The disclosure and our analysis
What the review found
Colocation Scout read third-quarter disclosures from the 100 largest publicly reporting U.S. banking groups. Thirty-six described what they use AI for. For the largest groups we also went back through earlier 2026 presentations and announcements, which are dated where we use them.
The uses cluster in a few places: software engineering, tools for advisors and bankers, assistants for employees, customer service and lending. Much of it is aimed at the bank's own staff. Most of the models come from outside suppliers. Few banks put a dollar figure on the return.
The thinnest part of the record is where the computing happens. A handful of banks have described their own hardware, their own data centers or rented cloud capacity. None has disclosed a GPU rental contract or a colocation arrangement for AI.
Writing code
Software engineering produced the clearest adoption numbers. Citi CEO Jane Fraser said 14,000 of the bank's engineers use Devin, an AI coding tool. Citizens said about 95% of its software engineers use AI coding tools. M&T Bank said 100% of its developers use AI. SouthState said 80% of its monthly code is AI-assisted. BNY said in April that AI authored more than 40% of its code in the first quarter of 2026. [Source 1] [Source 2] [Source 3] [Source 4] [Source 24]
Pinnacle Financial gave the most specific result. CEO Kevin Blair said AI helped draft, automate and update tests for a systems conversion, with a person approving each step. For the 838 test scripts completed at the time of the call, Pinnacle estimated that AI cut the required time by roughly 70%. [Source 5]
Preparing advisors and bankers
Wealth firms described similar tools. Charles Schwab said it collaborated with Anthropic and is the first custodian for registered investment advisors included in Claude for Financial Advisors. Raymond James CEO Paul Shoukry said its Meeting Manager tool helps advisors prepare for client meetings, summarizes them and sends follow-up items into their client records. He said it saves advisors hours a week. Ameriprise said it began an early release of an AI meeting prep agent. It said nearly three quarters of its advisor practices use its AI-generated insights. [Source 6] [Source 7] [Source 8]
Citi announced Citi Sky in April. It called it an always-on, AI-powered member of its wealth team, built with technology from Google Cloud and Google DeepMind. Citi planned a phased rollout to clients starting in the summer. In October, Fraser said Citi Sky was in pilot. [Source 27] [Source 1]
Banks described the same idea for their bankers. Bank of America said it launched a generative AI tool for prospecting and onboarding in September. Citizens said about half of its commercial bankers and analysts use AI for research and meeting prep. Regions said it finished rolling out an AI coaching tool for branches, with more than 16,000 practice simulations. [Source 10] [Source 2] [Source 12]
Answering employees' questions
Bank of America said more than 200,000 employees use AI tools, across more than 300 approved uses. JPMorgan Chase CEO Jamie Dimon said in February that 150,000 people use the bank's internal AI platform every week. UMB said it rolled out Glean, an AI tool for day-to-day work, to nearly 1,000 employees. [Source 9] [Source 26] [Source 11]
SouthState said its policy and procedure assistant answers more than 1,800 questions a day on average. It said more than 4,200 employees actively use Microsoft Copilot. M&T said about 75% of its employees have access to AI tools. BNY reported about 220 AI solutions in production in April and more than 400 in October. [Source 4] [Source 3] [Source 24] [Source 22]
Serving customers
Fewer groups described AI that customers use directly. American Express said it launched its first conversational AI voice agents. It also said cardholders can find and book Resy restaurants inside ChatGPT. Customers Bancorp said it put voice agents on its main customer service line. Citizens said it built AI phone support for the most common reasons credit card customers call. [Source 13] [Source 14] [Source 2]
Capital One built Chat Concierge, an AI assistant for car buyers, and executives said at an NVIDIA conference this year that the bank used the same approach to launch AI call summarization for customer service. Synchrony disclosed an OpenAI partnership covering enterprise access and commerce through AI agents. Huntington announced The 1866 Studio, an AI platform for small business owners built with Google Cloud. [Source 25] [Source 16] [Source 17]
Lending and operations
Customers Bancorp CEO Sam Sidhu said cycle time for its AI loan underwriting workflow fell 95%. SouthState said AI performs about 90% of its automated credit spreads. KeyCorp CEO Chris Gorman said the bank is building AI into four areas: commercial onboarding and servicing, commercial originations, consumer servicing and software engineering. [Source 14] [Source 4] [Source 18]
BNY gave two of the more precise operating results. It said certain onboarding processes for corporate clients were completed more than 20% faster in the first quarter of 2026 than a year earlier. It said settlement inquiries handled by AI were investigated more than 80% faster. [Source 24]
Who supplies the models
Banks mostly named outside suppliers. BNY said its Eliza platform integrates models from OpenAI, Google and Anthropic, and that it entered a multi-year collaboration with OpenAI. American Express said it built its dining and hotel experiences using OpenAI models. Schwab named Anthropic. Citi and Huntington named Google. SouthState named Microsoft Copilot and UMB named Glean. [Source 24] [Source 13] [Source 6] [Source 27] [Source 17] [Source 4] [Source 11]
Capital One described a different route. Its executives said the bank relies on open-weight models and customizes them heavily. Open-weight models are ones a user can run and adapt on its own systems. JPMorgan CFO Jeremy Barnum said in October that the bank's strategy is to be able to run both proprietary and open models, to avoid being locked in and to manage cost. [Source 25] [Source 19]
BNY made a similar point about cost. It said Eliza is built to use the right model for each task and to manage token costs, which is what it pays to use different models. The sources we screened did not confirm that any group pretrained its own foundation model. [Source 22]
Where it runs
This is the part of the picture that matters most for data center demand, and it is the part banks say least about. The disclosures we found fall into three patterns.
