Skip to content
be-human.ai
← Back to the journal

Human oversight

SOC 2 Evidence for AI Agents: What Auditors Will Ask For

Learn which logs, access reviews, change records and approval records auditors usually request for AI agents, and how to keep them ready.

Be Human typographic cover with the title: SOC 2 Evidence for AI Agents: What Auditors Will Ask For
Conceptual illustration created with AI · Be Human

A SOC 2 report does not certify an AI agent on its own, but any agent that reads company data, calls tools or writes to customers falls inside the systems an auditor tests. For each control in scope, the auditor will ask you to show who could reach the agent, what it did, who approved changes to its tools and memory, and how exceptions were handled. Keep those records as a by-product of normal operation so they can be produced on request instead of assembled in a rush before fieldwork.

Key takeaways

  • Treat each agent as a system identity inside your SOC 2 boundary, with its own named owner, permissions and log trail.
  • Auditors sample four evidence families: activity logs, access reviews, change records and approval records for high-risk actions.
  • Evidence counts only when it is complete, time-stamped, attributable to a named identity and kept under a written retention rule.
  • Run an evidence drill every quarter: pick one agent action at random and check whether you can trace it from request to log to approval.
  • Ask any AI vendor which evidence it produces by default and which you must request, before you sign.

What SOC 2 checks and where agents fit

SOC 2 reports are built on the Trust Services Criteria. Security is always in scope, and availability, processing integrity, confidentiality and privacy are added when a company commits to them with customers. An agent usually touches several criteria at once: logical access (who or what can act), change management (how its behavior is altered), monitoring (how you see what it does) and incident response (how you stop it and recover).

The report type matters as well. A Type I report describes whether controls are designed appropriately at a point in time. A Type II report tests whether controls operated effectively across an observation period. For agents, Type II is the harder case, because an auditor will not accept a screenshot taken the day before fieldwork. They will want samples spread across the period, including any changes and incidents that occurred during it.

Ask your auditor early which criteria cover your AI systems and whether the agent appears in the system description. Scope decided late is expensive to correct.

The logs auditors will ask for

Logs answer the basic question: what did the agent do, when, and on whose behalf. The minimum useful record for each action includes a timestamp in one time standard, the agent's own identity rather than a shared service account, the tool or system called, the target object, the outcome, and a correlation identifier that links the action to the request or conversation that caused it. Tracing concepts such as traces and spans are a good model here, as described in the OpenTelemetry observability primer.

Be careful about what goes into the log. Full transcripts may contain personal data, and an audit log that leaks customer details creates a second problem. Decide which fields are stored in full, which are summarized or hashed, and which are redacted. The PII redaction guide walks through that tradeoff.

Auditors also test completeness. They may pick an action from a customer ticket and ask you to find the matching agent log entry. If you cannot connect the two, the control is weak no matter how polished the logs look on their own.

A practical checklist for each logged action:

  • Agent identity on every action, separate from human accounts
  • Tool name, parameters (redacted where needed), target object and result
  • Link to the triggering request or conversation
  • Model and prompt version active at the time of the action
  • Alerts raised, and whether a person acknowledged them

Access reviews show who can reach what

The access review is where auditors spend much of their time on agents. The question is not only whether access was granted correctly on the first day, but whether it is still appropriate. Each agent should have a documented list of permitted systems, actions and data sets, reviewed on a schedule by a named owner who did not build the agent.

The evidence for a review is an export of permissions as they stood at review time, the reviewer's decision on each line, the reason for any removal, and the ticket that executed the removal. Authorization guidance from OWASP stresses deny by default and enforcement where the resource lives, not only in the agent's instructions. If the prompt says the agent should not issue refunds above a limit but the tool permission allows it, the auditor will treat the permission as the real control.

Illustrative scenario: a support agent was given read access to order status and temporary write access to refunds during a pilot. The next quarterly review finds the refund scope still active. The owner removes it, records the reason in the review file and attaches the removal ticket. That sequence, more than the original grant, is what an auditor will sample.

