A DPO puts a dashboard in front of the auditor. Three charts: detections per month, distribution across applications, and a figure sitting at 97.
The auditor looks at it for ten seconds and asks one question: what is this evidence of?
That is where the conversation usually stalls, and the numbers are not the problem.
A figure is not evidence
Evidence is a relationship between three things: an obligation, a control, and an observation showing the control worked.
A dashboard typically supplies only the third. There were 312 detections and 97 per cent of them were handled. That is an observation, and it is not yet attached to anything.
The missing step: which article of which framework does this support, and above what line is it enough? Until somebody writes those two down, the reader has to guess. Auditors do not guess.
Three things that are usually absent
The control. A handling percentage might support data minimisation under Article 5(1)(c), or appropriate security under Article 32, or neither. It depends how you set it up. Write it next to the figure, with the article number, or the figure does no work.
The scope. The big one. If detection runs in the supported applications, then 97 per cent describes those applications. Not every AI service your organisation uses. Without that sentence the percentage is not wrong but it is misleading, which is worse, because it does not survive the first question.
The red rows. Every organisation has controls that measurement cannot substantiate. Someone using an AI tool on their phone falls outside the picture. Whether a processing agreement was signed is not something your system knows. Leave those out and the view is incomplete, and that is exactly what gets probed.
Two evidence chains you must not add together
The distinction most reporting misses, and the one that pays off most.
There is a visibility layer: which AI services are in use, how many carry a decision, what risk band applies. It sees a great deal, across hundreds of services, and can intervene nowhere.
And there is an intervention layer: sensitive data getting a highlight before it is sent. That layer can intervene, and only inside the applications it supports.
The temptation is to produce one number, because it reads better. Resist it. The two layers have different scope, and a blended figure hides precisely the distinction a regulator asks about. What you want is an accountability view stating, per control, which layer the evidence comes from. That both layers can be measured without attaching people to them is covered in AI monitoring versus employee monitoring.
Written out, it looks like this:
