Every Fortune 500 company now has a published AI governance policy. Most were drafted in 2023 or 2024, following regulatory pressure from the EU AI Act's risk classification framework and voluntary commitments made to the US AI Safety Institute. Two years later, there is enough enforcement data to distinguish the policies that function from those that are governance theatre.
The enforced policies cluster around three controls. First, data classification gates: which categories of data can be processed by external AI systems. Every company with a material AI governance programme has implemented — and is actually enforcing — a prohibition on inputting customer PII, regulated financial data, or attorney-client privileged material into public LLM APIs. The enforcement mechanism varies from technical controls (DLP tools that block prompts containing PII patterns) to contractual controls (vendor agreements requiring data residency and processing restrictions) to procedural controls (mandatory legal review before any new AI tool processes regulated data).
Second, model documentation requirements: for AI systems used in employment decisions, credit decisions, or regulated product recommendations, companies are maintaining model cards and bias audit records. This is driven by specific regulatory requirements — the EU AI Act's high-risk classification, New York City's Local Law 144 on automated employment decisions, and Colorado's AI law on consequential decisions. Enforcement here is audit-driven; companies facing regulatory scrutiny in these areas have been found with documentation gaps.
Third, incident response procedures: what happens when an AI system produces a harmful, incorrect, or legally problematic output. The companies with mature governance have incident classification frameworks (severity 1 through 4), rollback procedures for deployed models, and communication protocols for affected customers. The majority of companies with published policies do not have tested incident response procedures.
The gap between published policy and operational reality is widest in two areas: shadow AI (employees using personal AI tools for work tasks outside corporate controls, which surveys suggest affects 40 to 60 percent of knowledge workers) and supplier risk (third-party vendors using AI internally in ways that affect outputs delivered to the enterprise, where audit rights are often absent). These are structural problems that policies alone cannot address — they require technical controls that most organisations have not yet implemented.