Bank AI Spending in Q3 2026: What JPMorgan, BofA and Citi Actually Quantified

On its third quarter call on 13 October 2026, JPMorgan's finance chief said the bank's annual AI expense was "something like $700 million". That is roughly a third of the $2 billion figure the bank's own chief executive had put into circulation a year earlier. Bank AI spending got more talk than ever this quarter, and fewer hard numbers than you would expect.

Bank AI spending in Q3 2026: the short answer

What did the big US banks disclose about AI spending in Q3 2026? Very little that was audited. JPMorgan estimated its AI expense at about $700 million a year and declined to quantify benefits. Bank of America put AI benefits at $250 million to $300 million this year, rising to $700 million. Citi gave operational metrics but no dollar figure for either side.

That is the whole finding, and it is more useful than it sounds. Three of the largest technology spenders in any industry discussed AI at length on the same day. Each chose a different unit to measure it in. The unit a bank chooses tells you what it is confident about, and what it is not.

My view is that the choice of unit is the real disclosure here. A bank that gives you dollars of benefit is inviting you to check it next year. A bank that gives you engineer counts and percentages is asking you to accept a direction of travel. A bank that refuses to quantify is telling you the cost is too diffuse to isolate, which is honest and also unhelpful.

The per-bank ledger of tech spend and AI claims

The table below pulls together what each bank said on its Q3 2026 call, plus the one tech budget figure disclosed earlier in the year. Every figure is the bank's own, spoken by a named executive. Where an executive hedged, the hedge is kept.

Revenue and expense figures are as stated by each CFO on the call. Bank of America's reported expense includes a $1.2 billion regulatory settlement charge. Citi's prior-year quarter included a $726 million goodwill impairment on Banamex. JPMorgan's $19.8 billion is from its February 2026 company update, as reported by Business Insider.

JPMorgan: a $19.8 billion tech budget and a shrinking AI number

JPMorgan is the largest technology spender of the three by a wide margin. At its company update in February, CFO Jeremy Barnum said the bank planned a technology budget of $19.8 billion for 2026, roughly 10% above 2025, according to Business Insider's report of the update. He called technology "a major driver of our expense growth".

The Q3 numbers show that growth. Expenses were $27.3 billion, up 12% on a year earlier, against revenue of $50.9 billion, up 8%. Barnum attributed the expense rise mainly to volume and revenue-related costs, front office hiring and wage inflation. He guided the fourth quarter to about $27.5 billion.

He also said consensus for 2027 expenses sat around $114 billion and that the bank's preliminary read "isn't too far from that". He named "the effects of AI" as one reason the investment picture was dynamic. That is the closest the call came to tying AI to the expense outlook.

The $2 billion figure that became $700 million

The most useful moment on any of the three calls came from an analyst question. Mike Mayo of Wells Fargo asked whether JPMorgan spent "$2 billion in AI" and got "$2 billion of benefits". The figure had a public history. In October 2025, Bloomberg reported that Dimon said the bank spends $2 billion a year on AI and saves about the same, citing a Bloomberg TV interview.

Barnum corrected it. "I think it's not exactly right, and I think we corrected it later," he said. "The AI expense number was significantly less than that. I want to say something like $700 million annually." He added that the cost was "pretty well diffused inside the whole institution".

Read that carefully. The bank's own CFO is hedging his own number, describing it as hard to isolate, and calling the older quantification accurate "as far as it went". I think this is the most honest AI disclosure of the quarter, and it is also an admission that nobody outside the bank can audit the figure. It is the same problem covered in the gap between AI spending and measurable EBIT impact, here stated by a CFO in his own words.

Why the CFO stopped quantifying, and what he worries about instead

Barnum said he was "really trying to move away from" benefit quantification, and that "the benefits aren't observable in our current quarterly earnings". He described experiments rewriting minor legacy tools in a single pass, and said results in his own finance technology group were "increasingly exciting".

Then he named the cost side nobody else raised. "Some of those benefits are going to be offset with token costs, which is a whole other conversation," he said, adding that the bank's strategy focuses on "minimizing lock-in and maximizing optionality". Dimon followed with a point about model routing tools that send work to "the best, quickest, cheapest" model. Token costs eating into efficiency gains is the pattern laid out in why cheaper tokens still produce rising AI bills.

Dimon also drew a perimeter. "We're not going to allow just any outside agent to come in inside the company," he said, and warned of "rogue agents, and there could be thousands of them". For a vendor selling agents into a bank, that sentence matters more than any budget figure on the call. The control problem behind it is the subject of the piece on agent identity and non-human access.

Bank of America: the only bank that put dollars on AI benefits

Bank of America took the opposite approach to JPMorgan. Answering the same analyst, Moynihan offered a count and a dollar trajectory. "We are about 130 different projects implemented," he said. "We expect that number to go from the $250, $300 this year to $700 million next year in terms of projects we put in the pipeline."

He placed the AI work inside a larger figure. A key area, he said, "is in the development of technology, which is $4 billion and change a year". He also gave a ratio: "Overall, the benefits are 2x-3x over time of the expense you are giving out". No AI cost figure was offered to anchor that ratio.

I would treat the $700 million as the single most checkable AI claim made by any of the three banks this quarter. It is a dollar figure, it has a year attached, and it can be compared against the 2027 results. That makes it more valuable than any number of adjectives, even though it is a target and not a result. Counting projects is a weaker signal, for the reasons set out in the case against AI use case counts as a metric.

Two headcount numbers on one call

CFO Alastair Borthwick said in his prepared remarks that "our headcount was down nearly 3,500 from a year ago as we automate processes and simplify work across the company". Later in the Q&A, Moynihan described headcount as "down 3,000 round numbers" and "down by 3,000, 1.5% year-over-year".

