The AI Agent Inflection Point
According to Gartner, 57% of companies now have AI agents running in production environments — up from nearly zero three years ago. That number is expected to double by the end of 2026. We are past the pilot phase.
But most companies’ insurance programs were written before that inflection point. The policies renewing today were structured for a world where AI meant a chatbot answering FAQs or a model flagging transactions for human review. That world is gone.
AI agents are a fundamentally different class of technology. A traditional AI tool — a copilot, a recommendation engine, a classification model — assists a human who then makes a decision. An AI agent executes. It takes actions in external systems, calls APIs, sends communications, initiates transactions, and sequences decisions across multi-step workflows — without a human approving each step.
That distinction matters enormously for insurance. Liability frameworks are built around who made a decision and whether they acted reasonably. When an AI agent makes thousands of micro-decisions per hour with no human in the loop, traditional liability attribution breaks down. And most current policies were not written to address it.
Shadow AI: The Exposure You Can’t See
The problem isn’t only the AI your company has officially deployed. It’s the AI your employees adopted last Tuesday without telling IT.
Enterprise shadow IT research consistently finds that large organizations run more than 1,200 unsanctioned applications at any given time. AI tools now represent the fastest-growing segment of that shadow ecosystem. Employees are using ChatGPT to draft contracts, Copilot to summarize earnings calls, department-specific AI tools to analyze customer data — often without IT approval, security review, or any record of use.
Each unsanctioned tool is a potential uninsured liability. When an employee uses an unauthorized AI tool to generate a compliance document that turns out to be wrong, or to draft a client-facing communication that triggers a dispute, the company bears the exposure. But because the tool was never inventoried, never reviewed, and never disclosed, there’s no coverage position to stand on.
Shadow AI creates the worst possible insurance scenario: liability you’re exposed to but cannot document, quantify, or defend. Carriers exclude what they cannot price. And they cannot price what they cannot see.
How AI Agent Failures Create Insurance Claims
Not all AI risk is equal. The AI risk classification framework categorizes AI deployments into four tiers based on decision autonomy and external impact:
Tier 1 — Internal information processing. AI summarizing internal documents, organizing data, generating internal drafts. Human reviews all outputs before any external action. Low risk; most standard policies respond adequately.
Tier 2 — External-facing interaction. AI communicating directly with customers or third parties. The Air Canada chatbot case is the canonical example: an AI agent told a customer it could get a bereavement discount that didn’t exist, and the airline was held liable. The agent acted; the company owned the consequence.
Tier 3 — Transaction execution. AI initiating financial transactions, operational decisions, or contract actions with real-world consequences. A pricing agent that misapplies a discount at scale. A procurement agent that executes purchase orders. The error compounds across every transaction before it’s caught.
Tier 4 — Autonomous business operations. AI agents coordinating other agents, operating with broad authority across systems, making strategic-level decisions with minimal human oversight. The highest risk tier and the one most insurers currently lack a coherent coverage framework for.
Agent failures are qualitatively different from tool failures. A traditional software bug produces a discrete error. An agent failure cascades: the agent makes a flawed decision, acts on it, and that action triggers the next decision in the sequence — all before any human checkpoint. By the time the failure surfaces, the damage is compounded and the audit trail is complex. That is not a scenario traditional E&O or cyber policies were priced to handle.
The Exclusion Language Problem
Insurance carriers have responded to AI risk primarily through exclusion endorsements. The ISO exclusion forms — CG 40 47 and CG 40 48 — have been widely adopted, and their definitions set the terms for coverage disputes.
“Generative artificial intelligence” means a machine-based learning system or model that is trained on data with the ability to create content or responses, including but not limited to text, images, audio, video or code.
That definition is broad. It captures not just large language models but any system trained on data that produces outputs — which includes most AI agents. If your agent generates a decision, a recommendation, or a communication as part of its workflow, it likely falls within this definition.
The Verisk CG 40 47 analysis explores how this exclusion language interacts with commercial general liability coverage in practice. The short version: the exclusion is written to be inclusive, and courts will likely read it that way.
Berkley PC 51380 takes a different approach — offering a coverage endorsement rather than a blanket exclusion, structured around documented AI governance. The difference matters: Berkley’s form creates a pathway to coverage for companies that can demonstrate oversight and control. The ISO exclusion forms create a pathway to a denied claim for companies that cannot.
The key question every company should be asking before renewal: if your AI agent makes a decision that causes a loss, does your current policy respond? Most companies don’t know the answer. Many would not like it if they did. Review your AI insurance exclusions before your carrier does it for you.
Documentation That Changes the Calculus
Carriers don’t exclude AI because it’s inherently uninsurable. They exclude it because they can’t price what they can’t see. Documented, governed AI deployments are insurable. Undocumented, ungoverned ones aren’t.
This is the core insight that most companies miss. The carrier community is not trying to exit AI risk permanently. They are trying to exit unquantifiable AI risk. Those are different problems with different solutions.
What carriers need to see — and what structured underwriting questionnaires increasingly ask for — is a complete AI inventory, risk tier classifications for each deployment, a governance framework with defined oversight responsibilities, and evidence of ongoing monitoring. Not a policy document that says “we have an AI governance committee.” Evidence: logs, review records, escalation workflows, incident response procedures.
The gap between “we use AI” and “here is exactly how we use AI, how it’s governed, what controls are in place, and how we detect and respond to failures” is the gap between an uninsurable risk and a priced one. Brokers working with clients on AI risk management programs are finding that the companies that can close that gap are getting materially better terms — sometimes coverage that would otherwise be excluded entirely.
This documentation is the foundation of a carrier-ready risk assessment. It is also, not coincidentally, the foundation of a defensible position if a claim does occur.
What Companies Should Do Now
The window between “carriers are paying attention to AI” and “carriers have locked down their exclusion language for another cycle” is closing. Companies that move now can still influence their coverage terms. Companies that wait will be negotiating from a weaker position at renewal.
Take inventory. You cannot govern or insure what you don’t know exists. Find every AI tool and agent in your environment — including shadow AI. This means going beyond IT’s approved list to include department-level tools, individual subscriptions, and any third-party services that process your data with AI. Get an AI Risk Assessment to establish a baseline.
Audit your policies. Pull your current CGL, E&O, cyber, and D&O policies and look for AI endorsements — specifically CG 40 47, CG 40 48, and any carrier-specific AI exclusion filings. If you don’t know what’s in your policies, your broker can help. The Exclusion Hub provides plain-language explanations of the most common forms currently in circulation.
Document governance. Build the evidence package that demonstrates AI oversight: a complete inventory with risk tier classifications, defined ownership for each deployment, monitoring procedures, and incident response protocols. This is not a one-time project — it needs to be maintained and updated as your AI footprint changes.
Time it to renewal. Your next policy renewal is the practical deadline. Carriers review AI governance at underwriting, and they are tightening terms every cycle. Coming to renewal with documented governance is leverage. Coming to renewal without it is not.
Your next renewal is the deadline. If you don’t know what AI is running in your environment — or whether your policies cover it — start with an AI risk assessment.