
Reviewed October 30, 2026.
A vendor sends over its SOC 2 report, and your team is ready to approve a new AI feature. Before you move forward, ask a simple question: does the report address the service you are about to use?
The answer takes more than checking the vendor’s name on the cover. You need to understand the system described, the period covered, the controls examined, and what responsibilities remain with your company.
AI adds questions about how a system behaves. Those questions may overlap with SOC 2, but the words “we have a SOC 2 report” do not tell you whether a particular AI feature, model, or use case was evaluated.
Start With the System in the Report
What does SOC 2 cover? The AICPA describes SOC 2 as an examination of a service organization’s system description and controls relevant to security, availability, processing integrity, confidentiality, or privacy. The specific report determines the scope you can rely on. The AICPA SOC 2 guide explains this reporting framework.
Read the system description against the product you are buying. Look for the relevant service, infrastructure, data flows, and outside providers. If the AI feature is missing or the description is unclear, ask the vendor to explain the difference.
Example: A company reviews a report covering a customer-support platform. The vendor recently added an assistant that sends prompts to another provider. The buyer should ask when that feature launched and how its data flow relates to the reported system, rather than assuming that the platform’s existing report answers the question.
Check the Timing and the Evidence
A Type 1 report addresses controls as of a specified date. A Type 2 report also addresses their operating effectiveness over a specified period. Neither should be treated as a prediction that every future release will behave correctly.
Ask your reviewer to identify the opinion, relevant exceptions, and any controls the report expects customers to perform. Keep unresolved items with the vendor review rather than treating receipt of the document as approval. The AICPA SOC resource library includes report examples and resources for report users.
Sample Review Step: Record the report period beside the launch date of the AI feature. If the dates do not overlap, obtain an explanation of the change and the evidence available for it.
This simple comparison can prevent a team from relying on evidence that predates the feature under review.
Ask What Was Tested About the AI
What happens when the assistant is wrong? What can it do without approval? How does the provider check that a change has not weakened an existing safeguard?
For your proposed deployment, request evidence covering:
- The kinds of information the system can access and disclose.
- Whether it can take actions, and which actions need human approval.
- How incorrect or unsupported answers are identified and handled.
- What testing was performed, on which configuration, and when.
- How issues are corrected and changes are evaluated.
Example: An assistant that drafts an internal email presents a different approval problem from one that sends messages to customers. Ask the vendor to demonstrate the permitted workflow for your configuration. A test of drafting alone would not answer your question about sending.
These are practical review suggestions, not a universal list of mandatory audit tests.
Understand What an AI Framework Adds
“AI assurance” can refer to different work. Ask which framework, criteria, evaluation method, and deliverable the provider means before comparing offers.
NIST’s AI Risk Management Framework is voluntary and organizes risk management around Govern, Map, Measure, and Manage. Its Generative AI Profile provides additional guidance for generative AI risks. Using that guidance does not, by itself, establish that an independent examination occurred. See NIST’s framework overview and Generative AI Profile announcement.
A provider may instead offer certification or an assessment under another program. Obtain the underlying scope, criteria, findings, and limitations. Compare those with the question you need answered, rather than comparing logos.
Identify Your Own Responsibilities
Before rollout, assign responsibility for:
- Configuring and reviewing access.
- Approving new data sources and tool connections.
- Disabling the feature if it produces unacceptable results.
Document those responsibilities before rollout. Our user access review guidance is a useful starting point for assigning owners and reviewing permissions. Include tool connections and service accounts in the discussion.
Sample Control: The business owner approves new connections to the AI tool, and the system owner retains the configuration record and evidence of the review. The organization defines the review frequency based on its risks and commitments.
Make a Decision You Can Explain
Keep a short decision record:
- Intended use: the feature, configuration, data, and permitted actions.
- Evidence reviewed: what the SOC 2 report supports and what the additional AI evidence covers.
- Open gaps: unanswered questions, owners, and required follow-up.
- Decision: whether the use can proceed, under what conditions, and who approved it.
Resolve or explicitly address the open issues in the decision before production use.
SOC 2 and AI-focused work can be complementary. Their value depends on whether their scope and evidence fit your actual deployment.
Need help organizing a provider review or defining your SOC 2 scope? Contact a SOC consultant to discuss the questions your organization needs answered.