JPMorgan Stops Trying to Measure AI's Payback, Even as It Tops AI Rankings
JPMorgan Chase has spent the better part of two years positioning itself as the bank that has figured out generative AI: a $20 billion technology budget, nearly 1,000 internal use cases, an in-house "LLM Suite" reaching tens of thousands of employees, and a first-place finish on the Evident AI Index for banks for a fifth straight year, confirmed again on October 7. So it was notable when, on the company's Q3 earnings call on October 14, CFO Jeremy Barnum did something unusual for a bank that size: he admitted the math no longer holds together, and said he intends to stop trying to force it to.
From '$2 billion in savings' to 'not exactly right'
The retreat centers on a number. Earlier this year, CEO Jamie Dimon told investors AI-driven cost savings were roughly matching what the bank spent on AI, with figures in the range of $2 billion annually floated in that context. Pressed again on the call by analyst Mike Mayo, Barnum said the earlier figure was "not exactly right" and offered a new one closer to $700 million a year — before immediately undercutting even that number. AI spending, he said, is now "pretty well diffused inside the whole institution," making it difficult to cleanly separate from ordinary technology costs. He added that most of what the bank previously quantified was really older machine-learning work, like consumer fraud detection, not the generative and agentic AI that dominates the current conversation.
More striking than the revised figure was Barnum's explanation for why he plans to stop updating it. "I am just very reluctant to spend a lot of energy quantifying it," he said, arguing that the exercise "winds up wasting a lot of people's time inside the institution" who could otherwise be doing the underlying work. He acknowledged AI's benefits were not visible in the quarter's results and said returns take time to show up at an institution of JPMorgan's scale — a materially more cautious framing than Dimon's October 2025 comment that the bank's AI investment had "already paid for itself."
A cost-discipline strategy dressed up as flexibility
What Barnum offered in place of a dollar figure was a description of strategy: deliberate avoidance of dependence on any single AI vendor. "Our strategy is quite focused on minimizing lock-in and maximizing optionality and the ability to manage costs," he said, describing an "all of the above" approach that mixes proprietary frontier models with open-weight and open-source alternatives, some of which the bank runs in its own data centers. Dimon added that tools for automatically routing a given query to the cheapest or fastest adequate model — rather than defaulting to whichever model an employee happens to know — are coming, implicitly conceding that uncontrolled model choice has been inflating the bank's token bill.
Jeremy Barnum, CFO, JPMorgan Chase: "I am just very reluctant to spend a lot of energy quantifying it... it winds up wasting a lot of people's time inside the institution."
Why a bank this far ahead is backing off the scoreboard
The timing compounds the signal. JPMorgan posted adjusted earnings per share of $5.57, below the roughly $5.94 analysts expected, on revenue of about $50.9 billion against a roughly $51.2 billion estimate, with commercial and investment banking results softening after a strong second quarter. Overall expenses rose 12% year over year to $27.3 billion, which Barnum attributed mainly to front-office hiring and wage inflation rather than AI — a reminder that even at a bank running nearly 1,000 AI use cases, headcount growth in client-facing roles hasn't stopped. Dimon, separately, flagged a newer worry: thousands of potential "rogue agents" acting on customers' behalf, and the liability questions that arise when something goes wrong.
The broader lesson for enterprise AI buyers is less about JPMorgan's specific dollar figures, which the bank itself now says were imprecise, and more about what a retreat from measurement implies. JPMorgan has more AI infrastructure, more specialized staff, and more external validation of its AI maturity than almost any other financial institution, yet its own CFO now argues that isolating AI's return at that scale isn't worth the effort. That's a harder case to make for companies with far thinner AI footprints still being asked by their boards for a clean ROI number — and it suggests the industry's leading adopter has concluded that chasing a precise payback figure is less useful than simply controlling what AI costs.