Starling's 17-second investigation measures speed, not whether it was right

Seventeen seconds

Seventeen seconds is about how long it takes to read this paragraph aloud. Starling Bank says it is also now long enough for a complete financial crime investigation — opened, worked and concluded by software acting on its own, with no person driving it.

That is a remarkable operations achievement and an almost meaningless safety claim. Cycle time is the easiest thing in a bank to measure and the least informative thing to publish. It tells you the unit cost of an investigation has collapsed. It tells you nothing about whether the conclusions are any good.

What a financial crime investigation actually does

Strip out the jargon and the work is this. Monitoring software flags a payment or a pattern as odd. Someone looks at the account, the history, the counterparties and whatever the customer has said. Then they choose from a short menu: do nothing, ask for documents, restrict the account, close it, or report it to the authorities. Every option after the first lands on a real person’s ability to pay rent.

So the only question worth asking about an agentic investigator is not how fast it finishes but what it does at that fork, how often it chooses wrongly, and what happens next to the person it chose wrongly about. None of that is in the number that was published.

The queue was doing a job nobody designed it to do

Banks generate far more alerts than they can work. Investigator hours are the binding constraint, which means the backlog quietly decides how many customers actually get examined. It is a terrible rationing mechanism — arbitrary, slow, and worst for the people waiting — but it is a brake.

Make investigation nearly free and the brake goes. Lowering the alert threshold used to cost real money in analyst time; now it costs compute. Any monitoring team under supervisory pressure will take that trade, and should be expected to. The result is more investigations at the same precision, which arithmetically means more innocent people restricted, not fewer. Speed does not improve the hit rate. It only increases the number of shots.

The incentives are not symmetrical

The penalty for a bank that lets criminal money through is a regulatory enforcement action. The penalty for a bank that freezes a legitimate customer is a complaint, and maybe an ombudsman award months later. Those two are not remotely the same size, and every automated system built inside that gradient will lean the same way.

This is also why the usual reassurance — that a human signs off the agent’s conclusion — deserves scepticism rather than comfort. If the machine produces cases fifty times faster, the sign-off budget does not grow fifty times with it. A human approval given in four seconds to a conclusion generated in seventeen is not meaningfully a human decision; it is a countersignature. Supervisors should treat the ratio of review time to generation time as the real control, not the existence of a reviewer in the org chart.

The customer cannot see the file

In most other automated decisions, the error-correction mechanism is the affected person. You are declined, you see why, you produce the missing document. Financial crime work deliberately breaks that loop: a bank that has reported a suspicion generally cannot tell the customer what it suspected or why, because doing so risks the offence of tipping off. The customer is left arguing against reasoning they are not allowed to read.

That was already uncomfortable when a person wrote the reasoning. It is worse when the reasoning is a model’s chain of inference that even the bank may struggle to reconstruct in plain English eighteen months later for an ombudsman. The opacity now runs in two directions at once.

The strongest case for the machine

Here is the honest counter-argument, and it is a good one. Slow investigation is itself a harm. People have had wages, house deposits and business floats sitting behind a restriction for weeks because a queue did not reach them. Human investigators are inconsistent: the same facts get different outcomes depending on who picked up the file and how many were left in the tray. A system that resolves in seconds, applies the same logic every time, and never gets tired is a plausible improvement on all of that — and the people it would help most are the wrongly flagged.

I accept all of it. That is exactly why the omission is telling. If the 17 seconds is being spent clearing innocent customers as enthusiastically as it is spent catching guilty ones, the median time from restriction to release should have fallen off a cliff, and that is a far better headline than the cycle time. It was not the number offered. Until it is, the only verified beneficiary of the speed is the bank’s cost base.

Supervisors are alive to the general question — regulators and industry figures have spent this week arguing publicly about how far agentic systems should be allowed to act. The argument is still being had at the level of principle. It needs to be had at the level of published numbers.