Boundaries
Authority is not a feeling. It is a boundary.
In most AI tools, the limits are guidance the model is asked to respect. In MAIA they are a boundary it cannot cross — enforced in the system of record, below the application, where no prompt and no confidence can reach. MAIA can prepare anything. It can only commit what it was granted.
The line the AI cannot argue with
It cannot make a number true by sounding sure.
A confident wrong answer is the specific thing finance cannot afford. So confidence buys MAIA nothing at the boundary: the record keeps the last word, and there are things the AI simply has no power to do — no matter how it is prompted, and no matter how certain it seems.
It cannot exceed its grant
Post above the value it was trusted with, act outside the part of the business it was given, use a permission it never had — the entry is refused, not flagged after. (You saw this on the previous page: the €41,800 it would not post.)
It cannot rewrite history
Once a record is posted, its figures are frozen. A correction is a new, traceable entry — never a quiet edit to make the books look tidy.
It cannot hold an impossible state
Books that don’t balance, stock below zero, a period reopened after close — the ledger will not hold them, whoever or whatever asks.
Why below the application matters
Rules in the software can be talked around. Rules in the record cannot.
Put the guardrails in the application and a bug, an integration, or a persuasive instruction can slip past them. Put them below it, in the system of record itself, and there is no path by which a wrong state gets in — not through the screen, not through an import, not through an AI that is very sure of itself. That is the difference between trusting a model and trusting your accounts.