Short answer
Most AI companies begin by reviewing technology errors and omissions and cyber liability, then add general liability, D&O, crime, workers’ compensation, or other lines as their operations and contracts require. The harder question is not which model is in production; it is what the system may decide, spend, promise, publish, change, deploy, or transact—and whether the policy wording and evidence describe that reality.
Evidence frame
- Established
- Insurance responds through the wording, definitions, exclusions, conditions, endorsements, and limits of the policies a company actually buys. A generic AI label does not decide the result.
- Clara inference
- The most useful first underwriting artifact is an authority map: what the system can decide, touch, change, promise, spend, publish, or transact, and who remains accountable.
- Hypothesis
- A dated record that joins system authority to contracts, controls, incidents, and policy wording will make AI-company risk easier to review than a model inventory alone.
- Unknown
- No public guide can determine whether a particular loss is covered. The exact policy form, endorsements, application, facts, contracts, law, and carrier interpretation remain decisive.
The program starts with the work, not the model
A model is a component. The insured operation is the company that packages it into a product, connects it to data and tools, makes representations about what it can do, and asks customers to rely on the result. Two companies using the same model can present entirely different insurance questions.
| Line | Typical loss question | What to read carefully |
|---|---|---|
| Technology E&O | Did the product or professional service cause a customer financial loss? | Covered services, technology-product definitions, contractual liability, AI exclusions, and limits. |
| Cyber liability | Did a security or privacy event affect systems, data, or response obligations? | Network-security triggers, data definitions, system failure, third-party claims, and incident response. |
| General liability | Did the operation cause bodily injury, property damage, or covered personal and advertising injury? | AI exclusions, products/completed operations, physical operations, and contractual requirements. |
| D&O | Are directors or officers accused of mismanagement, disclosure, or governance failures? | Entity and individual coverage, claims-made mechanics, exclusions, and investor or regulator allegations. |
| Crime and other lines | Could deception, funds movement, employees, premises, vehicles, or equipment create a separate loss path? | Definitions, authorized-versus-fraudulent activity, social engineering conditions, and operational scope. |
This is a map for investigation, not a universal package. The U.S. Small Business Administration describes insurance needs as dependent on the business and its risks; an AI company adds a moving technical and contractual layer to that ordinary starting point.
Authority is the underwriting variable
The phrase “AI risk” is too broad to guide a serious review. A useful description starts with authority: the actions a deployed system is permitted and equipped to take, the resources it can reach, the people who supervise it, and the point at which a person can stop or reverse it.
- Low authority drafts, classifies, summarizes, or recommends while a person approves the consequential action.
- Operational authority changes records, routes work, communicates with customers, or commits the company to a process without reviewing every step.
- Economic authority can spend, price, negotiate, approve payments, change production code, or delegate to another system.
- Physical authority can command a robot, vehicle, drone, industrial process, or other machine in the world.
These categories are not policy definitions. They are a way to make the company’s operating reality legible before a policy form is interpreted. NIST’s AI Risk Management Framework similarly treats risk management as a lifecycle activity rather than a one-time model label.
The same deployment can create several claim paths
A customer-service agent that sends an incorrect refund may create a financial-loss dispute. The same agent, after a prompt-injection event, may expose customer data. A system connected to a warehouse or vehicle can add bodily injury and property damage. The cause may cross policy boundaries even when the company experiences one incident.
| Deployment event | Potential loss surface | Evidence that matters |
|---|---|---|
| Faulty output relied on by a customer | Professional liability, contract, or consumer-protection claim | Version, customer promise, approval path, logs, and remediation. |
| Agent manipulated into revealing data or credentials | Cyber, privacy, technology, and regulatory response costs | Identity, tool permissions, access logs, detection, containment, and notification. |
| Agent changes code, configuration, or production data | Service interruption, customer loss, or a security event | Change authority, review gates, rollback, testing, and affected systems. |
| Agent commits money or makes a binding promise | Crime, contract, E&O, or management claim | Transaction limits, dual control, authorization, contract allocation, and payment records. |
| Autonomous system affects people or property | General, product, auto, aviation, workers’ compensation, or equipment loss | Operating envelope, supervision, maintenance, site rules, and event telemetry. |
The point is not to force every event into every line. It is to stop treating a single “AI policy” as the answer to several different mechanisms of loss.
The form and the evidence matter as much as the limit
A limit is not a promise that every AI-related loss is covered. The definition of professional services, the covered operations, the insuring agreement, exclusions, conditions, retentions, and contract language decide how a claim behaves. Clara’s analysis of the 2026 AI exclusions shows why a policy label is a poor substitute for reading the actual wording.
The evidence file should be equally concrete. Keep a versioned record of the agents in production, their sponsors and purposes, tools and credentials, authority limits, human approvals, monitoring, shutdown and rollback, incidents, near misses, and material changes. Keep the customer promises and provider terms beside the technical record. A reviewer needs to understand what the company said the system would do and what it could actually do.
This is the practical bridge to RISK.md: a company-owned context packet can preserve facts, evidence, uncertainty, and change without pretending to make underwriting judgment on behalf of a market.
What Clara is testing
Clara’s question is not whether every AI company needs a new policy. It is whether the insurance process can keep pace with companies whose operating authority changes faster than an annual description of the business. That requires translating technical facts into underwriting evidence, then testing which facts change appetite, terms, controls, or claims analysis.
The current public work moves through that loop: the AI exclusions note reads committed market wording; Efficient Path studies what capable agents do when a goal and a weak boundary meet; and the manifesto sets the larger question. The answer is still being earned through evidence.
Common questions
What insurance does an AI company usually review first?
Technology E&O and cyber liability are common starting points because AI products can create customer financial loss and can process or expose data. General liability, D&O, crime, workers’ compensation, property, auto, and other lines depend on the company’s operations, people, assets, contracts, and jurisdictions.
Is there one AI insurance policy that covers everything?
No universal policy should be assumed. AI-related loss can implicate several lines, and the answer depends on the policy wording, exclusions, conditions, contracts, facts, and applicable law.
What matters more than the model name?
The deployed authority: what the system can decide, spend, promise, publish, change, deploy, or transact; which tools and data it can reach; and what controls and evidence surround those actions.
Do AI exclusions automatically mean an AI company is uninsured?
No. An exclusion may remove a defined exposure from one policy or coverage part while other wording, policies, endorsements, or affirmative coverage may respond differently. The actual form and claim facts must be read together.
What should an AI company prepare before an insurance review?
Prepare a current system inventory, authority and permission map, controls and testing evidence, incidents and near misses, customer contracts and promises, model-provider terms, revenue and operations information, and a clear list of unknowns.