Short answer: there are no new AI criteria in SOC 2. The AICPA has not issued AI-specific SOC 2 criteria, and there is no separate AI module. The existing Trust Services Criteria still apply, but auditors now apply them to the AI in your system and ask for evidence.
Auditors expect evidence for AI that is part of the in-scope system, whether you built it, embedded a vendor's model, or enabled it through tools your team uses. If you are new to the process, start with our SOC 2 readiness assessment guide.
How to tell whether AI is in your SOC 2 scope
Scope depends on your role. Expectations rise with each one:
- You only use AI tools internally. Your team uses AI tools in its daily work.
- You embed a vendor's AI in your product. A third-party model is part of the service you deliver to customers.
- You build and train your own models. Your own models are part of the in-scope system.
Record which role applies to you in your risk assessment.
What the auditor asks about AI, by criteria area
| Criteria area | What the auditor asks about AI | Evidence to keep |
|---|---|---|
| Control environment and communication (CC1, CC2) | Is there an AI acceptable use policy, and are staff trained on it? | The approved AI acceptable use policy and staff training records |
| Risk assessment and monitoring (CC3, CC4) | Are AI-specific risks assessed, and is model behaviour monitored? | AI-specific risks in the risk register and records of ongoing monitoring of model behaviour |
| Logical access (CC6) | Which AI tools and accounts are in use, including tools staff adopted without approval? | An inventory of AI tools and accounts |
| System operations (CC7) | Is the AI tested for prompt injection and other misuse, and how are AI incidents handled? | Prompt injection and misuse test results and AI incident handling records |
| Change management (CC8) | Are changes to models and prompts controlled, and which model version is in use? | Change control records for models and prompts and a record of the model version in use |
| Vendors and privacy (CC9 and the Privacy criteria) | Were AI vendors reviewed, and how is personal data sent to AI tools handled? | AI vendor due diligence and records of how personal data sent to AI tools is handled |
The gaps companies most often have
- No written AI policy.
- No list of the AI tools in use.
- Customer data pasted into AI tools without a vendor review.
- Model or prompt changes made without change control.
How SOC 2 relates to ISO 42001 for AI
ISO 42001 is the management system standard for AI. SOC 2 is an attestation report on your controls. They are different, and a company can do both. CyberWave GRC supports both: see our SOC 2 consulting page and our ISO 42001 consulting page.
SOC 2 is an attestation, not a certification. Our reports are issued by our licensed US CPA firm partner, and our lead auditor has handled over 200 SOC 1 and SOC 2 audits.
What to do before your next audit
- [ ] Decide your AI role (internal use, embedded vendor AI, or your own models) and record it in your risk assessment.
- [ ] Write and approve an AI acceptable use policy, and train staff on it.
- [ ] Add AI-specific risks to your risk register and monitor model behaviour.
- [ ] Build an inventory of AI tools and accounts, including tools staff adopted without approval.
- [ ] Test for prompt injection and other misuse, and define how AI incidents are handled.
- [ ] Put models and prompts under change control and record the model version in use.
- [ ] Complete due diligence on AI vendors and document how personal data sent to AI tools is handled.
To plan your budget, see our guide to SOC 2 audit cost.
Book a SOC 2 readiness call
If AI is part of your product or your team's tools, check your evidence before the auditor does. Learn more on our SOC 2 consulting page, or book a 30-minute SOC 2 readiness call.