The largest US banks just put hard numbers on their AI. The most useful sentence in the quarter was a warning about what those numbers do not mean.
The prevailing story says AI is a margin engine. Deploy it across the workforce, cut the cost base, keep the difference. The second quarter results tell a more precise story. The productivity is real. Whether it turns into durable margin is a separate question. The bank that spends the most on AI answered that question out loud.
The adoption is no longer in doubt
Bank of America reported more than 200,000 employees using AI tools, over 400,000 prompts a day, and more than 300 approved use cases in its second quarter 2026 results (Bank of America, Q2 2026 earnings). JPMorgan reported close to 1,000 AI use cases in flight, against roughly 2 billion dollars of annual AI spend, and said AI had cut headcount by up to 30 to 40 percent in some functions, with most of those staff redeployed (JPMorgan, Q2 2026 earnings call, 14 July 2026). Citigroup said nearly nine in ten employees now use its AI tools (Citigroup, Q2 2026 earnings).
These are not pilots. This is production, at the core of regulated institutions, reported to investors as fact.
Then the largest spender drew the line
Jamie Dimon said the part most AI decks avoid. In competitive banking, everyone will use AI, so dramatic margin improvement will not happen soon. His conclusion was blunt. The ultimate beneficiary of AI will be our customers (JPMorgan, Q2 2026 earnings call, 14 July 2026). He added that expecting sharp, permanent cost reductions is an illusion.
Read that as an operator, not as a headline. The cost comes out. The saving does not stay. Competition passes it through as lower prices and faster service. The productivity gain is real, and it is table stakes.
A productivity gain that every competitor can buy is not an advantage. It is the new cost of staying in the game. Advantage lives in what a rival cannot replicate by purchasing the same tools. Hikari Blue · operator note
What this changes for the people allocating the budget
The failure data already pointed here. MIT's NANDA study found that 95 percent of enterprise generative AI pilots delivered no measurable profit and loss impact (MIT Project NANDA, 2025). The banks sit on the winning side of that divide because they wired AI into real workflows, not slideware. But winning the productivity round is not the same as keeping the gain. The question a board should now force is narrow. What in our AI stack is ours, and what did we rent from a vendor every competitor also rents?
The same call gave the tell. JPMorgan's chief financial officer noted that routine work, such as summarizing an analyst report, does not need the most expensive frontier model (JPMorgan, Q2 2026 earnings call). That is model selection by task, not by brand. It confirms what the market already shows. The model is becoming an input, priced like one. If the model is an input every rival can buy, the durable part is the layer above it. The proprietary workflows. The institutional data. The orchestration and the audit trail that turn a generic model into your operation. That layer is built, not licensed.
In a regulated sector the point sharpens. The layer that makes AI defensible to a supervisor, who did what, on which data, under which control, is the same layer that makes the productivity gain hard to copy. Build it and you hold both. Rent it and you hold neither. You are then paying to run a system whose upside flows to your customers and whose know-how accrues to your vendor.
The question to bring to the next board
Do not ask how much AI is saving. That number is real and it is temporary.
Ask which of those savings a competitor cannot buy tomorrow. That is the only part that becomes margin.
Everyone will have the model. The advantage is in the layer you own around it.
- JPMorgan Chase (2026). Q2 2026 earnings call, 14 July 2026. CEO Jamie Dimon on AI and margins ("everyone will use AI, so dramatic margin improvement won't happen anytime soon"; "the ultimate beneficiary of AI will be our customers"), nearly 1,000 AI use cases, and headcount reductions of up to 30 to 40 percent in some functions. pymnts.com · cnbc.com (bank earnings, 14 July 2026)
- JPMorgan Chase (2026). Q2 2026 earnings call. CFO Jeremy Barnum on task-appropriate model selection, noting routine work such as summarizing an analyst report does not require the most expensive model. pymnts.com
- Bank of America and Citigroup (2026). Q2 2026 earnings. Bank of America: more than 200,000 employees on AI tools, over 400,000 prompts a day, more than 300 approved use cases. Citigroup: nearly nine in ten employees using its AI tools. ciodive.com
- MIT Project NANDA (2025). The GenAI Divide: State of AI in Business 2025. Finding that 95 percent of enterprise generative AI pilots delivered no measurable profit and loss impact, driven by approach and integration rather than model quality or regulation. MIT NANDA report (PDF)
The Hikari Blue team · Austin, July 2026