Rune Kvist and the weight of a promise

Rune Kvist flew from San Francisco to London with certificates that could have travelled perfectly well by email. AIUC-1, his company’s standard for AI agents, had acquired a decidedly physical form: engraved aluminium, packed in Pelican cases, delivered by hand. In an industry busy making text almost free, he had found a reason to carry a suitcase.

The recipients had put their software through independent scrutiny. Kvist wanted the result to hang where customers and security teams could see it. A certificate was a small object carrying a large claim: someone had looked closely enough to put a name to the judgment. The flight made the gesture harder to mistake for another attachment awaiting deletion.

It is an unusually tangible introduction to the cofounder and CEO of the Artificial Intelligence Underwriting Company. His work concerns agents, models and risks that resist easy description. His chosen delivery method involved metal and a suitcase. The contrast supplies a useful clue to his thinking. New capabilities need familiar reasons for people to trust them. Sometimes those reasons begin with someone showing up.

A long way from the farm

Kvist grew up on a farm in Denmark and went on to Oxford, where he studied Philosophy, Politics and Economics. Before his work involved the institutions around artificial intelligence, he was already interested in who gets through the doors of existing ones. He helped found Project Access, a nonprofit connecting university applicants with people who could explain how admission actually worked.

By January 2017, while finishing his degree, he was leading the organisation. Its concern extended beyond the application form. A student could have the ability to thrive at Oxford or Harvard and still decide that applying was for somebody else. Sharing the experiences of ordinary, capable students could make an unfamiliar university seem reachable.

The distinction matters to his story. A formal opportunity and a usable opportunity are different things. Information, introductions and encouragement can determine whether somebody acts. Project Access addressed the space between being allowed to apply and believing an application was worth making. That is a modest task beside the rhetoric of changing education, and a rather concrete one.

Years later, Kvist would be working on a different gap between possibility and action: what happens after software becomes capable enough to use, but before an organisation feels able to entrust it with real work. There is no need to pretend the two ventures were parts of a master plan. The recurring interest is visible in the work itself.

Publishing the instruction manual

At Aula, an education software company, Kvist became chief operating officer. The role put him inside the daily business of making an organisation function. His career included several years there before his early work at Anthropic, as well as a cofounding role at the generative AI company Momentum. The route to underwriting was through operations and products.

One small episode from Aula is revealing. In May 2019, he made its handbook public. It covered remote friendships, feedback and communication among colleagues who did not meet in person. Readers had seen the company’s anonymous employee engagement survey and wanted to understand how the place actually worked. He answered by opening the manual.

A handbook is an unfashionably literal response to a question about culture. It lets people inspect the machinery beneath the adjectives. How do colleagues disagree? How does information travel? Where do friendships find room? Kvist presented the document as something that would grow with the company, making it a record of a working organisation rather than a polished monument to one.

He later wrote that Aula’s platform had seen tenfold user growth in 2020 and shared a personal review of the experience. His description included finding his “water wings.” It is an appealingly ungrand image for an operating executive: staying afloat while the pool gets considerably larger.

The paper that changed his address

In late 2021, Kvist read the scaling laws research associated with Jared Kaplan and colleagues. A month later, he moved to San Francisco. The idea had a particular attraction for somebody trained to think about economics: if more resources produced predictably better models, capital would recognise the opportunity and arrive in quantity.

He later recalled arriving in a city he had never visited, looking for the people behind the work. Anthropic was still a small lab. Coffee and introductions eventually led him to Dario Amodei. The company was considering fundamental questions about deploying its models, making revenue and engaging with the world beyond the lab.

Kvist became Anthropic’s first product hire. That places him at a specific junction in the company’s early life: where research had to meet users and a commercial offering. He was looking at what the technology might become while helping think through how it would enter the world. The questions were practical and philosophical at once.

The move also helps explain why his later interest in insurance should not be mistaken for indifference to AI’s capabilities. He had rearranged his life because he expected those capabilities to matter. Once a model can do useful work, a further set of questions arrives: who will authorise its use, on what evidence, and with what consequences if it fails?

His brother-in-law brought the other half

At AIUC, Kvist works with Rajiv Dattani, his cofounder and brother-in-law. Dattani brought a different professional background: a former McKinsey insurance partner, a former chief operating officer at the AI evaluation organisation METR, and a member of METR’s board. Between them were experience inside a frontier lab and experience with the business of risk.

The pairing gives the venture a useful division of knowledge. Evaluating an AI system requires understanding the technology. Making an assessment matter to an insurer requires understanding how risk becomes a contract. Putting the two conversations together is the company’s job. A promising demonstration has to survive questions that a sales presentation would prefer to leave until later.

Kvist’s public writing supplies another connection: he serves on the board of the Center for AI Safety. His interest in what powerful systems might do has therefore taken both nonprofit and commercial forms. At AIUC, the commercial proposition is that enterprises need evidence about agents’ behaviour and a way to manage the consequences of failures.

“Progress requires security.”

Rune Kvist

The sentence is short enough to fit on the certificate case. It is also his answer to a familiar argument about whether security work obstructs technological progress. His view is that confidence helps people adopt useful systems. He is trying to build a business around that relationship.

What sits behind the trust mark

AIUC-1 makes that proposition specific. Its published framework addresses data and privacy, security, safety, reliability, accountability and societal considerations. These are different questions, even when they end up under the same badge. A system can produce a useful answer while handling information poorly; it can follow instructions in ordinary use and behave differently when someone deliberately tries to divert it.

The standard is maintained on a quarterly cycle. That rhythm is part of the proposal: the surrounding rules need to change as systems and their uses change. A certificate has to relate to a defined assessment, rather than suggest that software has been awarded permanent good character.

In July 2026, Kvist was among the coauthors of Underwriting the Agent Economy, a research report laying out an AI insurance stack. It covered the less glamorous necessities: incident data, catastrophe modelling, contracts, risk selection, pricing, monitoring and claims. The list is a reminder that an insurance market needs more than a confident opinion about the future.

In September, he and his coauthors Cristian Trout and Dattani developed a further proposal: a mutual insurer owned by frontier AI companies. Members would contribute capital and share exposure to one another’s failures, giving them a reason to insist on common standards and independent evaluations. They presented it as an institutional design, with inspections and incentives doing work that voluntary assurances alone might struggle to accomplish.

A bigger cheque, a longer assignment

On September 15, 2026, AIUC announced a $40 million Series A led by Ribbit Capital, with participation from First Harmonic and Terrain. Added to its $15 million seed round, that brought disclosed funding to $55 million. The stated next step was extending audits, standards and insurance from agents to frontier models.

The announcement named Cursor, ElevenLabs, Harvey, KPMG, Lovable, UiPath and Fin among the builders certifying against AIUC-1. It described business-specific testing across 5,000 combinations of risks and attacks, and input from more than 250 security and risk leaders. For Kvist, the undertaking had moved beyond explaining why somebody might eventually want such a service.

His next ambition is measured in a different unit. In a public account of his work, he described hoping this would become his last job, a mission he and his colleagues could pursue for decades. He also described the company’s values in terms of accountability and openness about mistakes. These are his stated intentions, and they set a demanding standard for a business selling trust.

The aluminium certificate makes a fitting object to keep in mind. It can be held, carried and hung on a wall. Everything that gives it meaning requires continuing work: an assessment, a judgment, someone willing to answer for it. Kvist has chosen a career in that continuing work. The suitcase is merely the part that fits in an overhead locker.