NIFFLER | Autonomous Revenue Recovery

Policy Guardrails

The line between what the AI wants to do and what it is allowed to do.

Why this exists

The AI is good at reading a situation and suggesting what to do. It is not something you want quietly moving money on its own.

So it does not get to decide. Whatever it suggests is checked against four plain rules first. These rules are ordinary code, not AI. They give the same answer every time, and the AI cannot talk them round.

The three verdicts

ALLOWED

Go ahead. The agent does what it planned.

DENIED

Blocked. The case stops here, and there is nothing for a person to do about it.

REQUIRES HUMAN APPROVAL

Blocked, but a person should take a look at it.

The four rules

Eligibility

Is this order still unpaid?

If the money already came in, there is nothing to recover. This holds even if the customer paid some other way that NIFFLER had nothing to do with.

DENIEDOrder already paid

Attempt limit

Has this already failed too many times?

Banks keep score of how often a shop's payments get declined. Keep retrying a card that has already been refused and the bank starts trusting that shop less, which makes future payments fail more often.

REQUIRES HUMAN APPROVAL3 or more failed attempts

Prior recovery link

Has a payment link already gone out?

If the customer pays both links, they get charged twice.

DENIEDLink already paid

REQUIRES HUMAN APPROVALLink outstanding, unpaid

Iteration limit

Is the agent going round in circles?

An AI can get stuck repeating itself forever. This cuts it off and hands the case to a person.

REQUIRES HUMAN APPROVAL10 or more steps on one case

How the answer is decided

All four rules run on every case, and the strictest one wins. If any rule says block, the case is blocked, no matter what the other three say. The agent only gets to do what it planned if every rule agrees.

You can see this happen on real cases in the Decision Explorer. Filter it to the cases where the rules overruled the AI.