The Artificial Intelligence Underwriting Company has raised a $40 million Series A led by Ribbit Capital, with First Harmonic participating, according to TechCrunch. The round is notable less for its size than for its category: the company underwrites AI agents. Its founders include an early Anthropic hire and a former METR chief operating officer, and the pitch is essentially that autonomous agents now need the thing every risky activity eventually needs — someone willing to price the risk and stand behind it.
Key facts
| Fact | Detail |
|---|---|
| Question answered | AI Agents Are Getting Underwritten: AIUC Raises $40M to Insure Them |
| Short answer | The Artificial Intelligence Underwriting Company raised a $40 million Series A led by Ribbit Capital to underwrite AI agents. |
| Reading time | ~3 min |
| Last updated | 2026-09-16 |
Why underwriting is a bigger deal than it sounds
Underwriting is boring until you notice what it forces. An underwriter cannot price a risk it cannot observe. To cover an agent, you need answers to questions that most agent products currently cannot answer cleanly:
- What exactly is this agent permitted to do?
- Under what conditions does it act without a human in the loop?
- What record exists of what it actually did, step by step?
- If something goes wrong, can the sequence be reconstructed?
Those are auditable properties, not marketing claims. A vendor cannot satisfy them by writing “trusted” or “secure” on a landing page.
That is the real signal here. The moment money depends on agent behaviour being describable, the describability stops being optional. And once a handful of large buyers start asking for it, it shows up in procurement checklists for everyone else — the same path SOC 2 and similar attestations took.
The three properties that keep coming up
Strip away the jargon and the same three requirements appear in almost every serious conversation about agent risk:
Scope. An agent should hold the narrowest set of permissions that lets it do the job. An agent with access to the whole disk is not a more capable agent; it is a larger blast radius.
Approval at the point of action. A one-time “yes, you may act” is not a control. It degrades into full autonomy after the first click. The meaningful version is approval per action, where a write, a delete or a command has to be shown and confirmed before it runs.
A reviewable record. If you cannot reconstruct what an agent ran, you cannot investigate an incident, and you cannot prove that nothing happened either. Logs are what turn “trust me” into evidence.
These are not exotic engineering requirements. They are the ordinary mechanics of doing risky work under supervision, and they happen to be exactly what a desktop agent can provide and a purely cloud-hosted one mostly cannot — because on the desktop the approval prompt sits in front of the person whose files are at stake.
What this means for how you pick a tool
The funding round is a reminder that the agent market is moving from capability questions to accountability questions. Capability still matters, but it is converging: most agents can call tools and follow a plan. What will separate tools over the next year is whether their behaviour survives inspection.
Practically, when you evaluate an agent for your own machine, ask whether the answer to each of the three questions above is a mechanism or a promise. If it is a promise, you are the underwriter, and you are carrying all of the risk.
MOOGH is built around the mechanism version: permissions stay scoped to what you approve, every action is shown before it executes, and the session is recorded so you can review it afterwards. You can verify the behaviour yourself on a disposable folder before trusting it with anything important — the Windows build is at https://agent.aimoogh.com/download.
Related reading: