A new survey says 55 percent of companies wrote the AI policy and only 19 percent can prove it was followed. Regulators, auditors, and enterprise clients are about to ask which group you are in.
Most companies can hand you an AI policy. Far fewer can hand you proof it was followed. That gap is the entire finding of a new survey, and it lands hardest on the industries that get audited.
Per Arctera's State of AI Governance 2026, 78 percent of organizations using AI expect communications risk to rise. The sharper number sits underneath that. 55 percent have the core pieces on paper, the policy, the training, the review step. Only 19 percent have the logging, retention, detection, and scoring controls needed to prove what actually happened. The survey was run by Hanover Research in May 2026 across 500 compliance decision-makers in finance, healthcare, and energy across the Americas and EMEA. These are regulated shops, and 81 percent of them cannot show their work.
The timing is the problem. Gartner projects that 40 percent of enterprise applications will carry embedded AI agents by the end of 2026, up from under 5 percent in 2025. More software is taking actions on its own, and fewer humans are watching each one. A policy that lives in a slide deck does nothing when an agent files something, sends something, or approves something at 2 a.m.
A policy tells people what they are supposed to do. Evidence shows what the system actually did. In finance, healthcare, and legal work, the second one gets tested. Regulators ask for records. Opposing counsel asks for records. Your enterprise clients ask for records in the security review before they sign. Intentions are not admissible.
The fix is an evidence layer, and most teams can stand one up faster than they expect.
| Control on paper | Evidence that holds up in an audit |
| Written AI use policy | An immutable log of every model call and who approved it |
| Staff training deck | Records showing trained staff actually followed the steps |
| "A human reviews the output" | Timestamps proving a human reviewed before anything shipped |
| Retention rule | Prompts and outputs you can retrieve on request |
Three moves, in plain terms. Log every model interaction, the prompt, the output, and the person who signed off. Retain prompts and outputs under the same rules you already apply to email. Make review a recorded event, not a good habit people mean to keep.
That is what AI drafts, you verify looks like when someone asks for proof. You verified, and you kept the receipt. The 19 percent who built that layer will pass the audit the other 81 percent are about to fail.
AI, Properly is the Ayvede field report.
Advise. Vet. Deliver.
Ask a question in writing