Cyber Insurance and AI Agents: The Controls Underwriters Ask About
Underwriters asking about AI agents look for access reviews, action logs, spending limits and human approval for payments. Here is how to prepare the evidence.

Cyber insurance underwriters who ask about AI agents want evidence that an agent cannot act beyond what the business has approved, that its actions can be reconstructed afterward, and that a person signs off on money movement and sensitive changes. The questions usually cluster around access reviews, logging, human approval for payments, and how quickly access can be revoked. A business that can show these controls in writing and in practice is in a much better position to discuss terms than one that describes its agent as a general assistant.
Key takeaways
- Underwriters care less about how capable an agent is and more about whether a loss would be bounded, detectable and explainable.
- Every agent needs its own identity, a named owner, a written job description and a current list of the tools it can reach.
- Access reviews, action logs and approval records should exist before the renewal questionnaire arrives, not be assembled under deadline.
- Payment and money-movement actions need a human approval step that is real, recorded and out of the agent's reach.
- Revocation is a tested procedure with named people, not a setting you hope to find when something goes wrong.
What underwriters are trying to learn
Underwriters are rarely asking whether your AI agent is clever. They are asking whether an incident involving the agent would be limited in scope, visible to your team and explainable to a third party. A cyber policy depends on controls that already exist for people and systems: who has access, what they did, who approved sensitive actions, and how quickly a problem was contained. An agent sits inside those same questions, with one difference. It can take many small actions in a short time, through tools that someone configured earlier and that few people in the company still remember.
For a business owner, the most useful frame is to treat each agent as a non-human account with a defined job, a defined set of permissions and a named owner. If you cannot name the owner, list the tools the agent can call, and show when someone last reviewed those permissions, the underwriting conversation will be harder than it needs to be. For the questions to take to your broker about how a policy treats automated actions, see AI Liability Insurance for Superintelligent Agents: Insurer Questions.
Access reviews: what can the agent actually do?
Start with identity. Most agents act through credentials: an API token, an OAuth grant, a service account, or a connection to a messaging channel. Underwriters commonly ask whether the agent has its own identity or borrows an employee's login. Borrowed logins cause two problems. The logs then show a person's name for work that person did not do, and offboarding that person can silently break the agent or leave it with access nobody intended.
An access review answers one question: do the permissions still match the job? A useful review covers the following ground:
- The complete list of tools and data sources the agent can reach, with the business reason for each one.
- Read access separated from write access, and a note on whether each write action can be reversed.
- OAuth scopes and API permissions compared against the narrowest scope that still does the work.
- Shared memory and knowledge stores, including who can add to them and who can delete from them.
- The person who approved the current permission set, and the date of the last review.
The OWASP Authorization Cheat Sheet is a practical reference for the principle behind these checks: access decisions should follow an explicit policy, be enforced on every request, and not depend on what the user interface happens to show. An agent that can call a tool, regardless of whether a menu displays it, needs the same discipline as any other account.
Logging: can you reconstruct what happened?
Logging turns an incident from a disputed story into a set of facts. Underwriters and claims handlers want to know what the agent was asked to do, what context it used, which tools it called, what each call returned, and which person or rule allowed the final action. A log that says only that the agent sent an email will not help when a customer disputes what the message said or why it was sent.
Useful agent logs share a few traits. Each action carries a timestamp, the identity that performed it, the tool name, the parameters, the outcome, and a link to the request that triggered it. Personal data in parameters is redacted where your policy requires it, but the record still shows what happened. The logs are stored where the agent cannot edit or delete them, and retention follows a written rule rather than whatever default a vendor ships with.
The OpenTelemetry observability primer describes the general discipline of collecting traces, metrics and logs so that a system's behavior can be understood from its outputs. The same thinking applies to agents: instrument the steps the agent takes, not only its final answer. If the trail stops at the answer, you cannot show an underwriter how the answer was reached.
A final message in a log tells you what was sent. The steps before it tell you whether it should have been.
Illustrative scenario: a small distribution company lets an agent answer routine questions from customers about order status. One afternoon a customer complains that the agent gave a delivery date that turned out to be wrong. Because each step was logged, the team can see that the agent read an outdated shipping table, cite the table's last update, and show that the table was not supposed to be a source for dates. The team fixes the source, records the change, and can explain the incident in the same afternoon.
Human approval for payments and money movement
Payments are where underwriters tend to be most direct. If an agent can start a bank transfer, change a supplier's bank details, issue a refund, or spend from a company card, the question is who stands between the agent's decision and the money. The usual answer is a human approval step, and the follow-up question is whether that step is real or only ceremonial.
