The Man With the Best AI Numbers in Banking Says They Don't Prove Much
JPMorgan got 200,000 employees onto AI without asking them to — the adoption result every firm is chasing. Then it went looking for the money, and found the harder problem underneath.
In September 2025, Derek Waldron typed a prompt for the CNBC cameras.
You are a technology banker at JPMorgan Chase preparing for a meeting with the CEO and CFO of Nvidia. Prepare a five-page presentation with the latest news, earnings and a peer comparison.
Thirty seconds later, there was a deck. A credible one.
It was a good demo. Every vendor has one.
A year later, Waldron stood up and explained why the demo was not the point.
He is JPMorgan's head of analytics, a former McKinsey partner with a doctorate in computational physics, and he was sitting on the most successful AI rollout in banking. Top bank on Fortune's AIQ 50. Third company overall, ahead of every technology giant but Alphabet. Nobody was better placed to take the victory lap.
Instead he said, publicly, that there was a gap between what the technology could do and what the bank was actually capturing in its results.
That sentence is worth more than the demo.
The adoption story deserves its reputation.
In the summer of 2024 the bank switched on its internal AI platform and issued no mandate. No usage targets. No email telling everyone to log in by Friday. Access went out in phases, only to people who asked for it. Staff were told to treat it as a research analyst — someone you go to for a number, a draft, or an idea — and the bank let it spread by word of mouth.
Eight months later, 200,000 people had signed up. Not one had been told to.
There is a real lesson in that. Refusing to mandate turns usage into an instrument: a tool that works fills up without a campaign, and a tool that doesn't empties out, which is worth knowing. Make adoption a target and people start managing the number instead of reading it.
It also has a ceiling, and JPMorgan's is the honest one. Roughly half the people with access don't use it on a given day. That is at a bank spending close to $20 billion a year on technology, with a platform built in-house and a CEO who compares AI to the printing press. Fifty percent is the top of the market.
So the bank had the number everyone else was chasing. Then it went looking for the money.
Picture a credit memo.
An analyst used to spend four hours writing one. With AI she spends one. Three hours saved, and she will tell you so honestly, because it is true.
Now watch the memo. It joins the queue for the credit committee, which sits on Thursday mornings. It sat in that queue before and it sits there now. The deal closes the same week it always closed. The client notices nothing. The revenue lands on the day it was always going to land.
The three hours were real. They were simply spent on the part of the process that was never the constraint. Thursday was the constraint.
And if every analyst gets three hours faster, something worse happens quietly. More memos reach Thursday's committee than Thursday can read. The queue did not disappear. It grew, in the one place nobody is measuring.
What the firm actually bought was three hours of that analyst's week. That is worth having, but only if somebody has decided what she does with them. If nobody has, it becomes a slightly earlier lunch, and it shows up in no account anywhere.
This is the quiet failure inside most AI programs, and Waldron named it: you can add up the hour saved here and the three hours saved there, get the arithmetic exactly right, and still find nothing at the end of the process. The saving is visible at the desk and invisible in the accounts. Nobody catches it, because the number at the desk arrives first and looks so convincing.
Which raises the harder question. If that arithmetic cannot be trusted, how do you prove anything at all?
JPMorgan's answer came from medicine.
Business numbers have rhythms of their own. Customers transact on Mondays, not Sundays. Paychecks land at month end. August empties out. So when a figure rises the week after you ship something, you cannot tell whether your change caused it or the calendar did. Nearly every AI business case ever written quietly assumes the former.
So they stopped releasing changes to everyone. Half the users get the new thing, chosen at random. The other half carry on as they were. Both halves live through the same Monday, the same payroll cycle, the same August. Whatever is left between them is the thing you actually built.
Run that with a specialist beside every product team and you manage about eight experiments a year. Put it on a platform people can use themselves and you get three hundred.
That is the method. Not a model, not a budget. A control group.
None of it required scale. Not mandating costs nothing. Holding back half your users costs nothing but patience. What it takes is deciding what would count as proof before you begin, rather than assembling it afterwards from a trail of enthusiasm.
Which is where Waldron was standing when he said it. On the best AI numbers in banking, pointing at what they did not show.
A year from now, the question will not be how much AI your firm uses. Usage looks after itself once a tool is any good.
The question is which of your projects are still running because someone proved they work — and which are still running because nobody ever built the test that could have killed them.
Sources: CIO (4 September 2026) · CNBC · JPMorganChase · VentureBeat · Fortune AIQ 50
Comments