Vanta and Drata automate evidence collection for the classic Common Criteria — access, change management, monitoring. Then your auditor (or your customer's CISO) asks how you control the LLM features you shipped this year. That section doesn't auto-collect. We give you the evidence pack that plugs into your existing SOC 2 program.
SOC 2 doesn't have an "AI section" yet. What auditors actually do is map AI-related risks onto existing Trust Services Criteria — usually CC3 (risk assessment), CC6 (logical access and data), CC7 (monitoring and incident response). Below are the questions clients tell us keep coming back from their auditor's PBC list.
The entity identifies risks that could affect achievement of objectives — including risks introduced by third-party model providers, prompt injection, hallucination affecting users, and PII leakage to LLM APIs.
"Show us your AI risk register and the controls you put in place for each row."Logical access controls restrict access to information assets. For LLM endpoints this means: which services hold the API key, what data classes can be sent in a prompt, and how that's enforced in code, not just in policy.
"Show the technical control that prevents customer PII from being sent to OpenAI in a prompt."The entity monitors system components and the operation of those components for anomalies indicative of malicious acts, natural disasters, and errors. Auditors want to see actual log records — model, prompt hash or redacted prompt, latency, output classification, user identifier.
"Pull a 30-day sample of LLM calls. Show the fields. Demonstrate retention."Examples your auditor will probe: a hallucinated output reaches a user; a prompt injection exfiltrates data; the model provider has a 4-hour outage. Each needs a documented detection path, a runbook, and an evidence trail of the last test or tabletop.
"Show the runbook for an AI-output-related customer incident, plus evidence it was reviewed in the last 12 months."Anthropic, OpenAI, Mistral and Azure OpenAI are subprocessors. Auditors want their SOC 2 / ISO reports on file, mapping of the data you send each one, and the relevant DPA executed.
"Provide the SOC 2 report and DPA for each model provider in your stack."You probably already have a GRC platform doing the heavy lifting. We don't replace it — we hand it the AI-shaped evidence it doesn't know how to collect on its own.
Both paths produce the same end artifact: an evidence pack you upload to Vanta/Drata as a "control narrative + evidence" bundle, ready for your auditor to walk. The difference is who writes the narrative.
Buy the templates first, decide you want it done for you? We credit the €97 against the managed pack. One email: marc@auditaisdk.com.
The honest version. Big GRC platforms cover a much wider scope; AI-specific consultancies cover deeper engagements. We're the narrow, fast option for the AI section specifically.
| Capability | AuditAI | GRC platform alone | AI-risk consultancy |
|---|---|---|---|
| Cost | €97 – €199 | €8k – €30k / yr | €20k – €100k |
| Time to evidence | Same week | Months | 8 – 12 weeks |
| AI risk register | Pre-mapped to TSC | Generic template | Custom |
| SDK-level call logs | Built in | Not collected | Recommended only |
| Replaces SOC 2 audit | No, plugs into yours | No, just evidence | No, advisory |
| Updates as ISO 42001 / EU AI Act mature | Versioned releases | Slow | Fixed engagement |
pip install auditai-sdk wraps Anthropic, OpenAI, Bedrock, Vertex and Ollama. The control narratives reference whichever you actually use, named explicitly. If you need Azure OpenAI added, mention it in the interview — that adapter shipped in v0.4.