JPMorgan and Wells Fargo Earnings Show Banks Building AI Agents, Models and Blockchain Rails In-House

Artificial intelligence (AI) agents, AI models and blockchain payment rails were each expected to take something away from the traditional bank. Agents would take the customer’s login, AI labs would take the technology, and crypto networks would take the movement of money. The third-quarter JPMorgan and Wells Fargo earnings calls on 13 October 2026 showed two of the largest banks in the United States, between them, building all three for themselves. The limit of that strategy sits where AI agents and bank accounts meet: the customer who hands a bank password to an agent the bank did not build.

Each piece of that picture has a direct consequence for anyone who banks with a large lender, uses a budgeting app or has tried a personal AI agent with their account details. Together they describe how the largest players in financial technology are absorbing the tools that were supposed to bypass them.

The deposit fear that moved bank shares

Analysts raised the agent question with both banks. JPMorgan’s chief financial officer, Jeremy Barnum, acknowledged that recent launches of personal AI agents had produced a brief reaction in bank stocks, as investors weighed whether agents would sweep idle cash into higher-paying accounts. He described the threat as two familiar pressures arriving together, savers looking for better rates and fintech apps offering a smoother experience, and said JPMorgan would rather win the role of a customer’s main bank than match every rate-sensitive deposit. The bank added about 650,000 net new checking accounts in the quarter.

Wells Fargo chief executive Charlie Scharf was more sceptical of the threat. He said high-rate alternatives existed well before agents did, that agents handing savings to unfamiliar institutions are a distant prospect, and that people keep money where their everyday payments run. The Bright Minded has mapped what AI agents would move in an agentic bank run, and JPMorgan’s answer adds the condition that scenario depends on: an agent can only move what the bank lets it reach.

The agent: JPMorgan decides which ones reach the account

Jamie Dimon said outside agents will not get open access to the bank, because rogue agents could number in the thousands and the risk to clients would be too high. Barnum explained how some agent services get in today. Credential sharing means the customer hands the agent a username and password, and screen scraping means software signs in as the customer and reads the account page.

JPMorgan considers both methods insecure and blocks them where it can, Barnum said on JPMorgan’s results call, and the bank is working on a secure alternative. Dimon said JPMorgan already uses agents inside the firm and plans to put them in front of customers, taking no position for or against agents in general. The line falls at agents that arrive carrying the customer’s own key.

The liability lands on the customer who shares a password

Barnum said a customer who shares credentials bears the consequences if the agent then acts against that customer’s wishes. Dimon told customers using agents to understand that liability transfers to them, together with the risk of their data being misused. The reasoning follows from how a login works: once an agent signs in with the customer’s password, the bank’s systems see the customer.

That blind spot sits at the core of the authorisation problem in autonomous payments. A bank can confirm that the right password was entered. It has no record of whether the account holder approved the specific transfer the agent then made.

The model: JPMorgan runs AI without depending on one lab

Barnum described a strategy built to avoid reliance on any single AI provider. JPMorgan combines its own models with open-weight and open-source ones, the kind any company can download and run, and some of them could operate inside the bank’s own data centres. Keeping the models close lets the bank control what it pays per query, since AI providers bill by usage.

Barnum also said JPMorgan is moving away from putting a figure on what AI saves. Earlier estimates mixed in older machine-learning and fraud work, the gains do not yet show up in quarterly earnings, and usage costs will absorb part of the productivity the tools create.

The rail: Wells Fargo moves corporate money on its own network

Scharf said Wells Fargo launched tokenised deposits on its proprietary platform last month for corporate and commercial clients. A tokenised deposit is bank money recorded on a blockchain ledger, so it can move at any hour while remaining a deposit at the bank. Mike Santomassimo, the chief financial officer, said the service runs at small volumes for now, with cross-border payments arriving in the recipient’s account within two minutes.

According to Wells Fargo’s tokenised deposit announcement, these tokens keep the protections and insurance eligibility of the bank’s standard deposits and settle at weekends and on holidays. The first corridor converts dollars to sterling for a limited group of clients, with wider access in other markets and currencies planned during 2027.

A stablecoin venture formed by 21 institutions

Stablecoins are the crypto industry’s own answer to instant money: tokens on a blockchain that hold a fixed value against a currency. Scharf said American use cases for them remain unclear, yet Wells Fargo is setting up a stablecoin issuer with 20 other financial institutions from several countries, aiming to issue from the first half of 2027. The group includes Bank of America, Goldman Sachs, Santander, Deutsche Bank and Lloyds Banking Group, and its first planned token is pegged to the dollar.

The cost: scale pays for building in-house, and headcount falls

Scharf argued that the benefits of scale in AI are underappreciated, since an agent built once for a call centre or an operations team can then serve the whole organisation. The same logic helps a bank of that size justify building agents, models and rails itself. Wells Fargo’s third-quarter results release attributes part of its lower expenses to an 8% cut in headcount from a year earlier.

Scharf told analysts the workforce has shrunk in each quarter since mid-2020 and now stands roughly 30% below that peak, so the decline began before the agents arrived. He told Reuters in November 2025 that technology, AI in particular, would reduce staffing further, and that he preferred to achieve most of the reduction through attrition. The return on building in-house is measured in roles the bank no longer needs to refill.

Between them, JPMorgan and Wells Fargo now build the agent, run the model and own the payment rail, and the one link they cannot control is the customer who hands a password to someone else’s agent.

Editor’s note

Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.