The insurer's new Cortex unit offers a more useful template for AI in insurance: pick a difficult risk problem, put underwriters in charge, and make the technology earn its place by helping write better business.

There is a particular kind of insurance AI announcement that has become familiar: a carrier launches a centre of excellence, signs a model vendor, shows a slick demo of a submission being summarised, and promises a future of faster decisions.

Markel has taken a different route.

In late July, the specialty insurer launched Cortex, a new business unit inside its U.S. Wholesale & Specialty division. Its stated task is narrow and commercially consequential: underwrite and service hard-to-place U.S. casualty risks with a new, AI-enabled operating model.

That distinction matters. Cortex is not presented as an innovation team trying to find problems for technology. It is a business designed to write business in one of insurance's least forgiving markets.

Markel began the project with Bain & Company in March and launched the unit four months later. On its Q2 earnings call, Markel Insurance CEO Simon Wilson said Cortex was already starting to write risks. The company says the model combines market-leading AI tools with Markel's underwriting expertise and historical data.

The claims will need time to prove out. Cortex is new, and casualty performance cannot be judged in a single quarter, let alone a few weeks. Still, the shape of the move is worth studying. It gives us a clearer view of where AI may create a real advantage in specialty insurance, and where the industry is still fooling itself.

Casualty underwriters are rarely starved of information. They are more often buried by it.

A complex submission can contain loss runs, schedules, contract wording, inspection reports, broker narratives, financial information, location data and a growing number of external signals. The hard work is deciding what matters, what is missing, what has changed, and how the individual risk fits into an already exposed portfolio.

That work is becoming harder. U.S. liability claims rose 57% over the decade to 2023, according to Swiss Re. Its Social Inflation Index found that social inflation added seven percentage points to U.S. liability-claims growth in 2023. Legal funding, aggressive plaintiff tactics and larger jury awards are changing the severity curve, particularly in long-tail bodily injury lines.

AM Best has also kept a negative outlook on parts of commercial casualty. It reported a 120% combined ratio for general liability in 2024 and pointed to reserve pressure, elevated severity and continuing caution in U.S. auto and excess casualty.

The temptation in this environment is to say that AI will make underwriting faster. It can. But speed is a poor objective if it simply allows a carrier to make bad decisions more efficiently.

The better objective is to make the first underwriting view more complete, more consistent and easier to challenge. A strong casualty underwriter should spend less time hunting through PDFs and more time asking the questions that determine whether a risk belongs on the book.

That is where the Markel example gets interesting.

Markel has been rebuilding its insurance operation more broadly. It reorganised the business into three divisions and 14 business units, each with a leader accountable for a 2026 financial plan and a five-year strategy. Cortex fits that management design.

It has a home in the wholesale and specialty division. It has a defined market problem. It has access to actual underwriting capability and capacity. In other words, the people using the AI are the people accountable for the loss ratio, broker proposition and return on capital.

This is a far better setup than the usual innovation theatre.

Most corporate AI programmes fail in one of two ways. Some sit too far from the business, producing useful but orphaned tools that do not survive the pilot. Others rush a generic assistant into a live workflow without changing the workflow itself. The tool saves a few minutes, but the commercial model stays exactly the same.

Markel is attempting something more ambitious. It is using AI to design a new operating model for a selected risk problem. It is not claiming that a model can autonomously judge an excess casualty tower. It is trying to put better information in front of a human underwriter sooner, then build the service and decision process around that fact.

There is an important commercial difference between those approaches. A technology pilot is measured in adoption, log-ins and perhaps hours saved. A business unit is measured in quality of risk, response time, quote conversion, expenses, claims outcomes and capital returns. The latter is a much less forgiving test. It is also the one that matters.

Cortex is only one part of the story. On the same earnings call, Markel described three levels of AI deployment.

First, it built a new strategic business unit from scratch. That is Cortex.

Second, it has been rewiring existing classes of business. Markel said it had reworked six classes across its U.S. and international operations, representing more than $500m in gross written premium, using Harvey AI. Management reported that the time to get an initial risk assessment in front of an underwriter had fallen by 50% to 90%, depending on the line, while internal accuracy measures exceeded 90%.

