自保險公司正迅速成為 AI 保險的試驗場

The most repeated line about captive insurance right now is that AI data centers are too big for the traditional market to absorb. Marsh, the world’s largest insurance broker and captive manager, expects explosive growth in companies setting up their own in-house insurers to carry data center risk. S&P Global Ratings, the credit rating agency, has described one Meta facility as sitting beyond the limits of what can be fully insured. There are now more than 6,000 captives worldwide writing roughly $240 billion in premiums, up close to one-fifth from two years ago, according to Captive Review.

A captive is an insurance company, usually with a single customer (its own parent), a narrow book, and no retail distribution. It still needs underwriting, actuarial reserving, statutory accounting, audit, legal review, and regulatory filings. What it does not have is forty years of mainframe, a claims platform nobody dares touch, or a thousand underwriters with habits. Which makes it the cleanest environment in insurance to find out whether software can run the whole thing.

Alexandre Musy, CEO of Huscarl, which is building an AI actuary for corporations and their insurance captives, puts the implication bluntly. “Captives are small, simplified insurance companies,” he said. “You can be certain that once AI-powered captives are live, AI-powered insurance companies will be coming next.”

Musy and CTO Paulien Jeunesse previously built what Huscarl describes as the world’s first parametric cyber insurance product for corporations at Descartes Underwriting, where Musy scaled it commercially across Europe, and Jeunesse, an actuary and AI scientist, built the underlying model.

The Industry Runs on PDFs

Reporting by Captive.com in August found AI in captives is still mostly a set of single-task tools for jobs such as pulling data out of documents, reconciling records, checking policy wordings, drafting reports or supporting claims administration.

Marcus Schmalbach, CEO of the risk-transfer platform RYSKEX, told the publication the obstacle is the operating model and the data rather than the technology.

A captive can be a sophisticated risk-financing vehicle and still run on spreadsheets, email and hand-reconciled PDFs, fed by a parent company, a broker, a fronting insurer, a manager, a claims administrator, an actuary, a reinsurer and an auditor, each sending a different format on a different schedule. Lower transaction volumes than commercial insurers meant there was never much economic pressure to integrate systems, and every captive is structured slightly differently, which defeats standardization. The processes stayed manual, in his account, because they were never designed for automation in the first place.

Illia Pinchuk, CEO of DICEUS, told Captive.com the pain concentrates in the handoffs between organizations, and pointed to a program covering 100 captives and 200 policies where centralizing administration let renewals run in bulk and produced certificates on demand.

Musy is more specific about why a shared data layer keeps failing.

It “would greatly benefit the industry,” he said, but a captive drags in actuaries, auditors, accountants, lawyers, investment managers and third-party administrators, and “the more participants there are, the harder it is to implement.”

The failure mode is social rather than technical, he argued. “There will always be someone who can’t or won’t use advanced systems,” he said, so the market defaults back to “PDFs and Excel files.”

That observation carries a commercial edge, and it describes a three-way scramble rather than a duel. Software vendors are selling a data layer to incumbent captive managers. A second group of AI-native captive managers is competing with those incumbents with the layer already built. Huscarl sits in a third position, supplying the actuarial and data layer into the channel rather than replacing it: the company is expanding a network of partner captive managers and brokers, and already works with a risk retention group and a single-parent captive at a company with more than $2 billion in revenue.

“The winning path is unclear today,” Musy said, though his hedge is that the mess itself may be the moat. AI, he argues, is “a great glue for messy data workflows,” and Huscarl “probably spent 50% of our engineering work since inception managing just that,” which now lets it process messy PDFs and spreadsheets “with full automation, regardless of how the industry evolves.”

So half of an insurance startup’s engineering effort went into reading documents other people refused to stop sending.

The Part That Cannot Be Automated Yet

Where the ambition splits is on judgment, and the sharpest version of the limit comes from a vendor.

Schmalbach’s rule, as he put it to Captive.com, is that the system recommends and the professional decides.

Software can flag that a reserve moved, that a claim looks unusual, that an exposure is deteriorating, but a qualified person still has to decide what it means and whether coverage applies.

An agent may spot a claim crossing a reporting threshold and prepare the notification; it should not deny the claim, change a reserve, or approve a material payment on its own, and it should not be able to move funds outside defined approval controls.

Every action should be logged and reviewable, with a named person accountable. AI should support decisions rather than quietly make them.

Musy’s ambition runs past that stage. The north star, he said, is “AI as the ultimate orchestrator, as every decision involves underwriting, actuarial, accounting, legal and financial considerations,” something closer to “a company brain for captives that steers them towards higher value creation for their owners.”

Huscarl says every study is reviewed and signed by a credentialed human actuary, and captive domiciles generally require a signed annual actuarial opinion anyway. The autonomy is in the work. The judgment, and the professional liability attached to it, stays with a person. Schmalbach gestures at the same review from the other direction, arguing that linking captive data to the parent’s operations could surface loss patterns before they show up in the annual actuarial review. The annual review stays. The question is who feeds it.

Why It Matters Past the Niche

The economics are what pull capital in. AM Best rates roughly 150 US captives and estimates they generated more than $8 billion in savings over five years, because premiums stay on the balance sheet instead of leaving. Marsh counted 118 new captives formed in 2025, up from 92 the year before, in a year when commercial insurance pricing fell about 4%. Companies are self-insuring while insurance gets cheaper, which suggests the shift is structural rather than a reaction to a hard market.

Investors have noticed. In September, captive manager Luzern Risk raised $45 million, and Huscarl raised $5.6 million from FRST and Y Combinator, while captives once reserved for the Fortune 500 are being pitched to companies with as little as $10 million in revenue.

What Musy is aiming at is not the captive manager but the commercial market itself. His goal, he said when announcing the round, is to let corporate risk managers act as their own chief underwriting officer, in “a future where self-insurance becomes the default, and commercial insurance becomes the exception.”

None of this makes the captive market large. It makes it legible.

An industry of thousands of small, simplified, lightly staffed insurance companies, carrying risk the mainstream market will not touch, is the most permissive place in finance to let software underwrite, reserve, reconcile, and report. Whatever survives there arrives next at companies with a million policyholders and an actuarial department that has been doing it the same way since before the internet.