What Happens When Underwriting Gets Cheap?

Aug 12, 2026

Artificial Intelligence May Change Competition More Than Demand

By James W. Moore

 

Key Takeaways

  • AI’s most important effect on insurance may not be labor displacement but capacity displacement: the creation of underwriting capacity faster than insurable exposure grows.
  • The first-order effect, lower expense ratios, is real but temporary as a competitive advantage. Once every carrier captures the same savings, the savings stop differentiating anyone.
  • Newly created capacity has several plausible destinations: previously uneconomic markets, deeper intelligence on existing risks, and a changed competitive battlefield once AI capability becomes universal.
  • None of those destinations fully explains what is happening. AI does not eliminate uncertainty. It changes where uncertainty lives.
  • The insurance industry has already lived through one version of this story, when alternative capital flooded the reinsurance market in the 2010s. The same mechanism is pressuring that market again today.
  • If uncertainty moves, capital follows. The consequences of AI will not stop at the primary insurance layer.

Last week I argued that insurance isn’t coal. Jevons Paradox, the observation that greater efficiency can increase total consumption of a resource, explains the modern AI boom remarkably well: falling inference costs have produced exploding usage, exactly as William Stanley Jevons would have predicted. But insurance demand doesn’t follow underwriting efficiency. It follows exposure, and exposure grows on its own timetable. Swiss Re Institute expects global non-life premiums to grow just 0.6 percent in real terms in 2026, against a long-term trend of 3.6 percent. Making underwriting cheaper does not change that timetable.

Suppose that argument is correct. It settles one question and immediately raises a harder one.

Imagine AI dramatically improves underwriting productivity over the next decade. Not incrementally, but structurally: submissions triaged in minutes, unstructured data read at scale, risk assessments produced faster than any human team could produce them. Underwriting capacity, the industry’s ability to evaluate and price risk, grows several times over. Meanwhile, insurable exposure grows a few percent a year, the way it always has.

Where does the extra capacity go?

That question is different from the one the industry usually debates. The usual debate is about labor displacement: how many underwriters AI will replace. The more consequential phenomenon is capacity displacement. AI doesn’t simply substitute for underwriters. It manufactures underwriting capacity, and unlike the coal-fired industry in Jevons’ England, insurance cannot simply consume more of its own product. You cannot convince society to buy five homeowners policies for one house. The capacity has to go somewhere else.

This article is a thought experiment about where.

The Obvious Answer

Start with the answer everyone already knows.

AI makes underwriters more productive, so carriers need fewer of them per unit of premium. Expense ratios improve. Margins improve, at least for a while. Every AI strategy deck in the industry contains some version of this slide, and there is nothing wrong with it. Cost reduction is almost certainly the first-order effect, and the carriers capturing it now are behaving rationally.

But first-order effects have a short shelf life in competitive markets. The interesting question isn’t what happens first. It’s what happens after every carrier has captured the same savings. Swiss Re Institute reached the same conclusion in its analysis of AI adoption this year: efficiency gains alone will not determine outcomes, because competitive dynamics are likely to pass much of the cost savings through to customers, limiting margin expansion. Cheaper underwriting, in other words, is not a strategy. It is the entry fee to the next round of competition.

So the capacity created by AI has to find another use. The rest of this article walks through the candidates.

What If Capacity Finds New Markets?

The first candidate is the most natural one. Capacity flows downhill, toward businesses that were never economical to write.

Every underwriting organization maintains an invisible threshold: the account size below which the cost of evaluating the risk exceeds the profit in the policy. Below that line sit millions of very small commercial accounts, short-duration exposures, embedded coverages, and highly customized risks that carriers decline not because the risks are bad, but because assessing them costs more than they are worth. AI moves the line. When evaluation costs fall materially, business that was uneconomic at human underwriting costs becomes attractive.

This is not hypothetical. Lloyd’s told the UK Parliament that AI-enabled enhanced underwriting is already allowing the market to offer coverage for risks that were previously prohibitive to write. The mechanism Lloyd’s describes, using AI to make large data sets analytically tractable, is precisely the threshold-moving effect at work in the market most famous for hard-to-place risk.

