SOC 2 Trust Services Criteria for AI Systems
Extended SOC 2 framework addressing AI-specific risks. Covers model governance, data provenance, and algorithmic accountability for AI service providers.
What SOC 2 + AI draws on
This framework pulls on all three pillars — which is why running them as three separate tools means reconciling three sets of evidence at audit time.
Compliance
Security
What your auditor cites,
and what produces the evidence
The regulator's text is quoted below in italic, exactly as written. What follows each one is what the platform records, detects, or proves — not a claim about your compliance status, which no tool can confer.
AI System Monitoring
Entity MUST monitor AI system components for anomalies including model performance and data drift.
Rules catch what you already thought of. Behavioural models baseline every user and application against their own history and their peer group, then score deviations in real time — so a service account reading tables it has never touched surfaces without anyone having written a rule for it first. That is what makes monitoring hold up as access patterns change, instead of decaying into a ruleset nobody maintains.
AI Input Validation
REQUIRES validation and logging of all AI system inputs to prevent adversarial attacks.
Rules catch what you already thought of. Behavioural models baseline every user and application against their own history and their peer group, then score deviations in real time — so a service account reading tables it has never touched surfaces without anyone having written a rule for it first. That is what makes monitoring hold up as access patterns change, instead of decaying into a ruleset nobody maintains.
Model Change Management
AI model updates MUST be authorized, tested, and documented with complete audit trails.
Change control fails at the evidence step far more often than at the approval step. Running SELECT 'CR:12345' WHERE 1 = 0 before a change ties every subsequent statement in that session to the request that authorised it — no agents, no application changes, no database configuration. Schema and configuration changes are captured as they happen, so an unapproved DDL is visible rather than discovered at the next review.
AI Output Integrity
Entity MUST implement controls to ensure AI output integrity and detect manipulation.
Rules catch what you already thought of. Behavioural models baseline every user and application against their own history and their peer group, then score deviations in real time — so a service account reading tables it has never touched surfaces without anyone having written a rule for it first. That is what makes monitoring hold up as access patterns change, instead of decaying into a ruleset nobody maintains.
What SOC 2 + AI covers
This instrument defines no data category of its own. Scope is whatever the service organisation defines in its system description and the criteria it selects. There is no fixed data class — which is why two SOC 2 reports can cover very different things.
How SOC 2 + AI is enforced
Every figure below is the ceiling the instrument publishes about itself, not a prediction of what anything would cost. Enforced by No regulator. Your customers, through the contract and the renewal.
A voluntary standard. No regulator enforces it.
SOC 2 is an attestation performed by an independent CPA firm against criteria you select. Nobody can sanction you for failing it. What happens instead is that the report carries exceptions, procurement asks about them, and the contract that required a clean report does not renew.
Where a customer contract warrants that you hold a clean SOC 2, failing it becomes a breach of that contract — which is a real exposure, just not a regulatory one.
Enforcement data reviewed August 2026. Several figures are indexed annually and move.
Built for SOC 2 + AI,
not configured for it afterwards
AI Trust Services Monitoring
SOC 2 controls extended for AI system components
SOC 2 + AI Evidence Package
AI-specific trust service criteria documentation
AI Service Data
Identify model endpoints, training pipelines, and outputs
AI Anomaly Detection
AI-powered detection of unusual AI system behavior
Other AI & Technology frameworks
Walk into the SOC 2 + AI audit knowing the answer
228 cited requirements across 57 frameworks are mapped to the controls that evidence them. A fixed-fee gap assessment tells you which of them you can already prove today.