The access review guide provides a repeatable template for this cycle.

Change records for prompts, tools and memory

Auditors treat changes to agent behavior as production changes. A new instruction, a new tool, a new memory source or a model version switch can change outcomes as much as a code release does. Each change needs a record: what changed, who requested it, who approved it, the evaluation run before release, the release date and the rollback path.

The NIST AI Risk Management Framework is a useful reference for organizing this work under its governance and measurement functions. It is a framework to adapt, not a checklist to copy. The practical requirement is that the change record points to the version that actually ran. If the logs show one model identifier and the change file names another, the evidence contradicts itself.

Memory changes deserve the same discipline. Adding a document to the agent's retrieval store changes what the agent knows, and removing a person's data from memory is an erasure event with its own evidence trail.

A release routine that produces clean change records:

  • Log every change request with an owner and a stated reason
  • Run a fixed evaluation set before release and store the result with the change
  • Record version identifiers for prompts, tools, model and memory sources
  • Keep the previous version available for rollback, and note that the rollback was tested
  • Close the record only after the first monitored period confirms the change behaved as expected

Approval records for high-risk actions

Some actions should never run without a human decision: moving money, changing contract terms, deleting records or messaging a regulated group. For these, the approval record is audit evidence in its own right. It should show the named approver, the time of approval, the exact action and parameters the approver saw, and what happened next. An approval given through a shared inbox, or a loose confirmation in a chat window with no link to the action, will not satisfy a careful reviewer.

Decide the approval list before the agent goes live. A simple decision guide:

  • Does the action move money, change a legal commitment or delete data? Require approval.
  • Does it reach customers or regulated personal data for the first time? Require approval until evaluation results support relaxing the rule.
  • Is it reversible within a known window and low in impact? Log it and sample it in review.
  • Is it read-only within an approved data set? Log it and review by exception.

The approval gate guidance explains where the stop point belongs in a workflow.

How Be Human keeps evidence ready in managed operations

Be Human designs AI agents with their identity, memory, channels, tools, permissions and operating context under human control. In managed operations, the work that keeps an agent useful also produces its evidence: monitoring creates activity records, recurring permission reviews produce access review files, controlled improvements go through change records, and incident response leaves a documented trail. Be Human does not promise an audit outcome, since the auditor and your defined scope decide that. What it does is set up these records before the first production action, so your team can hand over samples instead of reconstructing history.

Build your evidence pack in one quarter

An evidence pack does not need to be large. It needs to be accurate, current and easy to navigate. A workable sequence for the next quarter:

  • Name one owner per agent and write down its permitted systems and actions
  • Confirm with your auditor which SOC 2 criteria your agents fall under
  • Turn on identity-tagged logging and set a retention rule that your policy and contracts support
  • Schedule the first access review and file its export, decisions and removal tickets
  • Trace one customer-facing action from request to log to approval, and time how long that takes
  • Store the change history for every prompt, tool and memory update

If you are not yet sure which agent task deserves this process first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Does a SOC 2 report certify an AI agent? No. SOC 2 is an attestation report about controls at a service organization, not a certificate for a single product or system. An auditor tests whether the controls around the agent, such as access, change management, monitoring and incident response, operated as described.

Do we need a separate SOC 2 report for each agent? Usually not. Agents normally sit inside the system description of the company or service that operates them, and the auditor samples evidence across that scope. The question to settle early is whether each agent is included in the boundary, and which criteria apply to it.

How far back will auditors look at agent logs? For a Type II report, they will look across the full observation period and sample from it, so retention has to cover that window. Your retention rule should be written down, applied consistently and aligned with your contracts and privacy obligations, including any limits on how long personal data may be kept.

Can a vendor's SOC 2 report cover our agent? It can cover the vendor's own controls, which is useful, but it does not cover your configuration, your permissions or your approval decisions. Read the report's scope and its complementary user entity controls to see which obligations stay with your team.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →