minefield of whose risk is pooled for what cost. In this minefield, the industry that benefits the most is insurance. Even though it does lag in time, it does not lag in importance. Often it is what allows the technology to not only scale but possess staying power.

Ultimately, the trailing edge is much of what it takes to deploy a technology at scale. Which means in turn that as AI models become more commoditized, value could shift towards complements such as insurance and governance.

Once again, history shows the pattern clearly: insurers have repeatedly built a large part of the trailing edge themselves:

New technologies and the insurance that trailed them

Leading technology	Use case	Trailing edge: the insurance that followed
Steam boilers	Powering factories, locomotives and steamboats (1850s onwards)	Boiler insurance combined with inspection in 1866, about a decade after boilers spread; insurer inspection standards become the specifications for boiler design
Electricity	Lighting and powering buildings and factories (1880s onwards)	Underwriters Laboratories in 1894, funded by fire insurers to test electrical safety, one year after the 1893 World’s Fair; first safety standard in 1903
Automobile	Personal transport (1886 onwards)	First car insurance policy in 1897, 11 years after Benz’s patent; first compulsory insurance law in Massachusetts in 1925, about 40 years after
Internet	E-commerce, fintech, and online dating (mid-1990s onwards)	Cyber insurance: $15.3 billion in global premiums in 2024, under 1% of global property and casualty premiums; about 30 years on and still not fully developed

Several conclusions follow: First, the lag is real and often longer than expected. For cars, it took about a decade before the first policy and four decades before the first mandate. The trailing edge starts small and voluntary and ends as compulsory infrastructure. Second, insurers also wrote the rules, as a result of their business model. The early boiler and electrical safety standards (see above) were private policy, written by insurers years before governments acted. Insurance and policy grow together. And third, the trailing edge can stay small for a long time. Cyber insurance, for example, is still a comparatively minor line three decades after the rise of the commercial internet.

Mapped onto AI, the aforementioned three-wave paradigm is straightforward. The technology is AI models and agents, deployed across every sector at once. The picks and shovels are chips, data centers and energy, growing alongside it. The trailing edge is 1) AI insurance, which prices and transfers the risk, and 2) AI policy, which sets the standards and liability rules that decide who carries it. However, there exists one key difference from earlier eras that cannot be ignored. In the past, if a boiler broke, it would subsequently need repair or replacement. In contrast, an AI agent today acts (autonomously). Its mistakes can repeat across countless transactions before anyone notices, which makes AI risk unusually correlated and also makes attribution—knowing what the system did and why—part of the trailing edge itself.

By now, I think it is common consensus that AI is the next general-purpose technology revolutionizing the world we live in. If the historical pattern holds, we will witness its trailing edge soon emerge: the industry that insures AI and writes its rules.

Is AI’s trailing edge already here?

AI’s trailing edge is forming now because the gap is opening from not one but two directions. Incumbent insurers are pulling back; simultaneously, neither the black letter practice of the law nor the spirited theory of jurisprudence has settled whether liability accrues with the companies that build AI or those that deploy it.

In late 2025, several large US insurers asked regulators for approval to exclude AI-related liabilities from corporate policies. Their central concern was an excess of correlation: the insurance industry can absorb a single $400 million loss, but it cannot absorb thousands of simultaneous claims caused by one AI provider’s error. Brokers are uneasy too: on October 15th 2026, the CEO of Marsh (a leading insurance broker firm) warned that broad AI exclusions could create “real confusion and uncertainty.”

New entrants are moving into the gap:

Select players in AI insurance, 2026

Player	What they do
Munich Re	Established reinsurer offering AI performance cover, and since 2026 AI liability cover for small and mid-sized businesses
AIUC	Insurance combined with an audit standard (AIUC-1) for AI agents
Harbor	Inventories a company’s AI agents, measures their risk, blocks rule-breaking actions and insures what remains

Harbor is a leading indicator of where the broader market is heading. The San Francisco company came out of stealth on October 13, 2026 with a $10 million seed round led by Base10, Abstract and 8VC. It claims its first customers are banks and insurers in the US, Europe, and Latin America. Its model follows the historical template of the trailing edge: measure the risk; reduce it; insure what is left. This is, to belabor the point, the same logic insurers applied to steam boilers in the nineteenth century.

The AI insurance market is still small. Deloitte, modelling based on cyber insurance’s early trajectory, projects about $4.7 billion in annual AI insurance premiums by 2032 at roughly an 80 percent annual growth rate. In 2025, cyber insurance reached an estimated $16.3 billion. In the future, I believe that AI insurance will be the layer that allows a multi-trillion-dollar technology to reach not only widespread adoption but also deeper market penetration.

The good news for builders is that this liability market will probably not be winner-take-all. One reason is founded on the market’s structural nature: correlated risk has to be spread across many balance sheets, which is why Lloyd’s (a global insurance and reinsurance marketplace) syndicates; why reinsurers stand behind most of the new entrants. A single winner would concentrate the risk insurers fear most.

Another reason arises from a study in how new industries gain legitimacy. Aldrich and Fiol argued in 1994 that new industries lack legitimacy above all, and that founders win it collectively by converging on shared designs and acting together. In any young industry, competitors and innovation overall are evidence that the industry is real—that it is its own category.

The last few years were about what AI can do. I firmly believe that the next decade will reward those who build what has to come after it: the insurance, the standards, the rules—the third wave, the trailing edge—that make AI safe to use.

Nora Panholzer (MBA ‘27) is from Vienna, Austria. She studied Law (LLM) and Business (BSc) at Vienna University of Economics and Business. Prior to HBS, Nora worked as a consultant at McKinsey and Company.