Three Ways to Connect an AI Agent to Your Bank, Ranked by Who Pays When It Goes Wrong


person using laptop computer holding card

Bank liability has always turned on one question: did the person who moved the money have your permission? Until this month that question was about spouses, co-workers and fraudsters. Now the biggest US banks are answering it for AI agents in advance, in writing, and the answer is that the agent is you.

Citi added a new Section H to its Online User Agreement in November 2026. PNC changed its terms on November 18 and Citizens emailed customers the same week, according to TechCrunch, which also reports that JPMorgan and Bank of America announced similar positions in October. Wells Fargo’s revised Online Access Agreement takes effect November 24 with a full section on AI agents.

The coverage has mostly stopped at “customers are on their own.” That is not quite right, and the gap between the headline and the law is exactly where your setup choices matter. Federal rules still decide some of these losses, the contract decides others, and the way you connect an agent to your account decides which bucket you land in. Below are the three connection methods ranked by who pays when something goes wrong, then the setup that keeps you on the right side of the line.

What Citi’s Section H says, in its own words

The clause is short and blunt. The trigger sentence reads: “If you disregard this caution and provide your Online Credentials to any third-party AI Agent or otherwise permit such agent access to your accounts or allow such agent to use any digital service and to perform any transactions on your behalf, Citi will consider any access and actions taken by the agent as authorized by You.”

Three more provisions do most of the work:

  • Leaks are on you too. “If the Online Credentials that you provided to the third-party AI Agent are further shared, leaked, or otherwise distributed to other persons or AI agents, Citi will consider any access and actions taken with the Online Credentials as authorized by You.”
  • Misread instructions are on you. You are responsible for “unintended actions performed by the third-party AI Agent, or actions taken by the third-party AI Agent that exceed the authority granted by you except as otherwise permitted under applicable law.”
  • Monitoring is on you. “You agree it is your responsibility and not Citi’s to monitor the activities of the third-party AI Agent.”

Two details are easy to miss. First, the definition of “AI Agent” covers software that takes “autonomous, semi-autonomous, or non-autonomous action,” which is broad enough to include a budgeting app that just reads balances. Second, Section H says it does “not apply to your use of any third-party AI Agent, tool, function or feature provided by Citi through Citi Online or Citi Mobile.” The bank’s own agents, if and when it ships them, get different treatment.

Citi’s Agreement for Online Funds Transfers adds the matching language for money movement: agent-initiated transfers “are deemed to have been initiated by me and authorized directly by me, without further investigation.”

Wells Fargo goes further in tone (“Your use of any AI Agent is AT YOUR OWN RISK”) and in mechanics. Its Section 15(f) says “You must retain evidence of such revocation and must provide such evidence to us, at our request,” and 15(g) reserves the right to require “that you register your AI Agents with us.” Its Section 2(d) treats delegate transactions, “including those you did not want or intend,” as authorized by you.

The line a contract cannot move

Here is what the headlines skip. For consumer accounts, the Electronic Fund Transfer Act and Regulation E set the floor, and a bank’s terms cannot drop below it. The CFPB’s Electronic Fund Transfers FAQ addresses almost this exact situation. Question 8 asks whether a bank can rely on an agreement “such as waiving Regulation E liability protections if a consumer has shared account information with a third party.” The answer is “No,” because “an agreement cannot restrict a consumer’s rights beyond what is provided in the law.”

That does not mean Reg E rescues every agent mistake. The rule cuts both ways, and the cut runs right through the scenarios people care about.

Regulation E’s definition of an unauthorized transfer excludes one initiated “By a person who was furnished the access device to the consumer’s account by the consumer, unless the consumer has notified the financial institution that transfers by that person are no longer authorized.” The official interpretation, comment 2(m)-2, is the one that should worry agent users: “If a consumer furnishes an access device and grants authority to make transfers to a person (such as a family member or co-worker) who exceeds the authority given, the consumer is fully liable for the transfers unless the consumer has notified the financial institution that transfers by that person are no longer authorized.”

Swap “family member or co-worker” for “AI agent” and you have the bank’s best argument. The banks did not invent a new liability rule; they wrote down how an old one applies to software you hand your login to. Nobody has tested that reading in court, and no regulator has commented on the agent clauses yet, as Startup Fortune notes. So the table below is this site’s reading of the rule text, not legal advice.

What went wrong Likely bucket under Reg E (consumer accounts) Why
Agent did what you asked, badly (wrong payee, wrong amount) You pay You furnished access and authorized the action; the transfer is not “unauthorized”
Agent exceeded what you told it to do You pay until you notify the bank Comment 2(m)-2, the “exceeds the authority given” case
A fake agent or phishing page tricked you into handing over your login Unauthorized; Reg E limits apply FAQ Question 6: a consumer “fraudulently induced into providing account information has not furnished an access device”
The agent vendor was breached and a third party used your stored login Contested You furnished access to the vendor, not the attacker. Citi’s leak sentence says “authorized”; that sentence carries no “applicable law” qualifier
The account is a business account No Reg E floor Reg E protects consumer accounts; business transfers run on the contract and commercial law

The last row matters for this audience. If you run a small company and a founder has wired an agent into the operating account, none of the consumer caps apply, and Wells Fargo’s Section 17(c) already says that letting an agent in “may adversely impact the commercially reasonable security procedures that you’ve otherwise agreed to, and you accept the risk of such access.”

