Essay
AI is often treated as a separate governance category. That is the wrong starting point. It does not sit neatly beside security, privacy, or operations. It modifies how existing systems behave - and that is why so many organisations are adding AI controls without actually improving governance.
We keep trying to put AI somewhere.
A new category. A new domain. A new layer to govern.
That instinct is understandable. AI feels distinct, so organisations isolate it. They create separate governance streams, separate policies, separate risk registers. It sits beside security, beside privacy, beside operations.
Contained. At least on paper.
But this is the wrong mental model.
Common framing
Looks tidy. Misses the real change.
More accurate framing
Harder to govern. Closer to reality.
AI does not sit cleanly in its own domain. It moves across the ones that already exist.
It does not stay where you put it.
Most organisations are not changing how they govern systems. They are adding AI governance alongside existing governance.
That sounds prudent. In practice, it leaves the underlying problem untouched.
The structure of the system is still there: data, interfaces, decisions, execution. What changes is how those parts behave.
Consider a claims processing system. The data pipeline, decision logic, interface, and execution chain were already there. AI does not add a new layer. It changes how the existing logic resolves, makes the reasoning harder to trace, and disperses accountability across the model, the data, the integration, and the business owner who signed it off.
The system looks the same. It does not behave the same.
Model: the issue is not that AI creates a new trust domain. It changes the behaviour of existing system layers, which is why governance often lags reality.
Three shifts matter most.
The same input no longer guarantees the same output. Testing shifts from verification to probability. You can still build confidence - but not on the assumption that the system will behave identically next time.
The system still produces outcomes. You just cannot always explain them. Why did it resolve that way? What influenced the output? Would it do it again under the same conditions? These questions become harder to answer with any precision.
When something goes wrong, no single owner is obvious. Was it the model? The training data? The prompt design? The integration? The business owner who approved deployment? Usually all of them, which means - in practice - none of them.
There is an obvious objection: AI does introduce genuinely novel phenomena - hallucination, emergent behaviour, training data contamination. These are real. But the governance challenge they create is not about categorisation. It is about changed behaviour in systems that already had structure, with owners who now cannot account for what shifted or where.
The question is not: "Do we have AI in this system?"
The better questions are:
What has changed in how this system behaves?
Where has confidence become weaker?
Can that behaviour still be interpreted, tested, and governed with enough clarity?
Most organisations have responded by creating AI controls. But controls are not the same as governance.
So you end up with something unstable: structured controls, unstructured behaviour.
| What changed | What weakened | Why current governance misses it |
|---|---|---|
| Repeatability | Confidence in stable outputs | Testing models assume deterministic behaviour |
| Interpretability | Clarity of cause and influence | Oversight relies on being able to trace decisions |
| Ownership boundaries | Clear accountability | Responsibility frameworks assume identifiable owners |
| System behaviour | Component-based governance assumptions | Controls are designed for parts, not emergent behaviour |
AI alters behaviour across existing systems. Governance stays organised around functions, tools, and ownership lines that assume a stability that no longer exists.
AI does not create the governance problem. It reveals one that was already there.
Most organisations do not actually govern behaviour. They govern components.
Tools. Policies. Platforms. Controls.
Not the live behaviour of the system as it operates under real conditions.
This was always the gap. Systems behave differently under load, under edge cases, under the messy reality that component-level controls were never designed to see. AI widens that gap and makes it impossible to ignore.
AI is not primarily a new category to manage. It is a stress test for whether the organisation has ever understood how its systems actually behave.
AI does not need its own domain of trust. It does not stay contained long enough.
What it does instead is more consequential. It moves across existing systems and changes their behaviour - making outcomes less stable, less legible, and harder to attribute.
Trust does not break because something entirely new was added.
It breaks because something familiar started behaving differently, and no one had built the governance to see it.
The organisations that get this right will not have the most comprehensive AI policy. They will be the ones that already knew how to govern behaviour - not just components - and adapted when the system changed under them.
Related: TrustSurface Framework · What Owning the Status Surface Looks Like
