Underwriting the Unknown.

Finance is the industry that industrialised being right. Credit scores, actuarial tables, value-at-risk: the entire edifice rests on the proposition that the past, properly weighted, prices the future. Machine learning is the strongest version of that proposition ever built, which is why banking adopted it faster and deeper than any headline suggests — and why its characteristic failure mode deserves a brief of its own.

Start with what has already happened, quietly. Credit decisions that took committees a week now take ninety seconds. Settlement stacks reconcile in minutes what took two days. Surveillance systems read every trader chat; fraud models arbitrate millions of transactions an hour against a learned portrait of forty million customers' normal. None of this had a launch video. The fintech decade, as Rohan Iyer puts it on the show, was a decoy: the apps won the interface while the incumbents rebuilt the engine room.

From this conversation

AI in Banking: The Quiet Disruption

S07 · E03 · 52m

▶ Watch the episode

The gains are real and the brief could end there, except for a pattern every risk officer I interview describes in nearly identical words. A model works. It works next year, and the year after. The people who remember why it might break get promoted, retire or leave. Documentation decays into ritual. And at some point the institution is no longer running a model it understands — it is worshipping one it inherited. The danger was never the model. It is the institutional forgetting that grows around a model that works.

“The danger was never the model. It is the institutional forgetting that grows around a model that works.”

The unknown in the title is not exotic. It is the regime change: the demand pattern, default cascade or volatility surface that appears in no training set because it has never happened. Human judgement handles these badly; models handle them worse, and with perfect confidence. The 2008 crisis was, among other things, a model-risk event in an industry that then wrote thousands of pages of regulation about capital and almost none about memory. The next one, the practitioners quietly agree, will be faster.

What does defensible practice look like? The leaders converge on a few disciplines. Synthetic crises run quarterly, models off, phones only — simulator hours for finance, expensive and non-negotiable. Dissent logs that make a model argue its case on a permanent record. Friction budgets that spend customer patience where measured risk actually lives rather than where compliance optics point. And a hiring rule borrowed from aviation: the people supervising the automation must be capable of flying the plane, which means keeping humans match-fit for work the machines normally do.

The regulatory gap is closing slower than the capability gap. Supervisors have engineers now, but examination frameworks still audit inputs and outputs while the risk has moved to the relationship between institutions and their own systems. A modest proposal circulating among the smarter central bankers: mandate memory. Require every systemically important model to carry, like a ship's log, a living record of why it might fail, maintained by named humans, examined annually. It would be resented, evaded and, eventually, credited with preventing something we will never get to see.

The brief's conclusion is uncomfortable for both camps. The luddite position — slow down — is not available; the efficiency gains are too large and too compounding, and the institution that declines them funds its rival's buyback. The evangelist position — trust the backtest — is how the last crisis was manufactured. What remains is the hard middle: adopt at full speed, and pay, in cash and senior attention, for the organisational memory that makes speed survivable. Underwriting the unknown was always the job. The unknown has simply changed employers.