Picture a freelance designer who reconciles her accounts at the end of every month: which client still hasn’t paid, whether a card charge went through twice, whether the power and phone bills are settled. She has heard an AI agent could do this for her, meaning an AI assistant that opens web pages, logs in and works through a task step by step. So she plans to hand it her online banking username and password, let it log in once a week, check the accounts and pay the bills while it’s there.

This month, several large US banks answered a question she hadn’t thought to ask yet: once the agent is logged in, whose actions are they?

The new bank terms: whatever the agent does, you did

Citi, one of the largest US banks, updated its online banking user agreement in November with a new section on third-party AI agents. It first cautions customers against using a third-party AI agent to access their accounts. Then it states that Citi will treat any access and action by the agent as authorized by the customer, that Citi has no obligation to investigate whether those actions really were authorized, and that monitoring the agent is the customer’s responsibility, not the bank’s. Another section of the same agreement already forbids customers from giving their username and password to anyone else.

Citi is not alone. Startup Fortune, a tech news site, reported on November 18 that the regional bank PNC changed its terms that day, and that Citizens Bank emailed customers that they are responsible for what AI assistants do. Wells Fargo goes a step further. According to American Banker, a US banking trade publication, once its new online access agreement takes effect on November 24, customers are responsible for revoking an AI agent’s authorization in time and must be able to prove they did, or what the agent does afterwards still counts as theirs.

Why “I watch every step” isn’t enough

The strongest objection goes like this: the agent asks before every payment, she sees each one with her own eyes, most of what she wants is just checking balances, and the amounts are small.

That argument misses three things.

First, an online banking login has no “look but don’t touch” version. The agent gets the whole account. If it can read statements, it can also send transfers, change the notification email and add payees. She may intend it only to check balances; the bank’s systems can’t tell the difference, and don’t need to.

Second, what she gives up is fraud protection. The US rule for electronic transfers is called Regulation E, and it is interpreted by the Consumer Financial Protection Bureau (CFPB). For an unauthorized electronic transfer, a consumer who reports within two business days is liable for at most $50. But the same rule draws a clear line around “unauthorized”: a transfer started by someone you gave your access device to doesn’t count, unless you have told the bank that person is no longer allowed. The official interpretation adds that even if that person exceeds the authority you gave, you are fully liable. What the new bank terms do is put AI agents explicitly into the box of “someone you gave your access device to.”

Third, the risk isn’t only the agent’s own mistakes. Startup Fortune reports that Citi’s terms also put on the customer what happens after an AI tool leaks credentials or gets compromised. Where the agent stores the password, which servers it talks to, whether a malicious instruction on a web page can trick it, she can’t see any of it, yet the terms make her answer for it.

Advertisement

The same job, by a different route

This doesn’t mean AI can never touch money. The difference is the channel. The same week, Google made the Task mode of Gemini, its AI assistant, free in six countries including the US, Japan and Korea, so it can book appointments, handle bills and shop for users. Google’s official blog says payments use tokenized cards from the card networks, meaning the merchant receives a stand-in number rather than the real card number. Users can set a spending limit, and Gemini asks before it clicks Buy.

How AI reaches your moneyWho the bank sees actingCan you revoke only the agent?Size of the loss if it goes wrong
Your online banking username and passwordYouNo; change the password and notify the bankThe whole account, without fraud protection
A tokenized card with a spending limitOne authorized transaction on that cardYes; disable that one numberCapped at the limit you set
Exported statements, or the bank’s own featuresThe bank itself, or no login at allNothing to revokeThe agent can’t move money

Put the designer’s needs into this table and the answer is clear. Checking accounts needs data, not a password: export the statements as files and let AI compare them. Bills go on the bank’s own autopay. If she really wants an agent to place orders, she gives it a capped tokenized card, not her banking password. Why an agent working across systems should get a fresh key at each stop, good only for that one job, is covered in more depth in When an AI agent works across systems, it needs a new key at every stop.

A cap limits the loss, not the mistake

A capped channel isn’t foolproof either. If the agent buys the wrong thing or books the wrong date within the limit, that is still a transaction you authorized; whether you get your money back depends on the merchant’s terms, not on the bank’s fraud protection. For how merchants view agents, see When an AI agent checks out for you, the seller’s terms decide who pays for mistakes. Gemini also offers an “Always allow” option so routine actions stop asking every time. Turn it on and the cap becomes your only safeguard, so set it at an amount you are willing to lose in full.

The legal side isn’t settled either. The same Startup Fortune report notes that no court or regulator has yet tested the reading that everything an agent does counts as authorized. The rules and terms above are American. In Taiwan or elsewhere, the answer depends on how your own bank’s terms treat a third party operating your account, and that is worth ten minutes of reading.

One thing to do this week: open your browser and password manager and check whether any AI tool, browser agent or extension has stored or used your banking password. If one has, change the password first. Then write to the bank through its support email or secure messaging, with one sentence: “I no longer authorize any third-party AI agent to access my accounts.” Keep that message. Under the rule above, notifying the bank is exactly what ends the earlier authorization; under Wells Fargo’s new terms, that message is your proof.

AI handoff card

I want an AI agent to help me with money tasks, such as reconciling accounts, paying bills or shopping online, and I want to decide first which kind of access to give it.
The method comes from this article: https://boosterminiclass.com/en/posts/ai-agent-with-your-bank-password-is-you/
The article's position: when an AI agent logs in with my online banking username and password, the bank treats everything it does as authorized by me, and fraud protection no longer applies. So I don't give my banking password to third-party AI. If AI is going to touch money, it only goes through channels that are capped and can be revoked on their own, such as a tokenized card with a spending limit, exported data, or the bank's own features.

First, ask me 3 questions:
1. Which tasks do I want to hand to the agent, and which of them only need to read data and which actually move money?
2. How am I planning to give it access: my banking login, a card saved in my browser, a capped tokenized card, or exported files?
3. Do my bank's or card issuer's terms mention AI, automated tools or third parties operating my account, and what do they say?

Based on my answers, give me:
- for each task, which kind of access it should use, and why;
- if I have already given my banking password to a tool, the steps to take next, including how to keep a record of notifying the bank;
- if I use a capped card, how to set the cap, based on an amount I am willing to lose in full;
- a one-line verdict: hand it to the agent / give it data, not the password / do this myself.
Do not ask me for any password or card number, and do not log in or pay for me. The final decision is mine.
Advertisement

Share

Share this mini class

If this lesson helps untangle a work bottleneck, share it with someone deciding how to use AI.

References