Superintelligence Insurance: What Your Cyber Policy Says About Agents
Before an AI agent gets write access, ask how your cyber policy defines agent actions, data handling and errors, and where exclusions and limits apply.

AI agent insurance coverage depends on the wording of your cyber policy, not on the name of the tool your team uses. Before an agent can send messages, change records or trigger payments, ask your broker how the policy defines an action taken by software, who counts as the actor, and which errors and data events fall inside or outside the cover. The written answers decide whether an agent's mistake becomes a claim you can make or a cost you carry alone.
Key takeaways
- A cyber policy written around employees, servers and vendors may not say how an autonomous agent's actions are treated, so get the position in writing.
- Exclusions and sublimits often matter more than headline limits when one agent fault repeats across many records or customers.
- Data questions reach beyond breaches. They cover agent memory, tool logs, prompt histories and the subprocessors behind your vendors.
- Insurers tend to ask for controls you can show: scoped access, approval gates, logs and a tested way to pause the agent.
- Do not grant write access until the policy answers are documented and every write action can be reviewed and reversed.
What does your policy count as an action?
Most cyber and technology errors policies were drafted with people and systems in mind. An agent sits in an awkward middle: it is software, but it acts through credentials, reads customer records and sometimes writes to them. Ask whether an action taken through a valid account by a non-human process is treated as authorized access, as an employee error or as a system failure. Each label leads to a different claims path, and the difference can change who pays for investigation, correction and notification.
Ask your broker these questions in writing:
- Does the policy define the insured, the user or the employee in a way that includes or excludes automated agents?
- If an agent takes a wrong action with a valid credential, is that treated as unauthorized access, a service error or neither?
- Does the cover reach damage to a third-party system that the agent changes through an integration you configured?
- Does the cover respond when the fault sits upstream, in a model provider or tool vendor, rather than in your own configuration?
Read the exclusions before the limits
A large limit says little if the exclusions remove the event you are most likely to face. Look for exclusions tied to failure to maintain security, unpatched systems, known vulnerabilities, contractual liability, professional services and intentional acts. Agents raise a specific question about intent. If an agent takes an action its operator did not intend, the insurer may ask whether the company knew the risk and failed to control it. Your written controls become part of that answer, which is why they belong in the same conversation as the policy.
The practical questions are these. Which exclusions refer to controls, and does the wording define what counts as a control? Does a contractual liability exclusion affect commitments an agent makes to customers, such as confirming a delivery date or a refund? Are sublimits set for business interruption, regulatory response or data restoration that one agent fault could trigger? What is the retention, meaning the amount you pay before cover applies, and does it apply per event or per claim?
Data handling: what the policy says about memory and logs
Insurers usually ask what personal or confidential data a system touches, where it is stored and who else processes it. An agent that keeps memory, reads email or writes call notes creates records in places a traditional application never did. Ask whether the privacy and breach response terms cover data held in agent memory, prompt histories, tool logs and vendor environments. Notification duties depend on the jurisdictions where your customers and staff are located, so align this section with your privacy counsel. Brazilian companies should also review how Superintelligence and LGPD: Keeping Agent Memory Lawful in Brazil applies to what the agent retains.
Do not assume that an agent's logs are harmless because they are internal. Logs can hold personal data, and an insurer or regulator may ask to see them after an incident. The policy should be read alongside your retention rules, so that the records you need for a claim exist and the records you no longer need are removed on schedule.
Data that the agent remembers is still data your company is responsible for.
Errors and omissions: an illustrative scenario
Illustrative scenario: a company lets an agent answer billing questions and update contact details. A customer writes that a new address is effective from a certain month, and the agent misreads the message and overwrites the wrong record. The next invoice goes to a former office. The customer complains, and staff spend several hours correcting the record, reviewing other updates made the same week and writing an explanation. No money is lost, but the cost is real.
In this scenario, the policy questions are concrete. Does the wording treat a wrong record update as a data event, a service error or nothing at all? Is the staff time spent on review and correction covered, or only third-party damages? Does the cover depend on a human review process that the company must maintain? Changes to money, legal terms or customer identity usually need a second check before the agent executes them. Two-Person Approval for AI Agents: Which Actions Need It explains how to choose those actions, and the answer should match what your policy assumes you will do.
Controls an insurer will ask you to show
Underwriters and brokers ask for evidence, not intentions. A promise that the agent is "monitored" does not answer a question about who reviews its actions, how often and with what record. Prepare a short, dated file for each agent that shows the controls in place. The Cyber Insurance and AI Agents: The Controls Underwriters Ask About guide is a useful starting point for the conversation.
- A written inventory of every agent, the tools it can call, the credentials it uses and the systems it can write to.
- Scoped permissions that separate read access from write access, with no shared administrator credentials.
- Approval gates for actions that change money, identity, legal terms or customer commitments.
- Logs that record which person or process the agent acted for, which tool it used and what changed.
- A tested way to pause the agent and its queued work without deleting the records needed for review.
- Vendor contracts that cover data processing, subprocessors and incident notice.
- A named owner who reviews agent errors and decides on correction and follow-up.
The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a common vocabulary for mapping and managing these risks, which can help your broker and your operations team speak the same language. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) gives practical guidance on checking permissions at each action rather than once at setup.
A decision guide before granting write access
Use the following order. Each question is a gate. If the answer is no, the agent stays read-only until the gap is closed.
- Can your broker confirm in writing how the policy treats actions taken by an agent? If not, keep the agent read-only.
- Have the exclusions for failure to maintain controls been compared with the controls you can actually demonstrate? If not, close the gap first.
- Does every write action appear on a list with an owner and a reversal path? If not, do not grant that action.
- Can a named person pause the agent within a short, practiced window? If not, run a drill before launch.
- Is the first write task narrow enough that any error can be found and corrected from the logs? If not, reduce its scope.
- Have the vendor's data processing terms and subprocessor list been reviewed against the data the agent will touch? If not, pause the rollout.
This order also gives you a way to say no to a vendor or a colleague without arguing about the technology. The question is whether the company can explain, correct and stop the agent's actions, and whether its insurance wording matches that reality.
How Be Human approaches the policy questions
Be Human's managed approach starts by defining one useful role for the agent, the memory and tools it needs, its approval boundaries and an observable success criterion. Those definitions give the policy questions concrete answers, because the broker can see exactly what the agent can read, write and change. Ongoing operations include monitoring, recurring evaluations, permission review, integration maintenance and incident response, which produce the records an insurer or auditor may ask for. Be Human does not set your coverage or give insurance advice, so your broker and counsel remain responsible for the policy. If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Does a standard cyber policy cover mistakes made by an AI agent? Not necessarily. Coverage depends on how the policy defines the actor, the error and the data event, and on which exclusions apply. Ask your broker for the written position on agent actions before you rely on it.
Should I buy a separate policy for AI agents? It may be worth considering, but first map what your current policy already says. Some companies find that existing cyber or technology errors wording needs clearer language or an endorsement, while others need a separate product. Your broker can compare the options against the actual scope of your agents.
What should I ask my insurer before giving an agent write access? Ask how agent actions are classified, whether errors in records are covered, which exclusions refer to controls and what evidence the insurer expects you to keep. Also ask how notification works if personal data is involved, and whether sublimits apply to a single agent fault.
Does insurance replace controls on the agent? No. Insurance can transfer part of the financial risk, but it does not stop an agent from changing a record or sending a message. Scoped access, approval gates, logs and a tested stop procedure remain your responsibility, and they are often what makes the cover workable in the first place.
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.


