AI governance has to work beyond the PDF

Essay

The question usually surfaces about twenty minutes in. The platform is selected. The budget is committed. The roadmap is built. Someone asks: when the model is wrong and it affects a customer, who owns that?

The room goes quiet. Not because nobody has an answer, but because the answer changes depending on who is asked.

That pause is where many AI programmes expose their real weakness: not model performance, but leadership accountability.

The pattern behind the pilot

When AI does not move from pilot to routine work, familiar explanations appear: unclear governance, weak culture, misaligned incentives or poor change management. Each can be true. But one question often sits underneath them: who decides?

When a model recommends one action and a person disagrees, whose judgement prevails? Who must review the choice? Who carries responsibility for the consequence? When a regulator, customer or colleague asks how a decision was made, can the organisation answer honestly?

A principles document rarely settles those questions at the level of daily work. People fill the gap themselves. They create shadow spreadsheets, back-channel approvals and unofficial review steps. Those workarounds are not simply resistance. They are evidence that the formal operating model has left responsibility unresolved.

Values need a decision mechanism

Transparency, accountability and human oversight matter. Yet each remains abstract until it changes who is allowed to act, who must approve and what happens when the normal rule no longer fits.

The Decision-Rights Map makes that conversation concrete. It uses two questions: how consequential is the outcome if the decision is wrong, and how reversible is that outcome? Those questions help a team place work into four operating modes:

• Human decides; AI informs. Consequences are significant and difficult to reverse, so human authority remains explicit.
• Human and AI collaborate. The system proposes or scores; a person validates, adjusts and owns the decision.
• AI executes; humans oversee. The task sits inside defined boundaries, with monitoring and an escalation route.
• AI acts autonomously. The work is low consequence and readily reversible, with the boundary still recorded and reviewed.

The map is not a universal scoring model. Context determines consequence. The point is to make authority discussable before a system starts making or shaping decisions.

Four leadership disciplines

Decision rights only hold if the surrounding leadership practice supports them.

1. State what AI is for

Leaders need a common position on the purpose of AI in the organisation, where it may shape work and where it will not replace human judgement. Without that position, every team quietly invents its own.

2. Assign authority in the workflow

Name the person who decides, the evidence they must consider and the route for exceptions. “Human in the loop” is not enough. The loop must have a role, a moment and an accountable owner.

3. Turn governance into a routine

Review model performance, inspect outcomes and let people closest to the work flag when a system is wrong. A governance practice changes with the system and the work; a document does not.

4. Make trust observable

Trust is built by what happens after the model is wrong: who is told, who reviews the incident, what changes and whether the people affected can see that learning took place. The organisation’s response teaches people whether the stated principles are real.

The test before the next pilot

Choose one live workflow and ask the people involved to explain who decides when AI is part of it. Ask what gets logged, when a decision is reviewed and who owns an exception.

If the answers conflict, the gap is not another paragraph in the governance document. It is a leadership decision that has not yet been made.

The documentation is not the governance. Governance is the pattern of authority people can rely on when the rule meets reality.