What did the AICPA publish?
A proposal to update the framework CPAs use to evaluate controls over information systems. The AICPA describes it as "a significant modernization" and says the aims are clarity, usability and relevance as AI and automation change third-party risk. The structural changes consolidate duplicated concepts, remove some criteria, add others where emphasis is needed and revise wording so each criterion reads the same way. The AICPA's SOC 2 Working Group is collecting feedback through 15 January 2027.
Which AI topics are named?
The release lists new points of focus addressing AI governance, model monitoring, prompt manipulation, data poisoning, transparency of AI-generated outputs, training data management and the risks of automated decision-making. Points of focus are not criteria; they are the examples an auditor and a service organization use to decide whether a criterion is met. Naming AI in them means an auditor can reasonably ask how an AI feature is governed, even where the criterion itself has not changed.
Why does this matter for a compliance platform specifically?
Because the platform is itself an AI system operating on your evidence. Every vendor in our lineup describes agents that collect evidence, draft policies, answer questionnaires or review vendors. Under the draft's points of focus, the questions a buyer should ask about those agents are the same ones an auditor may ask about any AI in scope: who governs the agent, how is its output reviewed, can a prompt change what it does, what data trained or tunes it, and which decisions does it take without a person. Vanta's AARM specification, Scytale's read-only MCP phase and Secureframe's agent workflows are each partial answers to one of those questions. None of the six pages we read answers all of them.
What should you ask each vendor before 2027?
Four questions, in writing. First, which agent actions are read-only and which change records, with the log that proves it. Second, how a human reviews agent output before it becomes evidence or a questionnaire answer; Scytale's trust center page says AI questionnaire answers are reviewed by a human expert, which is the kind of statement to look for. Third, what the vendor publishes about its own AI governance, and whether its SOC 2 report describes it. Fourth, whether the platform will map its own controls to the 2027 criteria and show you a gap list when the final text lands.
Does the draft change our scores?
No. Our criteria measure what vendors publish today. We have added the seven AI points of focus to the demo question list on the scoring page so you can test vendors against the draft now, and we will revisit the criteria when the AICPA publishes the final text next year.
What is the timeline?
Comments close 15 January 2027. The AICPA says that if the criteria are approved next year they would be effective for engagements beginning on or after 31 December 2027, with early adoption permitted. A company starting a Type II observation period in early 2028 would be among the first examined under the new criteria. A company renewing in 2027 would not, unless it chooses to adopt early.