SOC 2 Is Adding AI Criteria: What the AICPA's 2026 Proposal Means for SaaS Companies

What the AICPA proposed

The AICPA is the US body for accountants, and it sets the rules for SOC 2. A SOC 2 report is an independent review of how a company protects customer data. The AICPA released an exposure draft (a public draft for comment) that revises the Trust Services Criteria, the checklist SOC 2 auditors work from.

The headline change is a set of new points of focus for systems that use AI. They cover AI governance, model monitoring, prompt manipulation, data poisoning and training data management. In plain terms: who is in charge of AI and how you watch what it does. Also, how you stop people tricking it and how you keep the data it learns from clean.

It is a proposal, not a final standard. Comments are due by 15 January 2027. If approved as drafted, it would apply to SOC 2 audits starting on or after 31 December 2027. Companies may adopt it early.

What it means in practice

Points of focus guide what auditors look at. They are not a checklist you must tick one by one. But they shape the questions you will get. For a software company with AI features, expect auditors and customers to ask:

• Who owns AI risk, and is there a policy for how AI is built and used?
• How do you watch what the AI does and the quality of its answers once it is live?
• How do you protect against prompt injection (hidden orders slipped into text the AI reads) and other tricks?
• Where does training and fine-tuning data come from, and how is it protected from poisoning?
• How are AI features tested before release and after changes?

Why you should start now

Large customers already ask about AI in their security questionnaires, well ahead of any SOC 2 change. Controls you put in place now will help close deals today. They will also make the audit easier later. A Type II report checks that controls worked over a period of months. So waiting until 2028 to start would leave you short.

A practical starting checklist

You do not need to wait for the final rules. These seven steps are a sensible start.

1. List every AI feature, model and AI vendor in your product and operations.
2. Write a short AI use and development policy, with a named owner.
3. Add AI risks to your existing risk assessment.
4. Test AI features for prompt injection and data leakage (see AI red teaming).
5. Log and monitor AI inputs, outputs and agent actions.
6. Document training data sources and access controls.
7. Review AI vendors the way you review other critical vendors.

This also lines up with ISO/IEC 42001, the international standard for managing AI, if your customers ask for it.

How NextGarrison helps

We already help software teams prepare for SOC 2. We also test AI features for the exact risks named in the proposal. We can add AI controls to your readiness program and run the AI security test your auditor and customers will want to see. SOC 2 reports themselves are issued by independent CPA firms (licensed accounting firms). We get you ready. New to SOC 2? Start with what SOC 2 compliance is.