Your AI agent just booked a customer a refund they weren't owed, quoted a policy that doesn't exist, or handed a stranger a slice of someone else's data. Who pays for that?

For most of the last two years, the answer was nobody — because nobody would write the policy. A San Francisco startup called AIUC just raised $40 million to change that.

The Artificial Intelligence Underwriting Company — yes, that's really the name — closed a $40 million Series A led by Ribbit Capital, with First Harmonic joining in. That takes total funding to $55 million, on top of a $15 million seed led by former GitHub CEO Nat Friedman back in July 2025.

The pitch is almost boring in its simplicity: certify AI agents like an auditor, then insure them like an underwriter. The thesis underneath it is sharper — the thing blocking enterprise AI isn't intelligence, it's liability.

🧠 Why This Matters

Banks, hospitals and government agencies aren't hesitating on AI agents because the models are dumb. They're hesitating because the moment an agent makes a promise to a customer, someone is legally on the hook for it — and right now that someone is them.

Rune Kvist, AIUC's co-founder and CEO and one of Anthropic's earliest product hires, puts it plainly:

"Banks, hospitals, governments and militaries no longer decline to deploy AI because a model isn't smart enough. They decline because they've made commitments to their own customers about what a system will and won't do, and nobody can currently guarantee that."

— Rune Kvist, co-founder & CEO, AIUC (Jingletree)

That's the gap. A model vendor can tell you their agent is 98% accurate. Your compliance officer wants to know exactly what happens on the other 2%. AIUC's answer is a number with a dollar sign in front of it.

📊 Deep Dive

AIUC sells three things that stack on top of each other: a standard, an audit, and a policy.

The standard is AIUC-1, which the company describes as "SOC 2 for AI agents" — a phrase every security team already understands in their sleep. It's built on existing scaffolding: NIST's AI Risk Management Framework, the EU AI Act, and MITRE's ATLAS threat catalog. To earn it, an agent gets run through roughly 5,000 adversarial simulations — jailbreaks, prompt-injection attacks, hallucinations, data-leak attempts — all modeled on documented real-world failures. Software runs the attacks; humans sign off on the roughly 100-page report.

Then comes the part that makes it real. Pass, and AIUC will underwrite an insurance policy tied to those results. Break a covered promise, and the policy pays out.

Here's how the pieces line up against what you already know:

  • SOC 2 proves your servers are handled securely. AIUC-1 proves your AI agent behaves.
  • A model card is the vendor grading its own homework. An AIUC audit is a third party running ~5,000 attacks and putting money behind the grade.
  • Traditional cyber insurance covers the breach. AI agent insurance covers the agent's decisions.
  • The founders aren't lifelong insurance execs: Kvist came from Anthropic, co-founder Rajiv Dattani was COO of the AI-evaluations lab METR, and CTO Brandon Wang is a Thiel Fellow (Jingletree).

The first real proof point already shipped. In February, voice-AI company ElevenLabs became the first firm to earn AIUC-1 certification and buy the coverage, after putting its agents through more than 5,000 adversarial tests across six categories — data and privacy, safety, security, reliability, accountability, and societal impact.

"ElevenLabs is the first company to prove this model works at scale."

— Rune Kvist, AIUC (PR Newswire)

Reported clients now also include coding tools Cursor and Harvey and app-builder Lovable (Jingletree).

⚠️ The Catch

An insurance company is only as good as its ability to price risk — and AI agents are a risk class with almost no actuarial history. Cyber insurers have decades of breach data to lean on. AIUC is underwriting a technology that behaves a little differently with every model release. Price it too low and one bad quarter of agent failures eats the premiums; price it too high and nobody buys.

Then there's the snapshot problem. A 100-page audit describes an agent on the day it was tested. But agents get updated weekly, prompts drift, and fresh jailbreaks surface constantly — a clean bill of health in September says surprisingly little about December. And AIUC helped author the very standard it audits against, which is the same built-in tension SOC 2 auditors have lived with for years.

$55 million is real money, but it's a rounding error next to the balance sheets of the reinsurers who'd ultimately backstop any large claims. Scaling this means convincing much bigger players that the math holds.

🎯 What Happens Next

Watch two numbers. The first is the loss ratio: do the agents AIUC certifies actually stay out of trouble, or do claims start rolling in? Once that figure exists, it is the business. The second is competition — if AIUC-1 becomes the checkbox enterprises demand before switching an agent on, expect the big audit firms and insurance incumbents to want a piece, either by building rivals or buying their way in.

One tell that this is more than a novelty: Ribbit Capital's portfolio runs to Robinhood, Coinbase and Revolut. They bet on financial plumbing, not gadgets.

🧩 Bigger Picture

The industry spent two years arguing about how capable the models are. This $40 million is a bet that the next fight is over accountability — who certifies, who signs, and who writes the check when an agent gets it wrong. Every company wiring an AI agent into a real workflow eventually runs into the same question their lawyers ask first: and what if it screws up?

For a long stretch, the honest answer was a shrug. Now it's a policy number.

When AI finally has to buy insurance, you know it's grown up enough to get sued.

Sources