SOC 2 Evidence for AI Agent Vendors: What to Request Before Sharing Memory
Which SOC 2 reports, control mappings and access logs to request from an AI agent vendor before it holds your company memory, and what each one proves.

A SOC 2 report shows that an independent auditor tested a vendor's security controls against defined criteria, which makes it the most common third-party evidence buyers ask for. For an AI agent vendor, it answers only part of the question. Before you let a vendor hold company memory, request the full report with its scope and exceptions, the subservice organizations and controls you must run yourself, and access logs that show who and what touched that memory.
Key takeaways
- Ask for the full SOC 2 report with its scope, test period and exceptions, not a summary page or a trust badge.
- Confirm that the report covers the systems that store and retrieve company memory, including the tool gateway and logging pipeline.
- Map each control family to a specific agent risk: who can read memory, which tools the agent can call, and how changes are approved.
- Request the subservice list and the complementary controls you are expected to operate, and assign an owner to each.
- Treat SOC 2 as one input. Agent-specific evidence, such as tool-call records and approval logs, still has to be requested separately.
What a SOC 2 report can and cannot prove
A SOC 2 report is an examination by an independent CPA firm of whether a service organization's controls meet the AICPA Trust Services Criteria. Security is always in scope. Availability, processing integrity, confidentiality and privacy are added when the vendor chooses them. The report describes the controls, tests whether they were designed appropriately and, in a Type II report, tests whether they operated effectively over a period.
The report does not certify that a model gives correct answers, that stored memory is lawful in your jurisdiction, or that an agent will refuse a harmful instruction. Read it as evidence about the vendor's operational discipline, then ask separate questions about the agent's behavior and permissions. For a broader view of what auditors look for in agent systems, see SOC 2 Evidence for AI Agents: What Auditors Will Ask For.
Which documents to request first
Ask for these documents before any feature discussion:
- The complete SOC 2 report, including the auditor's opinion, the system description and the tests performed.
- The period covered by the report, plus a bridge letter from management if that period ended some time ago and the vendor has changed its systems since.
- The exceptions list, with management's response and remediation status for each exception.
- The subservice organizations list, stating which are carved out of scope and which are included.
- The complementary user entity controls, which are the controls the vendor expects you to run. Each one becomes a task for your team.
- The most recent penetration test summary, with its scope, the type of tester and whether critical findings were closed.
- The current subprocessor list and the regions where memory, logs and backups are stored. For a closer look at this list, see AI Vendor Subprocessors: Who Else Handles Your Company Data?
- Excerpts from the vendor's access review, change management and incident response policies.
Also check the system description itself. It should name the memory store, the retrieval index, the tool gateway and the logging pipeline. If any of those components are missing from the description, the report does not cover the part of the product that holds your data. Data processing terms should be reviewed alongside the report, as described in DPA Check for Superintelligence Vendors: Verify Before Signing.
How to map SOC 2 controls to agent risks
A generic report is organized around Common Criteria families. Translate those families into the questions that matter for an agent. The NIST AI Risk Management Framework is a useful reference for organizing these risks into governance, mapping, measurement and management activities.
- Step one: list what the agent will hold, usually memory, credentials, tool permissions and channel access.
- Step two: for each item, find the governing family. Logical access (CC6) covers who and what can read or write memory and call tools. System operations (CC7) covers monitoring and detection of unusual activity. Change management (CC8) covers prompts, tool definitions and model versions. Risk mitigation (CC9) covers vendor and subprocessor oversight.
- Step three: locate the control in the report, the test performed and the result.
- Step four: if a control is absent, carved out or has an exception, record it as an open risk with a named owner.
- Step five: decide whether the gap blocks the pilot, limits the agent's permissions, or needs a written compensating control.
Ask specifically whether a model upgrade counts as a change under the vendor's change process, and whether you receive notice before a new model version reaches your agent. A model change can alter behavior without any change to your configuration, so it belongs in the change record.
Which access logs to request and what they should show
A report can state that access reviews happen. Logs show what happened to your memory. Ask the vendor to demonstrate, for a sample period you choose, records that answer these questions: which identity read or wrote each memory item, which agent identity called which tool and with what scope, which actions required approval and who approved them, which records were deleted and when, and which administrators changed permissions.
The logs also need their own controls. Ask how long they are kept, who can read them, and whether personal data inside prompts is redacted before storage. The OpenTelemetry concepts primer describes traces, spans and logs as the way a system records each step of an operation, which is the level of detail to request. For the redaction question, see PII Redaction for AI Agent Logs: Keep Audit Detail Without Leaks.
Test the logs during the trial. Ask the vendor to reconstruct one agent action using only the log. If the reconstruction needs an engineer and a full day, the logs are not yet usable as operational evidence.
Illustrative scenario: reading an exception without rejecting the vendor
Illustrative scenario: an operations director at a mid-sized distribution company receives a vendor package for an agent that summarizes supplier emails and updates a shared company memory. The SOC 2 report is clean except for one exception. Access reviews for the memory database were not completed on schedule during part of the test period. Management states that the review cadence was changed and corrected.
She does not reject the vendor. She requests the review evidence from the corrected cadence, keeps the agent's write access to memory disabled until that evidence arrives, and adds a monthly permission review to her own calendar. The exception is recorded with an owner and a closing condition, which is the same treatment it would receive in an internal audit.
The following decision guide applies the same logic to other findings:
- If the exception touches the memory store or tool credentials, limit the agent to read-only access until remediation evidence arrives.
- If the exception touches only a component outside the agent's path, such as a marketing site, record it and continue the normal review.
- If a subservice organization that stores your memory is carved out, request its own report or confirm its controls directly.
- If the vendor cannot provide a bridge letter for a period that ended long ago, treat the report as historical evidence only.
How Be Human handles vendor evidence in managed operations
Be Human designs each agent around one defined role, with its memory sources, tools, permissions and approval boundaries written down before launch. In managed operations, we keep that record current, review permissions on a recurring schedule, and retain the logs needed to reconstruct an agent action. When a client asks a vendor for evidence, the same record shows which memory stores and tools are in scope. This article does not claim that Be Human holds a SOC 2 report. Ask us for the documentation that applies to the stack we would operate, and we will state plainly what it covers and what it does not.
If you are not yet sure which task should go to an agent first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Is a clean SOC 2 report enough before giving an AI agent vendor access to company memory? A clean report is useful evidence that the vendor runs defined controls, but it is not sufficient on its own. It covers only the systems inside its scope and does not judge whether an agent's answers or memory choices are correct. Pair the report with access logs, tool permission records and a documented approval process before expanding access.
What is the difference between SOC 2 Type I and Type II? A Type I report describes whether controls were designed appropriately as of a specific point in time. A Type II report describes whether controls operated effectively across a test period, with the auditor testing evidence from that period. For company memory that will be stored for a long time, a Type II report whose period covers your intended use is the stronger starting point.
Can a vendor refuse to share the full report? Vendors commonly share the full report under a non-disclosure agreement, and some ask you to sign it before release. If a vendor offers only a summary, a logo or a marketing statement, record that as a missing document. Ask what the vendor will provide and when, and make the answer part of your decision.
Do small companies need SOC 2 evidence from an AI agent vendor? Small companies still hold customer records, staff details and internal knowledge, so the core questions remain. The depth of the review can be smaller: a short document list, a clear view of what the agent can read and do, and a log check during the trial may be enough for a first, narrow role. Expand the review as the agent's permissions grow.
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.


