The scenarios sound like fever dreams from a cyberpunk novel: an AI agent draining cryptocurrency wallets by communicating in Morse code. Another wiping production databases through a code editor interface. Autonomous systems executing financial transactions that fall into a coverage netherworld—too novel for traditional policies, too risky to ignore.
Yet these aren't thought experiments. They're the kind of incidents that have already started piling up as companies push AI agents from controlled environments into the wild. And they've spawned a question that's proving remarkably difficult to answer: when an autonomous system screws up, who actually pays?
Enter Klaimee, a San Francisco insurtech that emerged from Y Combinator in Spring 2026 with what it bills as "AI Agent Insurance, Certification & Guarantee." There's just one hitch. Visit the company's homepage, and the message is blunter: "Insurance coming soon."
What Klaimee is selling today isn't quite insurance—not yet, anyway. It's something more like a stopgap: a risk evaluation framework that grades AI agents on a five-point scale, issues a procurement-friendly certificate, and bundles in a financial guarantee of undisclosed size. The actual liability insurance product that AI founders are hunting for? Still being assembled.
Certification as a Placeholder
The current offering works like this: Submit an AI agent for evaluation. Within 24 hours, Klaimee returns a letter grade—A through F—based on eight risk dimensions. Agents scoring B or better qualify for "Klaimee-backed liability insurance" once that product materializes. C grades might earn conditional certification with carve-outs. D and F grades don't certify at all.
The evaluation framework zeroes in on scope violation, data exfiltration, unauthorized actions, output integrity, adversarial manipulation, behavioral instability, model drift, and operational control failure. According to methodology documents the company has published, these scores map directly to future premium tiers—assuming, of course, that those premiums eventually exist.
CEO Ines Boutemadja, who previously ran operations at SafetyWing, has spent recent months in what she describes as relationship-building with U.S. carriers and Lloyd's syndicates. In a LinkedIn post from early May, she detailed the process of lining up underwriters for what she characterizes as "dedicated E&O for AI agents."
But the timeline gets fuzzy. A Launch YC post from around the same period claims Klaimee is "underwriting our first cohort now" and lists "liability insurance" as a current product. The homepage tells a different story. A subsequent LinkedIn article from Boutemadja references "design partners"—startup speak for early-stage customers willing to test something that isn't quite finished. The discrepancy isn't trivial. One message suggests the company is moving toward policy issuance; the other suggests they're still figuring out what those policies should look like.
The Market Klaimee Is Entering
The AI insurance landscape has gotten crowded, fast—perhaps faster than Klaimee's marketing materials acknowledge.
AIUC and ElevenLabs publicly claimed "first-of-its-kind AI Agent insurance" on February 12, pairing certification from AIUC-1 with backing from insurers they declined to name. Corgi Insurance, a full-stack carrier, rolled out modular AI coverage in early May—the same week Klaimee's YC batch went public—covering biased outputs, generated content harms, adversarial attacks, and autonomous system failures as an add-on to their Tech E&O product.
Armilla AI has been marketing "purpose-built AI insurance for generative AI and AI agents" with limits reaching $25 million per policyholder, backed by surplus lines carriers and Lloyd's, according to materials circulating since January. Redberry Labs offers both first-party and third-party coverage addressing data restoration, business interruption, incorrect funds transfers, and mis-selling scenarios. Grigio pairs real-time risk monitoring with coverage "tied to actual behavior profile" for voice and chat agents, with integrations for platforms like Bland, Vapi, and ElevenLabs.
Then there's RiskHelm, which takes a narrower approach: first-party indemnity for "execution errors" when an agent oversteps its authority. No third-party liability. No hallucination accuracy coverage. No protection if a vendor's infrastructure goes down. And Testudo launched generative AI liability coverage tied to the 2026 exclusions that reshaped the insurance market starting in January.
Those exclusions changed everything.
The Coverage Gap That Opened on New Year's Day
In January 2026, ISO/Verisk AI exclusions entered effect—endorsements CG 40 47, CG 40 48, and CG 35 08—that allowed carriers to carve out AI-related claims from standard Commercial General Liability policies. Trade publication Business Insurance reported in April that insurers and brokers were still adjusting to the shift, months after it took effect.
The practical impact was immediate. Standard policies that might have covered an AI agent's mistakes under broad "technology services" language suddenly had explicit carve-outs. Companies deploying autonomous systems discovered gaps in coverage they hadn't known existed. Boutemadja wrote in a May blog post addressing coverage gaps that existing E&O or cyber policies mostly don't cover AI agents—a striking claim, though one that depends heavily on how individual policies are written.
Regulatory pressure is mounting from another direction. The EU AI Act has begun phasing in enforcement, with governance and transparency rules for certain system categories kicking in by August. While some reporting suggests the schedule for high-risk classifications has been adjusted, compliance obligations are stacking up for any company deploying autonomous systems across borders. The timing is forcing companies to address insurance gaps they might otherwise have ignored.
What a Certificate Actually Buys You
Klaimee's approach targets a specific friction point in enterprise procurement. Risk officers won't sign off on deploying an AI agent without some form of coverage or guarantee. Vendors can't close contracts without satisfying those risk officers. Certification with a financial guarantee removes one blocker, even if full insurance remains aspirational.
The company hasn't disclosed the guarantee's limit, who backs it, or how claims get processed. No carriers, MGAs, or reinsurance partners are named publicly. No sample policy language. No limit ranges, no jurisdictional availability, no deductibles. Just a promise that the numbers will be sorted out later.
What Klaimee does offer is speed—24-hour turnaround from application to certificate—and a grading system that doubles as pre-underwriting for the insurance product they're still building. The methodology emphasizes kill switches, logging, tightly defined scope, human-in-the-loop controls for high-impact actions, adversarial testing, and third-party validation. These are sensible risk controls, though hardly revolutionary in the AI safety world.
In a LinkedIn analysis of recent incidents—including the Grok wallet drain via prompt injection and a Cursor-related database deletion—Boutemadja positioned Klaimee as the solution to coverage gaps that let those losses slip through. It's compelling framing. Whether the company can actually deliver on it is another question.
The Uncomfortable Questions
The market is moving fast enough that no single player can credibly claim to have gotten there first. Between February and May, at least half a dozen offerings emerged targeting AI agent liability, each structured differently, with different coverage triggers and different capacity sources.
Klaimee's certification-first strategy might solve the immediate procurement bottleneck while the insurance mechanics come together. Or it might just be a temporary workaround in a market where several competitors are already writing actual policies. The company's two-person team hasn't disclosed funding, customer names, or how many policies—if any—have been bound.
For founders deploying agents into production, the calculus is blunt. Certification beats nothing. A guarantee beats no guarantee. But limits and terms matter, especially when something actually breaks.
And in a market where ISO exclusions have already redrawn the boundaries of what standard policies will cover, the question isn't whether AI-specific coverage is necessary anymore. It's which product will actually pay out when the agent goes sideways—and which will find a reason not to.