There is an irony here worth pointing out. Part One argued that Jevons Paradox fails at the industry level, because cheaper underwriting cannot create exposure. At the segment level, the paradox quietly succeeds. Within the population of risks that already exist but were never insured, falling evaluation costs genuinely do expand consumption. For some of the smallest risks, transaction costs are part of the protection-gap problem, and AI attacks transaction costs directly. Capacity that finds new markets is Jevons operating inside the boundaries that exposure sets.

But new markets can only absorb so much. Micro-policies carry micro-premiums. Even a dramatic expansion of previously uneconomic business is unlikely to consume capacity growing at AI’s pace. Some of the capacity must go somewhere else.

What If Capacity Goes Deeper?

The second candidate points inward rather than outward. Instead of writing more policies, carriers learn more about the risks they already write.

Today’s underwriting is, for the most part, a point-in-time exercise. A risk is evaluated at submission, priced at binding, and revisited at renewal. Between those moments, the carrier’s understanding of the risk is essentially frozen, not because continuous evaluation lacks value, but because it was never affordable. Human attention is the scarcest resource in an underwriting operation, and no carrier could afford to spend it re-underwriting in-force business every week.

AI changes that constraint. Capacity that once went into processing new submissions can go into understanding existing exposures: continuous monitoring of commercial accounts, dynamic pricing that reflects changing conditions, portfolio optimization that treats the book as a whole rather than a collection of individual bets, scenario modeling that stress-tests concentrations before losses reveal them. The technologies matter less than the economic shift they enable. Intelligence per risk, once a luxury, becomes a deployable use of surplus capacity.

Notice what this does to the nature of the discipline. An underwriting operation that continuously evaluates its book, rebalances its concentrations, and reallocates its appetite is no longer doing something that resembles manufacturing, where value comes from throughput. It is doing something that resembles investing, where value comes from selection, timing, and portfolio construction. That resemblance will matter later in this argument.

What Happens When Everyone Has AI?

The third candidate is less a destination than a consequence. Follow the first two paths to their conclusion, and assume every competitor follows them too.

Much of the industry will rely on overlapping vendors, foundation models, data sources, and infrastructure. The efficiency gains will be broadly similar, and competition will hand the savings to customers, as it always eventually does. Expense ratios converge. Vendor AI becomes table stakes. At that point, many of the traditional sources of underwriting advantage tied to processing speed and analytical capacity have been substantially equalized by technology.

The industry has seen this movie before, one layer up. Through the 2010s, alternative capital flowed into reinsurance through catastrophe bonds, sidecars, and collateralized structures. Capacity expanded faster than the demand for it, and pricing responded the way pricing responds: Guy Carpenter’s U.S. property catastrophe rate-on-line index fell nearly 17 percent in 2014 and continued grinding downward until the market bottomed in 2017.

The reinsurers that prospered through that decade did not out-muscle the new capacity. They specialized, consolidated, and competed on capital efficiency and risk selection. And the precedent is not just a lesson from history. AM Best projects the reinsurance segment entering 2026 with record capacity, roughly 540 billion dollars in traditional capital plus 120 billion in insurance-linked securities, and property reinsurance rates fell 10 to 20 percent at the January renewals. The mechanism is running again in real time, one layer above where AI is currently building capacity.

The lesson of that precedent is uncomfortable for anyone counting on AI itself as an advantage. When capacity becomes abundant, margins compress, and competition reorganizes around whatever remains scarce, which sharpens the question this article has been circling from the start. If AI makes underwriting capacity abundant, what exactly remains scarce?

Where Does Uncertainty Go?

Here is where the thought experiment stops being about capacity.

Each of the destinations above is plausible. All three will probably happen simultaneously, to different degrees, at different carriers. And yet none of them explains the bigger change, because all three quietly assume that AI is doing the same job underwriters always did, only faster and cheaper. That assumption deserves scrutiny.

Insurance exists because uncertainty exists. Policies, premiums, reserves, reinsurance, and capital standards are all machinery for bearing uncertainty at a price. So the right question about any technology that transforms underwriting is not what it does to costs. It is what it does to uncertainty.

The intuitive answer is that AI reduces uncertainty, and in places it genuinely does. Risks that could only be estimated can increasingly be measured. But watch what happens across the full landscape rather than at any single point. Some uncertainty does disappear into measurement. Some becomes newly visible: patterns in data that no human would have surfaced.

And entirely new uncertainty appears where none existed before. In a recent survey of 600 corporate insurance decision-makers, the Geneva Association found that 71 percent had already deployed generative AI in at least one function, that more than 90 percent were interested in insurance protection for the resulting risks, and that two-thirds would pay meaningfully higher premiums for it. AI is not draining the pool of uncertainty. It is simultaneously draining one end and filling the other.

Some of the new uncertainty is exotic precisely because it has no loss history. Swiss Re Institute notes that underwriting the enormous next-generation data centers being built for AI depends on specialized technical assessment rather than empirical experience, because almost no empirical experience exists. And some of the new uncertainty is structural rather than novel. The Financial Stability Board warned in 2024 that widespread reliance on common AI models and data sources tends to produce correlated predictions across institutions.

Lloyd’s raised the same concern to Parliament: a market in which many insurers depend on the same cloud providers and similar large language models is a market with a new form of concentration risk. Follow the logic into underwriting, and it points somewhere uncomfortable. Each carrier’s individual risk assessments may look better than ever while the market’s errors become synchronized. Uncertainty has not vanished. It has moved from the individual account level, where it was priced, to the systemic level, where it is not.

Economists have understood for fifty years that insurance markets are shaped less by risk itself than by who knows what about it. Rothschild and Stiglitz built the canonical model of insurance under information asymmetry in 1976, showing that who gets served, at what price, and whether a stable market exists at all can turn on what each side knows. AI represents a potentially profound redistribution of risk information, but it will not make information symmetric. Someone will always know more, sooner. Asymmetry does not disappear. It shifts, and advantage shifts with it.

This, I think, is the real answer to the question the article opened with. Capacity is going wherever uncertainty went. The carriers deploying surplus capacity into micro-markets are pursuing uncertainty that became affordable to measure. The carriers going deeper on existing books are pursuing uncertainty that became visible inside their own data. When AI capability converges, the assets that still differentiate, proprietary data, feedback loops that improve models faster than the market improves, the portfolio judgment of underwriting-as-investing, are all instruments for reducing one’s own uncertainty faster than competitors reduce theirs. And the emerging risks nobody yet knows how to price are simply uncertainty at its newest, waiting for the first organization confident enough to underwrite them.

AI doesn’t eliminate uncertainty. It changes where uncertainty lives. The strategic race of the next decade may belong not to the organizations that automate underwriting first, but to the organizations that discover where uncertainty moved before anyone else does.

What Happens to Capital?

One consequence follows so directly that it deserves its own brief mention.

Everything in an insurance organization’s financial architecture is calibrated to a particular map of uncertainty. Pricing reflects what is unknown about individual risks. Retentions and deductibles reflect what policyholders and carriers each prefer not to know the cost of. Reinsurance structures reflect where carriers believe their own knowledge ends. Capital allocation reflects all of it at once. Redraw the map, and every one of those calibrations comes up for review. If risks that once required broad risk transfer become measurable, retention economics change. If new, poorly understood accumulations emerge, demand appears for protection that does not yet exist.

Working out what that means is not a task for this article. It is the task the whole industry’s capital providers are about to inherit.

Conclusion

Let’s return to the question we started with. AI will reduce underwriting expense. Everyone expects it, everyone is pursuing it, and it will happen. The harder question is the one that comes after: what to do with the capacity AI creates? Push into markets that were never economical? Go deeper on the risks already on the books? Build the data and feedback loops that survive convergence? Or look for uncertainty in its new locations before competitors find it? These are not predictions. They are options, and different organizations will exercise different ones simultaneously. The time to weigh them is while they are still choices.

And if AI truly changes where uncertainty lives, the effects cannot stop at the primary layer. The organizations that finance insurance risk will eventually have to ask what these changes mean for their own assumptions. That raises an entirely different set of questions. Not for insurers. For the organizations that insure insurers.

Those questions are where we turn next.

Sources

 

Related Posts

468