Loading…
Compliance
SOC 2 Type II is a prerequisite for most enterprise SaaS contracts. The Trust Service Criteria: Security, Availability, Confidentiality, Processing Integrity, and Privacy, apply equally to every component of your product, including the AI layer. An AI system with no audit logging, unmonitored outputs, and a standard consumer LLM API will not survive a SOC 2 audit.
We design AI systems with SOC 2 controls built into the architecture from day one. Every AI interaction is logged and monitored. LLM vendors are evaluated against their SOC 2 scope reports. Access controls, encryption, and incident response procedures are documented before your audit window opens, not after your auditor asks for evidence you don't have.
Tell us what you're building.
The five Trust Service Criteria each place specific requirements on AI components. These are the controls auditors test during a Type II examination.
Encryption in transit (TLS 1.2 minimum, TLS 1.3 preferred) and at rest (AES-256) for all customer data. Access control enforced with MFA on every privileged path. Audit logging captures every AI interaction, who sent what prompt, which model responded, and what data was accessed. The CC6.1 criterion requires documented logical access controls with role-based permissions and automatic de-provisioning.
Uptime SLAs that match the commitments in your customer contracts. For AI components this means graceful degradation when the LLM provider is unavailable, the system must not fail open or expose raw errors to end users. Circuit breakers, retry logic with exponential backoff, and fallback responses are required to meet A1.2 monitoring and incident response obligations.
Data classified by sensitivity, with LLM API calls routed through enterprise tiers that include a Data Processing Agreement. Under C1.1, confidential information must be protected during processing and transmission. Standard consumer ChatGPT and Claude.ai tiers do not carry the contractual commitments required by C1. Enterprise API tiers with a DPA are required.
PI1.2 requires that processing is complete, valid, accurate, timely, and authorised. For AI systems this means output monitoring for anomalous completions, evaluation suites that run before any prompt changes go to production, and documented testing procedures that an auditor can inspect. Unmonitored prompt changes are a finding.
Data minimisation in prompts. Personal data should not appear in LLM calls unless required for the task. PII handling policies must be documented and enforced technically, not just stated in a privacy policy. Under P3.1, the system must collect only the personal information necessary for the specified purpose.
These are the specific controls we design and build for AI systems going through SOC 2 Type II audits. Each one maps directly to Trust Service Criteria that auditors test.
Append-only logs of every prompt, completion, user identity, model version, and timestamp. Written to a separate logging service with write-only access from the application layer. Directly addressable by SOC 2 auditors reviewing the CC7.2 anomaly monitoring criterion.
We confirm that every LLM vendor in your stack has an active enterprise agreement with a Data Processing Agreement. We document which endpoints are covered, which are not, and flag gaps before your audit window opens.
Input validation to prevent prompt injection, output filtering for sensitive data leakage, and rate limiting on inference endpoints. These address CC6.6 (logical access from outside the system boundary) and CC6.8 (unauthorised use of data).
Automated sampling of AI completions with alerting for anomalous patterns: unusual output lengths, unexpected data classifications, or high refusal rates that may indicate adversarial input. Maps to CC7.2 ongoing monitoring.
Documented incident response procedures, access revocation runbooks, and change management processes for AI system updates. SOC 2 auditors test whether controls are operational. Runbooks are the evidence.
Multi-region failover for AI endpoints, queue-based inference decoupling to absorb LLM provider latency spikes, and health check monitoring. We set up alerting thresholds that align with the uptime commitments in your customer contracts.
Each vendor below holds a SOC 2 Type II certification, but the specific services and tiers in scope vary. Verify scope against the vendor's current SOC 2 report before relying on their certification for your own audit.
Note: Consumer-tier LLM products are typically outside the scope of enterprise SOC 2 reports. Confirm scope in writing before using any vendor for customer data.
Enterprise tier includes DPA. Standard API tier has limited contractual commitments for SOC 2 confidentiality criteria.
Enterprise tier includes DPA. Claude.ai consumer accounts are not suitable for customer data.
SOC 2 Type II certified. Bedrock and SageMaker are within scope for most AWS SOC 2 reports.
SOC 2 Type II certified. Vertex AI is within scope. Standard Gemini consumer products are separate.
SOC 2 Type II certified. Azure OpenAI Service is within scope for enterprise agreements.
SOC 2 readiness work has a specific audience. We decline projects that are a poor fit so neither side wastes time.
Internal tools not covered by customer contracts
SOC 2 is about the commitments you make to your customers about how you handle their data. If the AI system is purely internal and you have no enterprise customer contracts requiring SOC 2, this engagement is not the right fit. Start with security fundamentals instead.
Teams at pre-revenue stage without enterprise prospects
A SOC 2 Type II audit requires 6–12 months of operational evidence. Building the controls before you have enterprise customers who require it adds cost and process overhead that slows down early-stage teams. Wait until your sales pipeline has concrete SOC 2 requirements, then build.
Anyone expecting certification from us
We implement the technical controls that go into a SOC 2 audit. The actual SOC 2 report is issued by a licensed CPA firm after an audit of your full system, not just the AI components. We prepare the AI layer; your audit firm evaluates the whole.