AI Agent Liability Insurance: Cowbell's New SME Policy

Cowbell said on November 10 that it will insure small businesses against what their AI agents do unsupervised, from payments released without sign-off to the usage bill a looping agent runs up and the customer lawsuits that follow.

The cyber insurance provider, based in Pleasanton, California, said Cowbell AI Liability Insurance is written for US businesses whose agents act for them in customer service, payments and core operations. "SMEs are giving AI greater responsibility for how their businesses operate," said Trent Cooksley, Cowbell's co-founder and chief operating officer.

The launch comes as general liability carriers write AI out of their forms. ISO's CG 40 47 endorsement, available since January, removes bodily injury, property damage and advertising injury arising from generative AI from a CGL policy. A review of state filings found 2,369 generative AI exclusion records in force across 49 states by mid-2026 and no admitted standalone product to replace the cover, Insurance Business reported in September.

Key Takeaways

The policy carries ten first-party coverages and five third-party ones, running from incident response and business interruption to contingent bodily injury and property damage.
Two of the ten reimburse an insured's own AI usage bill: runaway cost from a looping agent and LLMjacking through stolen credentials or API keys.
A claim is triggered when an AI tool does any of five things: exceeds its authorized scope, acts without required approval, exposes data, produces inaccurate or unsafe output, or fails its controls.
Rating uses three AI-specific factors, for observability, autonomy and governance, fed by connecting the insured's own tools such as Gemini, Claude and ChatGPT.

What the policy pays for

Everything in the policy hangs on one defined event. An AI Incident, in the release's wording, is one in which an AI tool exceeds its authorized scope, acts without required approval, exposes data without authorization, produces inaccurate or unsafe outputs, or fails to operate within established controls. Adversarial manipulation, unstable behavior, model drift and algorithmic discrimination also count.

Cowbell's product page lists ten first-party coverages that respond to that event.

First-party coverage	What Cowbell says it pays for
AI Incident Response	Investigating, containing and isolating the AI tool, and preserving evidence
Data Exposure Response	Legal guidance, notifications, call center services and identity monitoring
Output Correction	Correcting or withdrawing inaccurate or unsafe output and notifying recipients
Proof of Loss Preparation	Forensic accounting, with the insurer's consent, for a business interruption claim
Reputational Harm	Crisis management and public relations
System Restoration	Restoring data, systems, workflows and settings damaged by a covered AI tool
Business Interruption	Lost profit and extra costs after a waiting period
Direct Transaction Loss	Unrecovered money lost through manipulation, unapproved actions or control failures
Runaway Cost	Unauthorized AI usage charges from looping, escalating tasks or failed spend limits
LLMjacking	AI usage charges from outsiders using stolen credentials or API keys

Source: Cowbell, AI Liability Insurance product page, November 2026.

The third-party side has five heads: liability, regulatory proceedings, algorithmic discrimination, copyright or trademark infringement, and contingent bodily injury and property damage "resulting from reliance on AI output or an autonomous digital action." That last head answers the cover CG 40 47 takes out of a CGL policy. Punitive damages, fines, penalties and consumer redress are included where insurable.

"When an agent goes rogue, a business can face its own losses and claims from customers and others at the same time," said Simon Hughes, Cowbell's chief commercial officer.

Cowbell has published no limits, retentions, sublimits, exclusions, price or carrier for the policy. When it launched Prime One in April, a non-admitted cyber form with affirmative AI cover for companies with $250 million to $1 billion of revenue, it named Chaucer Insurance Company as the paper and limits of up to $10 million.

The new policy will be sold only through wholesale brokers on a non-admitted basis. Cowbell's site says it may act as a general agent for one or more carriers, and that a non-admitted policy escapes many state rate and form rules, the filing gap actuary.info covered in August in its piece on Cowbell's OMNI underwriting system.

Three factors and the exposure they measure

Cowbell said it assesses the exposure with three AI-specific Cowbell Factors, used alongside other underwriting information. They score how a business monitors its AI tools, what authority those tools have to act, and the controls that govern them. An October 5 release gave them names: AI Exposure for observability, AI Vulnerability for autonomy and AI Assurance for governance.

The inputs come from the insured's own AI accounts. Businesses get their factors by activating Cowbell Connectors for tools such as Gemini, Claude and ChatGPT, the company said. It counted more than 60 connectors in October and said its risk data covers more than 55 million businesses.

All three factors grade how an agent is watched and constrained. Autonomy comes closest to a measure of exposure, because the loss in each of Cowbell's three claim scenarios scales with what the agent was allowed to touch: the payments an accounts payable agent can release, the spending controls on a quoting agent, the number of client tax returns a bookkeeping agent files.

The trigger leans on the same controls. Three of its five core events are defined against the insured's own rules: authorized scope, required approval and established controls. The governance an underwriter scores at binding is therefore also the yardstick for whether an incident happened at claim time.

One broker described the same direction this month. Ashley Bennett Jones, who leads AI risk and advisory in Lockton's cyber and technology business, told Insurance Business on November 2 that some carriers see agent-to-agent activity as something to underwrite and charge for differently. "There is less opportunity for the human in the loop to one, exist in a meaningful way, or two, mitigate risk," she said.

Jones also said flexible technology E&O wordings have reduced the incentive to buy standalone AI cover, and that interest had yet to turn into accelerating take-up. A slow start would leave Cowbell rating on its factors for some time before it has claims of its own, the problem set out in actuary.info's July piece on pricing AI liability with almost no loss history.

One incident, many claimants

Cowbell's scenarios each produce several payees from a single failure. In one, a quoting agent enters a loop, launches additional agents and sends customers incorrect quotes while usage charges build, until the business suspends the system. That is runaway cost, business interruption and output correction on the first-party side, and a possible claim from every customer who received a wrong quote.

In another, a bookkeeping agent applies outdated tax rates and files client returns without review. Clients incur penalties and interest and sue the firm. One error is repeated across every return filed, and how many retentions and limits it consumes will turn on how the issued form counts an AI Incident. Model drift, which the trigger names, has no single date of loss, a point covered in an earlier piece on drift as an insured peril.

The same arithmetic applies across the book. The three tools Cowbell named for its connectors come from three model providers, and a change at one of them that alters agent behavior would reach every insured running on it. Runaway cost and LLMjacking losses are also bills from those same providers. Actuary.info covered that provider concentration in July.

Insurers are deploying the agents Cowbell describes. Small commercial carrier MGT made its Aimee assistant available through ChatGPT so appointed independent agents can get business owner's policy quotes in conversation, Insurance Innovation Reporter said on November 6. The agent still reviews and binds in MGT's portal. MGT says its platform has handled more than $83 billion of quoted total insured value. Digital Insurance ran a piece the same day under the headline "AI agents that bind policies could leave insurers liable."

Allstate plans a consumer-facing one. Chief executive Tom Wilson said customers could expect a personal agentic assistant early in 2027, Carrier Management reported on November 10. Asked about regulatory and cyber concerns, Wilson said Allstate will "figure out stuff like what do we do with the regulators, what do we do with privacy, what do we do with cybersecurity, as we go."

Cowbell's policy becomes available on January 1. Insurance Business said final terms will govern at issuance.

Cowbell Launches a Standalone Liability Policy for AI Agents

Key Takeaways

What the policy pays for

Three factors and the exposure they measure

One incident, many claimants

Further Reading

Sources

Cowbell Launches a Standalone Liability Policy for AI Agents

Key Takeaways

What the policy pays for

Three factors and the exposure they measure

One incident, many claimants

Further Reading

Related guides

Sources