Governing AI in regulated organisations

In a regulated environment, the question is rarely "can this be automated?" It is "can we evidence how this was done, to whom, on what basis, and who was accountable?" A system that cannot answer that will not pass review regardless of how well it performs.

Five controls worth insisting on

These are the controls that tend to determine whether an AI deployment survives internal audit.

• Scoped access. Each agent has the narrowest set of tools and data its task requires, granted explicitly and reviewable.
• Explicit approval gates. Consequential actions require a named human approval, captured with the context the approver saw.
• Complete execution logging. What ran, on what inputs, using which sources, producing what output — retained for the period your obligations require.
• Source attribution. Any assertion an agent makes can be traced to the material it relied on.
• Defined evaluation. Agreed criteria for acceptable performance, measured before go-live and monitored afterwards.

Where organisations get this wrong

The most common failure is treating governance as a document rather than a runtime property. A policy stating that humans review AI output is not a control; an approval gate that cannot be bypassed is.

The second most common is granting broad access early "to see what it can do". Access granted for exploration is rarely withdrawn, and it is the hardest finding to defend later.

A note on certification

Certification demonstrates that a management system exists and is followed. It does not by itself demonstrate that a specific AI deployment is well controlled. Ask suppliers for both: their certification status, stated accurately, and the specific controls that apply to the system you are buying.