SOC 2 for AI companies

AI companies face a new assurance stack: SOC 2 for the platform, ISO 42001 for the AI management system, and procurement teams asking about both. Here's how they fit together.

AI companies typically need SOC 2 Type 2 plus ISO/IEC 42001: SOC 2 attests your platform's security controls; ISO 42001 certifies your AI management system (model governance, data quality, risk assessment). Several auditors run both engagements together, sharing evidence.

Why AI companies get asked for both

Enterprise buyers of AI products run two diligence tracks at once. Security review asks the classic questions — access control, data handling, incident response — answered by a SOC 2 Type 2 report. AI governance review asks newer ones — how models are trained and evaluated, how training data is sourced and protected, how you assess bias and safety risks — answered by ISO/IEC 42001, the 2023 international standard for AI management systems. EU AI Act procurement pressure is accelerating the second track: deployers of high-risk AI systems must demonstrate governance, and a 42001 certificate is the cleanest evidence.

The combined SOC 2 + ISO 42001 path

The two frameworks overlap heavily on the management-system layer: risk assessment, policies, internal audit, management review, vendor oversight. A combined engagement typically:

1. Scopes once. One system boundary covering the platform and the AI lifecycle (data ingestion, training, evaluation, deployment, monitoring).
2. Tests shared controls once. Access management, change control, logging, and risk assessment evidence serves both frameworks.
3. Runs two fieldwork tracks. The SOC 2 examination (AICPA attestation) and the ISO 42001 certification audit (accredited certification body) are separate opinions with separate reports — but one evidence set feeds both.

See our combined-audit guide for the general pattern — the mechanics are the same, with 42001's AI-specific controls (data quality, model validation, responsible-AI objectives) as the delta.

Which auditors cover AI

From our directory's public materials: A-LIGN and Schellman publicly list ISO 42001 alongside SOC 2; BARR Advisory lists ISO 42001 as well. When scoping, ask specifically: which entity performs the 42001 certification audit (it must be an accredited certification body, a different credential than the CPA attestation), whether the same evidence set serves both, and how AI-specific risks (training-data provenance, model evaluation, red-teaming) are tested.

Scoping advice for AI startups

• Sequence by revenue pressure. If deals stall on security review today, run SOC 2 first and add 42001 in the next cycle. If buyers already ask for AI governance, scope both together.
• Define the AI system boundary tightly. Training pipelines, evaluation harnesses, and inference infrastructure are in scope; exploratory research notebooks usually aren't. The boundary decision drives most of the cost.
• Document training-data provenance now. Data sourcing, licensing, consent, and PII handling are the controls AI auditors probe first — and the hardest to reconstruct retroactively.
• Don't confuse SOC 2 with AI safety. SOC 2 says your controls operated; it says nothing about whether your model is safe or unbiased. Say so plainly to buyers — credibility beats overclaiming.