AI Governance
An AI agent runs the same task ten times.
Nine are fine, the tenth is wrong.
So you investigate.
Same prompt, same system, same task.
But this time the agent took a different path.
It retrieved different context, called tools in a different order and produced a different result.
Now the real questions begin.
How do you reproduce the issue?
How do you prove what happened?
How do you know the fix worked?
And how do you stop the same behaviour reappearing somewhere else?
This is where AI governance currently falls short and has to move beyond policies and approval boards.
We need operational controls around:
Prompts and configuration
Model/Version/Quantization
Retrieved context
Tool calls and permissions
Outputs and actions
Human intervention
Evaluation against expected outcomes
In other words, every important agent run needs an evidence trail.
Then we need repeatable evals to test whether the system is becoming more reliable over time, not just whether it behaved correctly once and passed the test.
AI governance is not just about what an agent is allowed to do.
It is about being able to understand what it did and testing behaviour to improve it safely.
That is where governance, observability and engineering start to become the same discipline.