Skip to content
AI news

AI agents: why major banks are shutting them out (and what your business should learn)

Thierry GustinPublished 6 min read
Translucent bank vault door holding small glowing robots behind a barrier of light, illustrating AI agent security
Contents
  1. What happened at US banks?
  2. Why are banks shutting AI agents out?
  3. What is "read-only" access?
  4. Is AI already improving company margins?
  5. What this means for your business
  6. Train your teams before you delegate

The essentials in 30 seconds

  • On 22 October 2026, Axios reported that major US banks had detected third-party AI agents operating inside customer accounts, some apparently acting outside the user's instructions.
  • According to Reuters, banks are now blocking or restricting these agents; Citi confirmed it does not allow third-party AI agents to log into customer accounts or make transactions.
  • In many cases, customers had handed their own credentials to the agent; some banks are considering read-only access.
  • For a small business, five rules are enough to start: never share passwords with an agent, start read-only, use dedicated limited access, keep an action log, and require human approval for every payment.

What happened at US banks?

On 22 October 2026, Axios reported that several major US banks had detected third-party AI agents operating inside customer accounts. According to coverage by eSecurity Planet, some of these agents appeared to act outside the instructions given by the user.

Reuters then confirmed that banks are blocking or restricting these agents, as reported by MoneyCheck. The banks named are Morgan Stanley, which chose to block them, along with JPMorgan, Bank of America, Citi and Wells Fargo. Citi confirmed it does not allow third-party AI agents to log into customer accounts or carry out transactions, according to IBTimes UK.

This is not hacking in the classic sense. It is a question of access: who is acting in the account, with what authorisation, and who is responsible if the agent gets it wrong.

Why are banks shutting AI agents out?

An AI agent is not just a chatbot. It does not only answer: it acts. It logs in, clicks, reads and sometimes executes operations. When it does so with the customer's credentials, the bank can no longer tell the customer from the machine.

The most telling detail is this: in many cases, customers themselves had given their credentials to the agent, according to IBTimes UK. Nobody broke the lock. Someone handed over the key.

For a bank, this creates three problems:

  • no way to know whether an action comes from the customer or the agent;
  • no way to limit what the agent can do, since it has the same rights as the customer;
  • no way to guarantee it follows the instructions it was given.

What is "read-only" access?

Some banks are considering read-only access for agents, according to CPO Magazine. The agent could view information, such as a balance or a list of transactions, without being able to change anything or pay.

It is a simple, powerful idea. It separates two things that are often confused: seeing and acting. An agent preparing a report does not need to be able to make a transfer. The principle applies well beyond banking: email, CRM, accounting, social media.

Is AI already improving company margins?

The same month brought another signal worth noting. On 30 October, Apollo's chief economist Torsten Slok observed that AI-driven productivity gains are not yet visible in company margins outside the big tech firms, according to the summary published by AI SDG.

The two stories echo each other. AI is advancing fast and agents are becoming more autonomous, but value does not come automatically. It comes from structured use, on specific tasks, with guardrails. Plugging an agent in everywhere without rules mostly creates risk.

What this means for your business

You are not a bank. But you have bank accounts, an inbox, invoicing software and customer data. The same principles apply at your scale. Here are five practical rules.

  1. Never give your passwords to an AI agent. If a tool asks for your personal credentials to "act on your behalf", say no. Use the official connections provided by the vendor where they exist.
  2. Start read-only. Let the agent view, summarise and prepare first. Check its output for a few weeks before trusting it with any action.
  3. Use dedicated, limited access. Create a specific account for the agent with only the rights its task requires. If it sorts emails, it does not need access to your accounts.
  4. Keep an action log. You should always be able to see what the agent did, when and on what. No trail, no control.
  5. Require human approval for every payment. No transfer, order or invoice goes out without a team member approving it.

These rules do not slow AI adoption down. They make it possible, because they let you delegate tasks to an agent without betting on its reliability.

To go further, read Perplexity Teammates: AI coworkers that work continuously and IBM, Microsoft and the Paris Call: enterprise AI enters the governance era.

Train your teams before you delegate

Most of the incidents described here start with a human action: sharing a login, accepting a connection, letting a tool act without checking. That is why training is the first line of defence. A team that understands what an agent is, what it can see and what it can do makes better decisions.

I train small and mid-sized business teams to use Claude, ChatGPT and AI agents with clear rules. Explore AI training for businesses or book a free assessment.

Frequently asked questions

Why are banks blocking AI agents?

Because third-party AI agents were operating in customer accounts, sometimes outside the user's instructions. Banks could no longer tell the customer from the agent or limit its actions.

Can I give my login details to an AI agent?

No. In many of the cases reported in October 2026, customers had shared their credentials. Use only official connections and dedicated access.

What is read-only access?

Access that lets the agent view information without changing, paying or sending anything. It is the best way to start.

Can a small business use AI agents safely?

Yes, with five rules: no shared passwords, read-only first, limited access, an action log and human approval for every payment.

Sources
  1. eSecurity Planet, accessed 11 November 2026 (esecurityplanet.com)
  2. MoneyCheck, accessed 11 November 2026 (moneycheck.com)
  3. IBTimes UK, accessed 11 November 2026 (ibtimes.co.uk)
  4. IBTimes UK, accessed 11 November 2026 (ibtimes.co.uk)
  5. CPO Magazine, accessed 11 November 2026 (cpomagazine.com)
  6. AI SDG (Substack), AI News for October 30, accessed 11 November 2026 (aisdg.substack.com)
Portrait of Thierry Gustin

Thierry Gustin

AI trainer and author of the Amazon bestseller La Bible de Claude AI. I help teams integrate Claude into their daily work in English, French and Spanish.

LinkedIn

Would you like to know how AI can help your company?

Book a free 30-minute diagnosis: I'll work with you to identify where artificial intelligence can save your team time.