Key takeaways

  • The same six items come up repeatedly: an AI inventory, agent permissions, logs, human approval points, testing records and a named owner. If you can produce them, most questionnaires become easier to answer.
  • FINRA's 2026 report lists considerations for AI agents, including how to monitor agent system access and data handling, where to place human-in-the-loop oversight, how to track agent actions and decisions, and how to set guardrails. The report is not rulemaking.
  • The CSA AI Controls Matrix has 243 control objectives across 18 domains and maps to ISO/IEC 42001, ISO 27001 and the NIST AI RMF. It is a framework to map against, not a certificate.
  • ISO/IEC 42001 is a management system standard. Certification is optional and is issued by independent certification bodies, not by consultants.
  • Optional generative AI exclusion endorsements for general liability policies have been available from Verisk's ISO since January 2026. Read your renewal documents and ask your broker what applies to you.
  • This is not legal or insurance advice. Nothing here means a business is compliant with, or covered under, anything.

Why the questions are changing

Two years ago, a cyber insurance application asked about multi-factor authentication, backups and endpoint protection. Those questions remain. Now a second set sits beside them: Which AI tools do your people use? Do you run AI agents that act on their own? What can they access? Who approves what they do? In our experience, these questions arrive from several directions at once: insurance brokers at renewal, enterprise customers in vendor questionnaires, regulators in examinations, and boards asking for a briefing.

We are not insurers, auditors or lawyers, and we do not tell any business what its policy covers or whether it complies with a rule. What we can do is read the published frameworks and regulator materials and describe what they say. This article does that, then turns the pattern into a list of records worth having.

Six records that come up again and again

  • An AI inventory. A list of the AI tools, models and agents in use, including the ones staff adopted without asking: what each does, who uses it, what data it touches, and which vendor provides it.
  • Agent permissions. For each agent, the systems it can reach, the actions it can take (read, write, send, delete, pay) and the identity it uses to do so.
  • Logs. Evidence of what the agent did: actions, decisions, tool calls, and the data it accessed, kept for a stated period.
  • Human approval points. The steps where a person must review or approve before an action takes effect, and a record that the approval happened.
  • Testing records. What was tested, how, when, what was found, and what changed afterward.
  • A named owner. A person accountable for each AI system and for the program as a whole, with a defined escalation path.

Each of these appears, in different words, in the sources below.

What FINRA's 2026 report lists for AI agents

The Financial Industry Regulatory Authority (FINRA) publishes an annual regulatory oversight report that tells member firms what its staff are focusing on. The 2026 edition includes a section on generative AI that covers AI agents. It describes risks including agents acting autonomously without human validation and approval, agents acting beyond a user's actual or intended scope and authority, multi-step reasoning that makes outcomes difficult to trace or explain and so complicates auditability, and agents that may store, explore, disclose or misuse sensitive or proprietary information.

For firms to consider, FINRA lists how to monitor agent system access and data handling; where to have "human in the loop" agent oversight protocols or practices; how to track agent actions and decisions; and how to establish guardrails or control mechanisms to limit or restrict agent behaviors, actions or decisions.

Read those four items against the six records above and the overlap is plain: access and data handling map to permissions and the inventory, human in the loop maps to approval points, tracking actions and decisions maps to logs, and guardrails map to permissions and testing. FINRA's report is oversight guidance for its member firms. It is not rulemaking, and it does not create new requirements by itself. We cite it because it is a clear, primary-source statement of what an examiner may want to discuss, and because other industries borrow its framing.

CSA's AI Controls Matrix

The Cloud Security Alliance (CSA) publishes the AI Controls Matrix (AICM), a vendor-agnostic framework for cloud-based AI systems. CSA's page describes version 1 as released in July 2025 and updated in October 2025, with 243 control objectives across 18 security domains, and notes that a newer version, v1.1, supersedes it. It builds on CSA's Cloud Controls Matrix. CSA lists mappings to ISO/IEC 42001:2023, ISO 27001 and NIST AI RMF 1.0, and also to NIST AI 600-1, BSI AIC4 and the EU AI Act.

The AICM comes with a questionnaire, the AI-CAIQ, that organizations can use for self-assessment or to evaluate vendors, and with implementation and auditing guidelines for different roles: model providers, orchestrated service providers, application providers, AI customers and cloud service providers. That role split is useful. A business that uses a hosted AI service is an AI customer and has different responsibilities than one that builds and hosts its own model.

For a company preparing evidence, the AICM is a practical index. You can take your inventory, permission records, logs, approval records and test results and map each to the relevant control objectives. When a customer sends a questionnaire in different words, the mapping lets you answer consistently. Mapping evidence to the AICM is not a certification and does not show that a business meets any particular requirement. It organizes the record.

