Notes

Pending Outlines

Outlines for pieces that have to come from the work, not generated. High level ideas, not yet ready.

1. Who pays for fraud

   From the work: allocation, total cost of fraud, and what charging a P&L actually rewards. Write from rooms and numbers, not the general case.

   • Every fraud loss is paid by someone. The live question is which P&L is made to look responsible.
   • Loss is only the visible part of the bill: operating expense, false positives, late write-offs, disputes, brand.
   • An allocation model is an incentive machine. Who is overcharged starts caring about prevention; who is undercharged starts caring about customer experience.
   • ROI in fraud is a story with assumptions. A handsome ROI with miserable customer impact is a transfer, not a success.
   • If you cannot say, in language a product owner will accept, why their customers are in the queue, you have a dashboard rather than a strategy.

2. The control that performs

   The difference between a control that changes what someone may do, and the artefacts that prove such a control exists.

   • Two systems: the one that changes behaviour, and the one that produces evidence. Careers are often made in the second.
   • Test: if this control disappeared on Monday, what would break on Tuesday? If the answer is a missing slide, it is a ritual.
   • A real control has a cost someone does not want to pay. A long inventory that makes nobody angry is a warning.
   • Governance, used precisely, is plumbing: who may do what, how we know, what happens when they did not.
   • Generative models make the artefacts cheaper. They do not watch a queue or own a miss.
   • Prefer a smaller inventory: named owners, a failure mode in a sentence, a metric that would move if the control stopped working.

3. What a model does not know

   Where models help in this work, and where they cannot replace the person whose name is on the exception.

   • Banks do not lack text. They lack people who are still in the room when the text is wrong.
   • Fluency in the house dialect is not knowledge. Tired people waving through a pack that sounds right are the actual control environment.
   • What the model does not know is local: residuals in the loss line, vendor feeds patched with a default, the report a senior person last trusted in 2019.
   • Use: draft, summarise, point at fields. Do not use as system of record, policy, or attestation.
   • A number without a lineage a sceptical person can walk is a paragraph, not analysis.
   • Stay exception-native. Who owns the miss?