Those are company-reported figures, not independently audited results. They should be treated as an indication of operational potential rather than proof of underwriting superiority. But they are directionally credible. Document extraction, structured risk summaries, missing-data checks, referral routing and first-draft correspondence are exactly where current models are most useful.

Third, Markel launched an internal AI accelerator fund. It selected nine ideas from the business and put them into production in April. The examples were practical: an equine rating engine that produces initial quotes in seconds, a tool to respond more quickly to the marine-war market, and a system that uses broker renewal data to prioritise opportunities with the highest probability of success.

The pattern is sensible. Use a focused new unit for a high-conviction opportunity. Improve existing workflows where the friction is obvious. Give frontline teams a modest pool of capital to test smaller ideas. Central leadership funds and sets standards; business leaders choose the problems and own the outcomes.

The strongest use cases are not mysterious.

AI can read and structure submission packs. It can spot missing loss information. It can compare a submission against appetite rules, create an evidence-backed risk summary, identify changes at renewal, prepare a pricing scenario pack and draft a clear request for further information. It can also give underwriters a more useful view of emerging portfolio concentration and claims signals.

It should not decide the appetite. It should not bind a material casualty risk without a human decision maker. It should not turn an opaque score into a substitute for judgment, especially in long-tail classes where the apparent signal may be swamped by changes in law, litigation tactics, medical costs or jury behaviour.

The difference is subtle but important. AI can create a better starting point for a decision. It cannot eliminate the decision.

The best implementation will also join underwriting to claims and risk engineering. Casualty organisations have years of learning locked in adjuster notes, litigation files, inspection reports and risk-control recommendations. Much of that knowledge is hard to retrieve at the moment a new risk arrives. If AI can make that learning searchable and usable, it may improve selection, pricing, terms and client loss prevention at the same time.

That is a more defensible advantage than a chatbot sitting on top of a document repository.

The lesson from Cortex is organisational before it is technical.

Start with one selected casualty niche where the market has genuine pain and where the insurer has a reason to believe it can win. The niche might be a specialist general liability class, a complex commercial auto segment, environmental liability, construction-related risk or a part of professional liability. The answer will depend on the carrier's own claims data, broker relationships, risk-engineering capability and reinsurance structure. It should not be selected because a market segment sounds exciting.

Then give a small team a real mandate. It needs senior underwriters, actuarial support, claims input, risk engineering, product and data capability. It needs a defined appetite, controlled capacity and a clear route to the casualty P&L. Most of all, it needs the authority to change the end-to-end workflow rather than bolt software onto the old one.

Finally, measure the right things. Premium volume is a lagging and often misleading measure in the first year. Watch the time to first meaningful response, quote turnaround, data-chase volume, referral rate, appetite exceptions, underwriting-file quality, broker experience and emerging portfolio concentration. Track process accuracy separately from underwriting quality. A system that extracts 95% of a document correctly can still support a poor risk decision if the appetite is wrong.

There should be hard controls from the first day: approved data sources, evidence citations in risk summaries, output logging, review of material model errors, clear escalations for high-limit and out-of-appetite risks, and a manual fallback route. An AI system that cannot explain the evidence behind its suggestion has no place near a complex casualty decision.

The phrase that matters here is not "straight-through processing." It is the confidence to write the right risk.

Hard-to-place casualty business is hard for a reason. The risks are incomplete, the loss tails are long, and the market has repeatedly learned that apparent pricing adequacy can vanish as severity develops. More automation will not change that.

What it can change is the quality and speed of the underwriting conversation. It can surface missing information earlier. It can make appetite more consistent. It can expose portfolio drift before it becomes a reserve problem. It can give a senior underwriter more time for judgment, negotiation and risk selection.

That is the promise in Markel's Cortex experiment. Not artificial intelligence replacing underwriters. A tighter underwriting machine, built around them.

If Cortex produces strong risk-adjusted returns over time, other carriers will copy it. If it does not, Markel will at least have run the right experiment: a bounded commercial test in a real market, with real accountability. That is already a better standard than most AI programmes in insurance have set for themselves.

All performance figures attributed to Markel are company statements. Cortex launched in July 2026, so it is too early to assess its underwriting results.