ISO/IEC 42001

ISO/IEC 42001:2023, "Information technology — Artificial intelligence — Management system," was published on December 18, 2023. ISO describes it as defining requirements for establishing, implementing, maintaining and continually improving an AI management system within an organization. It is a management system standard, in the same family of thinking as ISO/IEC 27001 for information security. It specifies how an organization manages AI through policy, roles, risk assessment, controls and continual improvement, rather than prescribing technical settings for any one system.

ISO states that certification is optional: an organization may choose it when it wants independent confirmation that its AI management system meets the standard's requirements. Certification is performed by independent certification bodies. Consultants, including us, cannot certify anyone. What we can do is help a client review its practices against the standard's structure and assemble the records an auditor or customer may ask for.

NIST AI Risk Management Framework

The U.S. National Institute of Standards and Technology released AI RMF 1.0 on January 26, 2023. NIST describes it as intended for voluntary use, to help organizations build trustworthiness into the design, development, use and evaluation of AI products, services and systems. The framework is organized around four functions: Govern, Map, Measure and Manage. NIST also published a Generative AI Profile, NIST-AI-600-1, on July 26, 2024, to help organizations identify risks specific to generative AI. NIST's page states that AI RMF 1.0 is being revised as part of the White House AI Action Plan, so check for the current version before citing it.

The AI RMF matters for evidence because other documents point to it. The CSA matrix maps to it, and questionnaires from larger customers often use its vocabulary. Practically, Govern is where the named owner, policy and inventory live. Map is the use case and context record. Measure covers testing. Manage covers the response when something goes wrong.

A change in general liability forms

A trade publication for independent insurance agents reports that Verisk's ISO is rolling out three optional endorsements for commercial general liability policies that let carriers exclude generative AI exposures, with a January 2026 edition date. According to that report, one form (CG 40 47) excludes bodily injury, property damage, and personal and advertising injury arising out of generative AI under the commercial general liability coverage part; one (CG 40 48) excludes only personal and advertising injury; and one (CG 35 08) applies to products and completed operations. Verisk's stated purpose is to give carriers a way to generally exclude this emerging exposure.

Two cautions. These are optional forms that each carrier decides whether to use, and carriers may also file their own wording, so what applies to your policy depends on your policy. And we are reading a trade report, not the forms. If your renewal documents include an AI-related endorsement, ask your broker what it means for your business. We cannot interpret exclusions or tell you what is covered.

How to prepare the records

Start small and make each record something a person can maintain. A spreadsheet that is current beats a platform that is not.

  • Inventory: one row per AI tool or agent. Columns: name, vendor, purpose, owner, users, data classes touched, systems connected, date added, date last reviewed. Include AI features inside software you already own.
  • Permissions: for each agent, a short table of systems, actions and the identity used. Prefer a dedicated identity per agent over a shared or borrowed human account, and apply least privilege.
  • Logs: decide what is logged, where it is kept and for how long. Confirm what your platforms actually record. Many produce activity metadata but not full prompt text, and some do not record which model produced an answer. Write down the gaps; do not discover them during an examination.
  • Human approval points: list actions that require a person (payments, external messages, deletions, access changes, regulated communications), who approves, and where the record is kept.
  • Testing records: dates, scope, method, findings and fixes. Include tests of prompt injection and over-permissioned tools. Our article on securing AI agents with the OWASP Top 10 for Agentic Applications covers what to test.
  • Named owner: a person per system and an executive sponsor for the program, plus an incident path for AI-related events.

Then check the claims. Compare what the business told its insurer or customers on application forms with what the records show. A mismatch found internally is an easy fix. A mismatch found by someone else is not.

What this does not mean

Having the six records does not make a business compliant with FINRA's report, ISO/IEC 42001 or any other source, and it does not mean any loss would be covered. Carriers write their own questionnaires and policy terms; regulators apply their own rules. Your counsel and your broker are the right people to draw conclusions. What good records do is shorten the conversation, make answers consistent, and show that AI is managed rather than improvised.

How Clayton Cybernetics helps

Our team inventories AI use, including shadow tools, documents agent permissions and identities, reviews logging against what each platform can actually record, tests agents for over-permissioning and prompt injection, and assembles an evidence file mapped to the CSA AI Controls Matrix, ISO/IEC 42001 and the NIST AI RMF. We assemble the records that insurers, brokers, enterprise questionnaires and examiners ask for. Your broker and counsel draw the conclusions. A free security assessment is the place to start.

Sources

This article is general information, not legal or insurance advice. It reports what the listed sources say as of November 8, 2026 and does not state that any business is compliant with a rule or covered by a policy. Regulators and standards bodies revise their materials; confirm current versions. Ask your broker and counsel about your own situation. Last reviewed .

Keep reading

Need help with this?

American‑made Super Intelligence: AI your team can learn, trust and deploy.

Super Intelligence