LLM apps · RAG · AI agents

SOC 2 for AI companies: what changes when your product runs on models.

LLM providers, vector databases, prompts-as-code and autonomous agents all live inside your SOC 2 system boundary. Here's how to control them — and prove it.

Where AI products trip up SOC 2

AI componentSOC 2 concernControl approach
Third-party LLM APIsSubservice organization, data retention, training on customer dataVendor risk review, SOC 2 report review, zero-retention / no-training terms, carve-out in system description
Prompts & system instructionsUnreviewed changes alter system behaviourPrompts in version control, peer review, change tickets, deployment approvals
RAG / vector storesCross-tenant data exposureTenant isolation, access-filtered retrieval, encryption, logging
Model inputs/outputsPrompt injection, sensitive data leakageInput validation, output filtering, DLP, red-team / pen testing
AI agents & tool useOver-privileged actionsLeast-privilege tool permissions, human approval for high-impact actions, audit logs
Fine-tuning & training dataConfidentiality and privacy of customer dataData classification, consent, retention and deletion procedures

Writing the system description for an AI product

Your SOC 2 system description must explain how AI components process customer data. We draft it so it is accurate for auditors and reassuring for buyers — describing model providers, data flows, retention, and the complementary controls you expect from customers and subservice organizations.

Going beyond SOC 2

SOC 2 answers the security question. Buyers increasingly also want AI-specific assurance. We layer AI governance aligned to NIST AI RMF, and ISO 42001 when certification is needed, onto the same control set.

Frequently asked questions

Does SOC 2 cover AI systems?

SOC 2 has no AI-specific criteria, but your AI features sit inside your system boundary, so the Trust Services Criteria apply to them. Auditors increasingly ask how AI components, model providers and training data are governed under existing controls.

Are OpenAI, Anthropic or other model providers in scope for our SOC 2?

They are typically subservice organizations or vendors. You need vendor risk assessments, a review of their own SOC 2 reports, data-processing terms and a clear description of how they fit into your system — usually using the carve-out method.

What AI risks do auditors and buyers ask about?

Common questions cover customer data used for training, data retention by model providers, prompt injection, output data leakage, access to model and prompt configurations, change management for prompts and models, and monitoring of AI behaviour.

Ready to get audit-ready — and AI-ready?

Book a free 30-minute scoping call. You'll leave with a clear view of your gaps, timeline and realistic budget — no obligation.