Season 06 · Episode 08 · Banking · 45 min
AI in Banking: The Fraud Arms Race.
Crescent Capital's Chief Risk Officer spends her days fighting machine-generated fraud with machine-built defences. She explains why the attackers iterate faster, and where the defenders still hold the high ground.
October 30, 2024·With Farida Khan
▶45:47§01Chapters
Introduction: The Other Side Ships Faster
Anatomy of a Synthetic Identity
Voice Clones and the Death of the Callback
Fighting Models with Models
The Friction Budget
What Customers Should Actually Do
§02Show Notes
Fraud is the one corner of banking where the adversary has no compliance department, and Farida Khan is unusually honest about what that asymmetry means. Criminal crews A/B test their scripts, share model weights, and ship daily. Her team ships quarterly, after committee. This episode is a field report from the losing tempo of that race, and a surprisingly optimistic argument about why tempo is not the whole game.
The strongest material covers synthetic identities: customers who do not exist, assembled from real credit histories and generated faces, aged patiently inside the system for years before they borrow and vanish. Farida walks through how one ring aged four hundred synthetic customers for three years, and how the tell that unravelled them was not biometric but behavioural, a pattern no human investigator would have thought to look for.
The closing section on the friction budget deserves to be taught in business schools: every security measure spends customer patience, patience is finite, and the art is allocating it where the models say risk actually lives. Her practical advice for listeners, on voice cloning and family code words, is the most forwarded clip of the season.
“Real customers are boring in ways no generator has learned. What the fraudsters couldn't fake was boredom.”
§03Transcript Extract
NISHANT:
You said something at a conference last year that got you in trouble: the fraudsters have better MLOps than most banks. Do you stand by it?
FARIDA:
I got in trouble because it's true. A fraud crew tests a new script on a Monday, reads the results on Tuesday, and ships the improved version Wednesday. My equivalent change goes through model risk review, compliance sign-off and a release window. They are running a startup. I am running a utility. The honest starting point for this whole conversation is that asymmetry.
NISHANT:
So how do you win a race you're structurally slower in?
FARIDA:
You stop racing on their track. I cannot out-iterate them, but I can see what no crew can: forty million customers' worth of normal. Fraud has to look like someone, and looking like someone is much harder than sounding like them. The synthetic ring we broke last year had perfect documents and perfect faces. What they couldn't fake was boredom. Real customers are boring in ways no generator has learned.
NISHANT:
Explain the friction budget. I think it's the most useful idea you've given me.
FARIDA:
Every check I add costs the customer patience, and patience is a budget I spend, not a resource I own. Spend it on a pensioner moving her savings after a phone call from her 'grandson', absolutely. Spend it on someone buying coffee, and I have taught forty million people to ignore my warnings. The models exist to tell me where the budget buys the most safety. That allocation problem is the actual job now.
§ Keep listening
More on Banking.
S07 · E03 · 52 min
The Quiet Disruption
Why the next generation of banks will be code-first, and what legacy institutions are doing to survive.
S07 · E06 · 46 min
Powering the Algorithmic Grid
AI is balancing the world's most complex machine — the electricity grid — in real time. What happens when it fails?
S07 · E05 · 55 min
The Death of the Script?
Generative tools are rewriting Hollywood's economics. We talk to the people on both sides of the picket line.