From SOC 2 to AI: Why Traditional Compliance Falls Short

The compliance frameworks we inherited

SOC 2 Type II is one of the most successful compliance frameworks ever created. When a SaaS vendor hands you a clean report, you know an independent auditor observed their controls, security, availability, processing integrity, confidentiality, privacy, operating effectively over six to twelve months.

The framework assumes a clear model: humans design systems, humans operate them, humans review controls, and auditors verify the humans did what they said they would do.

ISO 27001 extends this with a comprehensive information security management system. ISO 42001, published in 2023, was the first AI-specific management system standard, providing an organisational framework for responsible AI governance.

These frameworks served us well when the operator was always a person. The operator is no longer always a person.

Five places where SOC 2 breaks down

The Trust Services Criteria were designed for a world where changes happen at human speed, operators have identities, behaviour is deterministic, and vendor relationships are stable. AI agents violate every one of these assumptions.

SOC 2 trust principles to AI equivalents

Each of the five SOC 2 Trust Services Criteria has an AI-native counterpart. The concepts translate, but the implementation must be fundamentally different.

The compliance evolution

From cloud-era controls to AI-native evidence generation, the frameworks are converging, but the tooling gap remains.

The three-way standards gap

SOC 2 cannot certify AI controls. ISO 42001 provides governance but not evidence. The EU AI Act mandates evidence but not infrastructure. The result is a gap that no single framework closes.

What AI-native compliance requires

Closing this gap requires compliance infrastructure designed for autonomous systems from the ground up, not human-speed frameworks with AI amendments bolted on.

Cryptographic evidence for non-deterministic systems

You cannot prove that a non-deterministic system will produce a specific output. But you can prove the process: what inputs were received, what policies were evaluated, what tools were called, what approvals were granted. Each event is recorded with a SHA-256 hash linked to the preceding event, creating a tamper-evident chain.

This is what VOLT specifies. It shifts the integrity guarantee from the output to the process. A SOC 2 auditor cannot verify that an AI agent will always produce the same output. But a VOLT Evidence Bundle lets that auditor verify the complete, unmodified chain of events that produced a specific output, the same verification your auditor can run directly.

Continuous observability across every decision layer

For AI agents, "controls operating effectively" means policy gates evaluated correctly, permissions enforced properly, human oversight triggered when required, and escalation paths followed as designed, across every decision, continuously, at machine speed.

AOCL defines 11 control layers (L0 through L10) producing structured, traceable decision records at every policy evaluation, every delegation, every human-in-the-loop approval. This is compliance evidence, not application logging.

Graduated certification for a spectrum of maturity

Not every organisation needs the same evidence rigour on day one. VOLT defines a graduated certification framework with four tiers, each building on the last, giving organisations a realistic adoption path from basic event recording through to automated, auditor-ready evidence packages.

The regulatory landscape

EU AI Act, key requirements

• Art. 12: Automatic logging enabling tracing of operations
• Art. 14: Human oversight with real-time intervention capabilities
• Art. 15: Accuracy, robustness, and cybersecurity measures

Penalties for non-compliance with high-risk system requirements can reach up to EUR 15 million or 3% of worldwide turnover.

NIST AI 600-1, published in January 2024, specifically addresses generative AI risk and underscores the regulatory direction toward verifiable AI operations. For government buyers, FedRAMP's evolving AI guidance adds another layer of compliance expectation.

The compliance industry is aware of the problem. The Big Four are building AI audit practices, but their auditors consistently report the same challenge: the frameworks they are trained to apply were not designed for autonomous systems, and the tooling to generate audit evidence does not exist at the rigour they require.

Compliance as competitive advantage

We have seen this pattern before. In the early cloud era, the first SaaS vendors to achieve SOC 2 Type II won enterprise deals that their competitors could not even bid on. Compliance was not a cost centre, it was a market access credential.

The same dynamic is emerging for AI. Enterprise buyers are beginning to ask questions that existing frameworks cannot answer: How do you prove what your agents did? Can you demonstrate continuous human oversight? Can you export a tamper-evident audit trail for a specific decision?

Organisations that can answer these questions with verifiable evidence, not slide decks and promises, will win the enterprise deals that demand it.

The complete posture

The compliance stack that gets there is not SOC 2 alone, not ISO 42001 alone, and not EU AI Act readiness alone. It is all three, unified by an AI-native evidence layer.

The frameworks we inherited are not broken. They are incomplete. The compliance infrastructure must evolve with the systems it governs.

Quox (quox.ai) builds trust infrastructure for AI agent operations. Its open protocols, AEE for standardised agent messaging, AOCL for orchestration control and observability, VOLT for cryptographic evidence chains, and WARD for independent witnessing, provide the accountability architecture that autonomous systems require.