A real approval step has four features. The approver sees the specific action, the amount, the recipient and the reason, not a generic message asking for permission. The approver is a named person, and for higher amounts or higher risk the approver is not the same person who requested the action. The decision and its time are recorded. The agent cannot approve its own request through another tool, a saved shortcut or an instruction it has been given to skip the step.
Retries also matter. Agents and the systems around them sometimes send the same instruction twice, and a payment processed twice is a loss that approval alone does not prevent. Stripe's documentation on idempotent requests shows how a provider can recognize a repeated request and avoid processing it again. Ask whether every payment call carries an idempotency key and whether the log shows both the original attempt and the retry.
Some businesses set different approval levels by amount or action type. For a guide to deciding which actions need a second person, see Two-Person Approval for AI Agents: Which Actions Need It. For where approval steps should sit inside a workflow, see Human Approval Gates for Superintelligent Agents: Where People Sign.
Revocation: how fast can you stop the agent?
Underwriters ask how quickly you can cut an agent off, and the honest answer names specific steps. Which tokens are revoked? Which channels are disconnected? Who has the authority to pause the agent, and who is the backup? What happens to work the agent has queued but not yet finished? Revocation is a runbook that touches several systems, so it cannot be a single setting in one place.
When a person who configured or supervised the agent leaves the company, the same questions apply with added urgency. The steps for removing that person's access to agent tools are covered in Offboarding Employees From AI Agent Tools: A Same-Day Checklist. Treat the agent's own credentials the same way: they belong to the business, they should be listed, and they should be removable without waiting for the original builder.
The pause procedure should be tested at least once and the result written down. A test that reveals a missing credential or an unknown queue is valuable, because it is much cheaper to find during a calm afternoon than during an incident.
A step-by-step plan to prepare the evidence
If a renewal or a new policy is coming, this sequence produces most of what an underwriter will ask to see:
1. Write a short description of each agent's job and name the person who owns it. 2. List every tool, data source and permission the agent holds, with the business reason for each. 3. Mark every action that moves money, changes customer records, or sends messages outside the company. 4. Confirm that each marked action has a human approval step, and that the approval is recorded with the approver's name and time. 5. Export a sample of recent logs and check that each entry can be traced to a request, a tool call and an outcome. 6. Run an access review, record what was removed or narrowed, and keep the sign-off. 7. Test the pause and revocation procedure, then store the record with the other evidence.
The point of this plan is not to produce a document for one meeting. Underwriters, auditors and your own team will ask the same questions again, and the answers should be easy to refresh.
Where a managed approach fits, and a first step
Be Human designs, implements and operates personalized AI agents whose identity, memory, channels, tools, permissions and operating context stay under human control. A first engagement defines one useful role, the memory and tools that role needs, the approval boundaries, and a success criterion that can be observed. Managed operations can add monitoring, recurring evaluations, permission review, integration maintenance and incident response, so the access reviews and log checks above happen on a schedule rather than once a year. Be Human does not promise insurance terms or any particular underwriting outcome. What it can help a business produce are the records and controls an underwriter will ask to see.
If you are unsure which task should be the first one to hand to an agent, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Does cyber insurance cover losses caused by AI agents? Coverage depends on the wording of your specific policy and on the insurer, so no general answer applies to every business. Ask your broker directly how the policy treats losses where an automated system took the action, and whether exclusions or conditions refer to human approval, logging or access controls. Write down the answers and keep them with your agent documentation.
What logs do underwriters usually want to see from an agent? They usually want to know what the agent did, when, with which identity, through which tool, and with what result. They also want to know which person or rule permitted sensitive actions. A sample of logs that can be traced from request to outcome is more persuasive than a long description of logging in the abstract.
Do I need a human to approve every action an agent takes? Not usually, and requiring approval for everything tends to make the approval meaningless. Focus human approval on actions that move money, change sensitive records, or communicate outside the company, and let low-risk reads and drafts run within their permissions. Review the boundary periodically as the agent's track record and the business's needs change.
What should we prepare before a renewal questionnaire arrives? Prepare the evidence before the questionnaire does: an owner for each agent, a current list of tools and permissions, a record of the last access review, a sample of traceable logs, a list of approval-gated actions, and a tested pause procedure. Keep these documents in one place and refresh them on a schedule. Answering from existing records is faster, more accurate and easier to defend than reconstructing them under pressure.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