The first is a bank's own hardware alongside the cloud. BNY said it built on-premise GPU infrastructure powered by NVIDIA SuperPODs to supplement its cloud services. PNC CEO Bill Demchak said in October that the bank had put GPU chips in its data centers. He pointed to depreciation from earlier data center rebuilds as where the cost shows up. He did not give the chip count, the timing, or the purchase or lease terms. [Source 24] [Source 21]
The second is a bank data center used for some workloads. Barnum said JPMorgan runs some open models in its own data center. He did not say which workloads, or who owns the servers. [Source 19]
The third is the cloud. Capital One executives said the bank is all in on the cloud, which gives it access to thousands of NVIDIA GPUs for both training and running models. They did not name the cloud provider, the price or the reservation terms. Citi built Citi Sky on Google Cloud. [Source 25] [Source 27]
Northern Trust discussed running workloads across a mix of its own infrastructure and partners such as Microsoft, without saying where individual workloads run. Fifth Third CFO Bryan Preston signaled spending to come, saying deploying AI takes near-term investment in compute capacity and infrastructure without naming a vendor or location. [Source 29] [Source 23]
What is missing matters as much. A GPU rental gives a bank access to computing power. A high-density GPU colocation contract provides space, power and cooling for equipment in a facility run by a provider. A bank can use both, just as many enterprises weigh colocation against cloud as workloads grow. No group disclosed either one for AI. Saying hardware is on-premise does not say who owns the building, or whether the equipment was bought or leased.
Some data center figures in bank filings are about customers, not the bank. First Citizens reported $3.023 billion in data center loans. That measures credit exposure to borrowers. [Source 28]
What it costs
No group gave a separate, current AI spending total. The figures on record are technology budgets. JPMorgan said in February that it expected to spend about $19.8 billion on technology in 2026. BNY said it spends about $4 billion a year. KeyCorp put its 2026 technology budget at $1 billion. Raymond James cited more than $1.1 billion a year. [Source 26] [Source 22] [Source 18] [Source 7]
JPMorgan showed why an AI number is hard to pin down. In October an analyst suggested the bank spends $2 billion a year on AI. Barnum said an earlier figure was significantly lower, something like $700 million a year, and that it was stale. He said AI spending is now spread across the whole institution and cannot be separated out. [Source 19]
What it returns
Twenty-two groups said something about results in the third quarter. Usage counts and time saved on a single task were common. Dollar figures were not. Citi's Fraser said financial benefits from AI and automation combined significantly outweigh year-to-date investment. She gave no dollar amounts. Customers Bancorp estimated 76,000 cumulative employee hours saved, based on employee reports it did not validate. It also estimated more than $15 million in annualized staffing capacity and did not isolate how much came from AI. [Source 1] [Source 15]
JPMorgan was direct about the limits. Dimon said in February that the 150,000 weekly users believe they save four hours a week. He said that time is not counted in the bank's project returns and does not show up as lower headcount. In October, Barnum said the benefits are not observable in the bank's current quarterly earnings. He said visible change takes a long time at an institution of its size. [Source 26] [Source 19]
Cullen/Frost CFO Dan Geddes said the bank has not yet reached the point where AI projects fund themselves. [Source 20]
What would close the gap
A filing that ties a named workload to a facility and a purchasing arrangement would show where bank demand for compute, space and power lands, and how much of it reaches available data center capacity. Pairing that cost with a workflow result like Pinnacle's or BNY's would let investors judge the return.
How we did this
Colocation Scout reviewed the 100 largest U.S.-headquartered banking groups that publish annual and quarterly reports, ranked by assets at September 30, 2026. The list covers bank holding companies, savings-and-loan holding companies and standalone banks. That brings in groups such as American Express, Ameriprise and Raymond James.
For all 100 we screened available periodic filings, earnings releases, earnings presentations and earnings-call transcripts for the quarter ended September 30, 2026, through November 17, 2026. We also screened seven supplemental company documents. At least one source was screened in whole or in part for 98 groups, and all four source types were screened for 75. Some call transcripts came from third-party transcript publishers.
For 42 of the groups, including the largest, we also reviewed disclosures from 2025 through the same cutoff. The earlier items used here are BNY's April presentation, JPMorgan's February update, Citi's April announcement and a Capital One conference presentation from 2026. That deeper review is not complete, and it is not a reading of every document those banks published.
The counts describe third-quarter disclosures found in the group of 100 and reflect how we classified each one. Review was AI-assisted and focused on relevant passages and their context. Reported results are management's figures and estimates.