The gap is small and probably a matter of rounding and timing. It is still worth noting, because headcount is the metric most often used as proof of AI efficiency, and the two senior executives did not give the same number. Moynihan also said the bank had added 1,000 extra staff for a remediation programme the transcript calls "Mythos", "not in the headcount but in consultants". Headcount that falls while consultants are added is a lower-quality efficiency signal than it first appears.

Revenue grew slower than expense

The quarter itself did not show AI efficiency in the totals. Revenue was about $30 billion, up around 3%. Reported expense was $19.7 billion, including a $1.2 billion regulatory settlement charge. Excluding that, adjusted expense was "a little more than $18.5 billion", up slightly less than 7% from a year earlier.

Borthwick said, "We expect to return to operating leverage in Q4." That phrasing concedes the bank did not have it in Q3. Net income was $6.9 billion, or $8.1 billion excluding the charge. Digital engagement was strong: over 50 million active digital users, and digital channels generated 71% of sales.

Citi: operational AI metrics without a dollar figure

Citi's Jane Fraser gave the richest set of AI operating metrics of the three, all in response to a Mike Mayo question. The pattern is consistent: specific percentages and counts, without a dollar base for any of them.

She said the bank was "deploying AI at scale with measurable results". She cited product launches running about 30% to 40% faster, client response times improved by 21%, and "around 100 of our most manual processes" being re-engineered end to end. She said AI service agents for clients were live in 82 countries.

14,000 engineers and a 30% vendor cut

The most concrete claim was about engineering. "14,000 of our engineers use our main agentic coding tool," Fraser said, naming it in a word both automated transcripts render as "Devon". She tied it to insourcing. "We've reduced our tech vendor spend by 30 percent year over year as a consequence. And that's still got a long way to run."

A 30% cut in vendor spend is a real number with a real mechanism. Without the base, though, you cannot size it. For a firm selling software or services to Citi, this is the sentence to read twice. The bank is saying that agentic coding lets it build in-house what it used to buy, and it is measuring the result in reduced vendor spend.

"Outweigh" is not a number

Fraser's summary claim was that "this year alone, we're seeing the financial benefits from AI and automation significantly outweigh the investments we've made year to date". That is a strong statement. It also has no figure on either side of the comparison, so it cannot be tested from outside.

The CFO, Gonzalo Lucchetti, referred to "the 100 plus end-to-end processes" and to "incremental investments over three years of about $5 billion" in the same answer. The transcript does not make clear how much of that $5 billion is AI. Citi's clearest hard efficiency figure is elsewhere: "over $1 billion of severance" recognised year to date, "largely across functions, technology and operations."

The quarter's totals were solid. Revenue was $23.8 billion, up 8%. Expenses were $14.0 billion, down 2% on a reported basis, with an efficiency ratio of 58.5%. Excluding the $726 million Banamex goodwill impairment in the prior-year quarter, expenses rose 3%. Citi guided to a full-year efficiency ratio below 60%.

Bank AI spending against expense and revenue growth

If AI were already moving bank cost structures, the place to see it would be the gap between revenue growth and expense growth. On Q3 2026 numbers, only Citi grew revenue faster than expense, and Citi had the cleanest restructuring story behind it. That story predates the current AI wave.

None of this proves AI is failing to save money. It shows that, at the scale of a bank's total cost base, AI benefits in the hundreds of millions are too small to see in quarterly totals. Barnum said exactly that. BofA's $700 million target is about 1% of a quarterly adjusted expense of $18.5 billion multiplied by four. It matters, and it is not yet a structural change.

For context on how banks compare with other regulated sectors on adoption, see the comparison of AI adoption in banking and healthcare.

Where this ledger is weakest

This post makes claims about what three banks said. The evidence for that has real gaps, and I would rather state them here than have a reader find them.

The transcripts are not the official record

The quotes come from third-party transcripts, two of which are automated. They render the same Citi tool as "Devon" and spell the Citi CFO's name two ways. The official Citi transcript PDF and the JPMorgan and Bank of America earnings releases were located but could not be text-extracted for this post. Every figure here should be checked against the replay or the official transcript before you rely on it in a board paper.

One quarter is not a trend

The brief for this post asked for a quarter-on-quarter comparison. The honest version is narrower. JPMorgan's AI figure can be compared with October 2025, and its budget with February 2026. I did not verify Q2 2026 AI statements for Bank of America or Citi, so their columns are a single snapshot. The ledger should be rebuilt after Q4 results in January.

Diffused cost is a genuine accounting problem

The strongest counter-argument to this whole post is Barnum's. When AI is embedded in every system, ring-fencing its cost is partly arbitrary, and a precise figure would be false precision. On that view, JPMorgan's refusal is the most rigorous position, and BofA's $700 million is the least. I find that argument persuasive on cost. It is weaker on benefits, because a bank that cannot size the benefit also cannot defend the budget. The wider problem of which AI metrics hold up is covered in the ranking of AI revenue disclosure metrics.

How to read the next bank AI spending disclosure

The Q4 calls in January will carry 2027 guidance. Bank of America said it would likely give next year's guidance then. These are the questions that separate a quantified claim from a narrative one.

If you sell technology to banks, the vendor line is the one to watch. Citi has told you it is cutting vendor spend with agentic coding. JPMorgan has told you outside agents will not get in without controls. Those two sentences describe your pipeline better than any AI budget figure.

Frequently asked questions

Where to start if you sell into banks

Read the Q4 2026 calls in January with one test. Write down each AI number each bank gave this quarter, from the table above, and check which ones reappear with an actual result attached. A figure that quietly disappears is information.

Then look at your own bank accounts. If a customer is cutting vendor spend through insourcing, ask your contact this week which of your services sits on that list. The answer will arrive before any bank's AI budget does.

Related reading