JPMorgan Chase is blocking external AI agents from logging in to customer accounts with shared banking credentials and carrying out transactions. The bank says autonomous tools pose security risks for both customers and the lender.
CEO Jamie Dimon laid out the position on the bank’s third-quarter earnings call on October 13, 2026. According to CNBC’s reporting, he told analysts the bank will not let just any outside agent come inside the company, because it is too risky for clients and for JPMorgan.
This is not a blanket ban on AI-powered finance. JPMorgan supports secure, customer-authorized data sharing through open banking providers and says it wants to enable agentic connections safely. But under the current approach, external agents cannot use a customer’s login to send money or pay bills.
Why JPMorgan is drawing the line
AI assistants are moving from chat tools to agents that complete tasks. In banking, that could mean checking balances, spotting better savings rates, or preparing payments. With enough access, an agent could also start transfers on its own.
That brings risks a chatbot never carried. An agent might misread an instruction, go beyond its intended scope, or fall for malicious commands. If it logs in with a customer’s credentials, the bank may struggle to tell the customer’s own actions from the software’s.
Dimon warned against “rogue agents” operating at scale, with thousands of autonomous systems probing the bank’s infrastructure. The concern goes beyond model accuracy. It is about verifying access, controlling permissions, and protecting customers when outside software acts with some independence.
A system that shows financial information carries one level of risk. A system that moves money carries another. An unauthorized transaction can mean losses, fraud investigations, and disputes over who is responsible.
What agents can and cannot do
JPMorgan separates access to information from authority to act. According to CNBC, external agents may reach some customer data, such as balances, through connectors and open banking providers once the customer authorizes it. They cannot use the bank’s credential-based login to act for customers.
- Customer-authorized data access: An approved connection may let an app retrieve permitted information, such as balances or transaction history.
- Direct account operation: Third parties cannot use customer credentials to log in and run retail accounts, including sending money or paying bills.
A read-only link lets an AI assistant analyze spending and organize information. Transactional access lets an agent affect a customer’s money directly. That gap will likely sit at the center of agentic banking design. JPMorgan has announced no timeline for safeguards that would let outside agents interact directly with accounts.
Who pays when an agent makes a mistake?
The decision raises a hard question for consumers. Many may assume an action by their chosen AI assistant equals an action in their bank’s own app. Banks may see it differently. They may distinguish between transactions made through their security processes and activity started by outside software using shared credentials.
Dimon urged customers to understand the risks of authorizing outside agents, including what he called a possible “liability transfer.” CFO Jeremy Barnum also warned about customers handing credentials to an agent that then does something unwanted.
These comments do not mean customers are automatically liable for every agent error. Actual liability depends on the circumstances, the law, account agreements, and payment rules. Still, the uncertainty matters in a young market.
Picture an assistant told to find a better savings rate. The customer wants recommendations. The agent might treat the request as permission to open an account or move funds. If the result is unwanted, someone must work out what was authorized and whether controls failed. The bank, the agent’s developer, the data intermediary, and the customer could all be involved. Clear permissions and clear accountability will be essential for trust.
JPMorgan builds its own AI tools.
The restrictions do not mean the bank is backing away from AI. Dimon said JPMorgan uses AI agents internally and plans customer-facing applications. Barnum said the bank is piloting an internal tool called Smart Cash. It is designed to forecast a customer’s cash flow and move funds to maximize yield.
Based on CNBC’s reporting, Smart Cash is a targeted pilot, not a rollout to all customers. It does point toward more proactive banking, where AI helps customers anticipate expenses and put idle cash to work.
The real contrast is between tools built inside a bank’s controlled environment and agents from independent firms. A bank can design internal systems around its own authentication, monitoring, and risk controls. Outside agents may span many services, models, and permission setups. That does not make them unsafe by default. It does mean banks need a reliable way to confirm what an agent may do and who answers for it.
Could AI agents shake up deposits?
The debate is not only about security. Agents could change how people compare products and move money. Switching savings accounts today means researching options, checking eligibility, and starting transfers by hand. An assistant could monitor rates and shift funds with far less effort.
That matters because deposits fund banks. If customers react faster to rate differences, some institutions could face more pressure to keep balances. Meta’s Muse assistant added to market worries about money flowing from low-yield accounts to better-paying ones. CNBC also reported that Instinct, a personal AI concierge from Spear Street Technology, reached a $10 billion valuation in September 2026.
Treat this as a future risk, not a settled outcome. Agents still face permission limits, eligibility rules, transfer restrictions, and the need for customer approval. Many consumers may hesitate to hand big financial decisions to software. Even so, easier comparisons could push banks to compete harder on rates, fees, and service rather than relying on customer inertia.
Regulators step up scrutiny.
JPMorgan’s move comes as regulators in the US and Europe look harder at advanced AI in finance. Reuters reported on October 13 that US regulators, including the Federal Reserve, are asking banks about advanced AI models and agentic systems. Reported concerns include how autonomous tools are controlled and audited, reliance on outside tech providers, cyber threats, and recovery if critical systems are compromised. European authorities have also pushed banks to strengthen cyber-recovery plans.
This is not a universal ban on AI agents in banking. It reflects concern about deploying capable systems without hurting resilience or stability. Agents can chain many actions together, so banks must watch permissions and decision paths as well as outputs.
Possible safeguards include identity checks, transaction limits, human approval for sensitive actions, audit trails, and quick ways to suspend access. These are general control ideas, not a confirmed description of JPMorgan’s future design.
What it means for agentic banking
Three issues will shape what comes next.
- Permission-based access: Providers need secure ways to show what a customer has authorized. Sharing a password is not the same as granting limited, transparent, and revocable permissions.
- Liability and consent: Customers should know when an agent only analyzes, when it prepares an action, and when it can execute one. Banks and AI firms need workable dispute processes.
- Competition: Proprietary tools give banks control over security and experience. Yet customers may want independent assistants that work across several banks.
The likely future is not all-or-nothing. A tiered model could apply stricter controls as agents move from reading data to moving money.
For now, JPMorgan has drawn a boundary around outside agents using customer credentials to run retail accounts. It backs secure data sharing and says it wants safe agentic connections. It has not said when outside agents might execute transactions through a safeguarded system.
Agentic banking will depend on who can access accounts, what they may do, and who is accountable when something goes wrong. For banks, AI firms, and regulators alike, earning customer trust may matter as much as building smarter agents.


Comments
No comments yet. Be the first to share your thoughts.
Leave a comment