Every platform shift has a moment when the incumbents who own the rails decide the terms of access. For agentic commerce, that moment arrived this week: the largest banks in the United States told outside AI agents they can look, but they cannot touch the money.
The line banks drew
On JPMorgan's third-quarter earnings call on October 13, CEO Jamie Dimon said the bank will not allow "just any outside agent" inside the company because it is "too risky for clients and for JPMorgan" — adding that customers should know "you don't want rogue agents, and there could be thousands of them."
His sharper line: anyone giving an agent access to their account should know "there's a liability transfer."
A JPMorgan spokeswoman told CNBC that the bank blocks outside agents acting for retail customers and prohibits third parties from entering accounts by harvesting customer credentials. What it does accept: customer-authorized data sharing through open-banking connectors. In practice, an agent can see balances in read-only mode — via Plaid or similar — but cannot send money or pay a bill.
Charles Schwab joined the posture days later. CEO Rick Wurster told Bloomberg the firm is actively blocking unauthorized AI agents — including Meta's Muse — from moving client money. "We don't want someone just giving their password to Muse, and letting Muse come in and take action on that client's behalf," he said. "We can't control it, the client can't control it, something could go wrong and who's liable?" Wurster said he gets a weekly report showing these blocks happening ten to twenty times per week for major agents.
Bank of America has separately emphasized the need for human oversight in AI-driven payments, while regulators in the U.S. and Europe are increasingly focused on the fraud and financial-stability risks of machine-driven activity.
Why now
The trigger is visible. Meta's Muse agent launched September 21 with agentic shopping through partners including Walmart, Best Buy, Sephora and Expedia — and Amazon already blocked Muse from Amazon.com over unauthorized access and stored credentials. Banks see the same pattern arriving at their rails: agents that log in as the customer, act as the customer, and break the bank's ability to know which is which.
The concern is not theoretical. A single compromised agent could scale mistakes or malicious actions across thousands of accounts — and the credential-sharing model it relies on is one banks have opposed for years. JPMorgan CFO Jeremy Barnum said on the same call that some money-moving services work with customers' own login details: "It's not secure," and the bank blocks them where it can "to protect customers and protect ourselves."
Open banking vs credential sharing
This is the structural fault line:
- Credential sharing (the agent logs in as you) makes the bank blind. It cannot distinguish a legitimate agent from a spoofed one, cannot scope permissions, and cannot revoke access without breaking the customer's login.
- Open banking (scoped, revocable, consent-based API access) keeps the bank in the control loop. It is why read-only access through authorized connectors remains permitted while write access does not.
The banks are not banning AI in finance — they are banning the unaudited version of it. JPMorgan is piloting Smart Cash, an AI tool that predicts cash flow and moves money between a customer's own accounts to maximize yield. Agents, yes — but ours.
The liability question
Dimon's "liability transfer" framing is the crux. When an agent moves money and something goes wrong, who pays? Congress is still debating how AI agent liability should even work, and no bank wants to be the test case for underwriting decisions made by an opaque third-party model. Until liability is allocated by statute or contract, the rational bank posture is: no write access for outside agents, full stop.
The playbook for fintech founders building on bank rails
- Design read-first. Build delightful client-side orchestration on data you can read through authorized open-banking connectors. Assume read-only is the default grant.
- Earn write access in narrow, auditable slices. Fine-grained scopes, explicit consent experiences, immutable logs, and escrowed funds make a bank comfortable treating you as an extension of its control plane.
- Build for liability allocation. Insurance-backed execution models and clear contractual responsibility for agent errors are now product features — they are what stand between you and the block list.
- Plan for the bank-owned agent. JPMorgan's Smart Cash shows incumbents will not cede the automation layer. Partner where you can differentiate on top of their rails; do not assume they will stay dumb pipes.
- Watch the regulators. The first clear liability rules — in the U.S. or Europe — will be the starting gun for write-access negotiations. Position your architecture to adopt them on day one.
The gate is real, but it is not closed forever. It is a compliance-shaped door, and the founders who build the key — consent, auditability, liability clarity — will be the ones who get through it.
Angeled.com connects founders with investors.
Sources: Bloomberg Tax/Bloomberg reporting, October 16, 2026; JPMorgan Q3 2026 earnings call, October 13, 2026 (via CNBC, CellCog); Digital Trends, October 16, 2026.
Read Next
The Finance Handoff Is Being Automated: Why Seed Capital Is Flooding Agentic B2B Payments
Bitcoin Life Insurance at $350M: Crypto's Grown-Up Financial Infrastructure Arrives
Apple Pay Finally Launches in India — Nine Years After the U.S. Debut
Samsung Partners with Solana to Bring USDC Transfers to 82 Million Galaxy Devices
Samsung Wallet Adds USDC for 82M Galaxy Devices: Stablecoins Just Went Mainstream
Fintech's Comeback: Where the Money Is Going Now




