What happens if the AI gets something wrong?

The first is the confidence threshold: uncertain output never goes out on its own. The second is validation against your systems, so an invoice total that does not match the order is stopped regardless of how confident the model was. The third is the human approval step for anything irreversible.

The question is not whether the AI is perfect, it is whether it is better than the current process on a Friday afternoon. A tired colleague handling the fortieth email of the day also makes mistakes, and those mistakes are not logged, not measured and not corrected systematically.

What we do promise is measurement: the error rate is visible on a weekly report, so you can see whether it is falling and where the remaining errors cluster.

Legally, you are, the same as when a colleague makes an error. That is exactly why the design keeps a person in the loop for anything with consequences, and why the automation records what it did and why, so responsibility can be traced rather than argued about.