What happens when the AI is not sure about a classification?
It says so. Maditon reports a confidence score and an explicit list of claims the verification pass could not support in the regulation text. Below the confidence threshold, or where a citation cannot be grounded, the classification is marked as requiring expert review rather than being presented as a settled answer.
Uncertainty is surfaced rather than smoothed over, because a confident wrong answer is the expensive failure mode in compliance work.
Concretely, a classification carries:
- a confidence score from the verification pass;
- a list of unsupported claims — statements in the classification that the verifier could not tie back to the retrieved legal text;
- a requires-expert-review flag, set when confidence is below threshold or when citations could not be grounded.
On consequential or low-confidence outputs a second model is consulted and its agreement or disagreement is recorded alongside the result.
This is what makes the output usable in an audit. An auditor is not impressed by a system that always answers; they are reassured by one that distinguishes between what it can evidence and what it cannot. The flagged cases are also exactly where paid legal advice is worth buying, which makes the flag commercially useful to you and not just epistemically honest.
Regulation references
- Regulation (EU) 2024/1689 (EU AI Act) · CELEX 32024R1689
Read next
A free Solo account runs a real classification on one of your own systems — no card, no sales call.
Run a free classification