The question about AI in finance processes is not what a model can do but what it may do. So AI here does not have a module but an authority axis — four levels, and the highest is a decision somebody makes and can withdraw.
Our AI product design is a target architecture. Nothing on this page describes an autonomous agent acting on its own in finance processes today.
The approval check is server-side and bound to roles. A recommendation is not an authority.
With too little history the forecast proposes no pattern and adoption stays disabled. An empty statement is better than an invented one.
Where the basis is not sufficient, the system says so — and proposes nothing. A recommendation without a basis is worse than none.
A person can reject any recommendation, and the rejection stays on the record. Overriding does not remove you from the evidence.
A system does not widen its own authority. A higher level is granted by an authority decision, not by the model.
AI is none of the 13 modules. It runs across all of them — together with cases, controls and evidence.
Where it worksWhat an AI level does in one module, it does the same way in every other — because the authority decision sits on the action class, not on the module.
No. Level 04 requires an explicit authority decision, per action class and per entity, and it is revocable. Without that decision the chain ends at Prepare.
Like every other decision: with person or action class, basis and time on the transaction. An AI action is not a special category in the audit trail.
The provider decision for document understanding is open and is carried as open. What we can say today is set out under what we can evidence.
Bring a transaction where somebody assembles numbers today before a decision is made. We work through which level holds there and where the authority would have to sit.