Three ways to connect an agent, ranked by who pays

1. Read-only through the bank’s own data API (lowest exposure)

PNC’s account-linking page describes an API built to Financial Data Exchange standards that lets you “link your financial information without having to share log-in credentials.” You approve the connection on secure.pnc.com, pick which accounts to share, and the app then shows up under Apps With Access in the mobile app’s Control Hub, where you can revoke it.

Most consumer AI finance tools that only read data now ride this kind of rail through an aggregator. Claude’s Personal Finance beta, which launched November 5 for US Pro, Max and Ultra users, connects through Plaid. Anthropic describes the access as read-only, which TechCrunch summarizes as meaning the tool “can’t move, manage, or invest users’ money.”

Does Citi’s clause still cover this? Read literally, “or otherwise permit such agent access to your accounts” reaches past password sharing, and the clause also says you are “authorizing Citi to disclose your information (including sensitive personal information) to the agent.” So the privacy exposure is yours either way. But a read-only token cannot initiate a transfer, which takes the money-loss rows of the table off the board. That is the whole point.

2. Agent shops, a card pays (middle exposure)

If you want an agent to buy things, keep it on a card network rather than in your online banking session. The agent never holds your bank login, purchases go through the card’s own authorization and dispute rules, and the bank’s agent clause has less to grab onto. We walked through what that dispute path does and does not cover in our piece on what you approve when ChatGPT Wallet asks for Face ID. The catch is the one Visa’s own agent benchmark exposed: an agent that buys the wrong thing has made an authorized purchase, and that becomes a merchant return, not a fraud claim.

3. Handing over your username and password (highest exposure)

This is the case every one of these clauses was written for. PNC’s page is candid that some apps still need you to “give your credentials directly to a third party,” and that those connections do not appear in the Control Hub; the only way to cut them off is to change your Online Banking password. The agent sees everything you see, can do anything you can do, and any copy of the password sitting on a vendor’s server falls under Citi’s “further shared, leaked, or otherwise distributed” sentence.

Bank of America’s October email put the practical version plainly: “If you do share your credentials,” customers should “keep in mind that you’re responsible for activity resulting from that access.”

A setup that holds up under these terms

If you or your team still want an agent near real money, this is the order that keeps losses small and keeps the Reg E protections you still have.

  1. Give the agent its own account. Open a separate checking account, fund it with what you are willing to lose in a bad week, and do not link overdraft protection to your main account. It is the same logic as giving a coding agent an account that isn’t yours: the blast radius is whatever that account can reach.
  2. Prefer read-only, API-based connections. If the sign-in redirects to your bank’s own domain and the app later appears in the bank’s connected-apps list, you are on the API path. If an app asks you to type your bank password into its own screen, you are on path 3.
  3. Know where to revoke. Check your bank’s connected-apps page and, for Plaid-based tools, the Plaid Portal’s disconnect flow. Deleting an app from your phone does not necessarily cut its data access.
  4. Revoke in writing, and keep the proof. Both Reg E’s “furnished” exclusion and comment 2(m)-2 turn on whether you “notified the financial institution.” Wells Fargo now demands that you keep evidence of revocation. When you stop using a credential-based agent, change the password, send the bank a dated secure message saying the agent is no longer authorized, and save it.
  5. Turn on transaction alerts and read statements. Under Regulation E’s liability limits, reporting a lost or stolen access device within two business days of learning about it caps your loss at the lesser of $50 or the amount taken, waiting longer raises the cap to $500, and failing to report an unauthorized transfer within 60 days of the statement can leave later transfers uncapped. Alerts are how you find out inside those windows.
  6. Write down what the agent is allowed to do. It will not bind the bank, but if you ever argue a transfer exceeded authority, or was not yours at all, a dated record of the agent’s scope is the evidence you will want.

For business accounts, treat step 1 as mandatory and step 3 as policy: no credential-based agent on any operating account, period. Without the consumer floor, the contract is the whole story, and our look at how insurers are excluding agent losses suggests your policy will not fill the gap either.

What the banks are really building toward

Read the clauses together and the direction is clear. Citi exempts its own agents. Wells Fargo reserves the right to require agent registration and to “Change, limit, restrict, or discontinue any application programming interface (API) or other access method for AI Agents.” TechCrunch’s reporting frames it partly as competition: agents that help customers dodge fees and low-yield deposit accounts are a threat to bank margins.

There is also no federal rule forcing a sanctioned path. The 2024 Section 1033 open banking rule, which would have required banks to share data through APIs, is enjoined and being rewritten; the CFPB sent its reconsideration proposal to OIRA on August 6, 2026. Until that lands, each bank decides what an acceptable agent looks like, and the sensible bet is that it will look like a registered, scoped, API-based connection, the same scoped-access pattern the agent payment protocols are converging on.

None of that helps someone whose agent overpays a utility bill next week. What helps is the boring part: an agent account with a small balance, a read-only connection where reading is all you need, and a dated record of every time you revoked access.

Ty Sutherland

Ty Sutherland is the Chief Editor of AI Rising Trends. Living in what he believes to be the most transformative era in history, Ty is deeply captivated by the boundless potential of emerging technologies like the metaverse and artificial intelligence. He envisions a future where these innovations seamlessly enhance every facet of human existence. With a fervent desire to champion the adoption of AI for humanity's collective betterment, Ty emphasizes the urgency of integrating AI into our professional and personal spheres, cautioning against the risk of obsolescence for those who lag behind. "Airising Trends" stands as a testament to his mission, dedicated to spotlighting the latest in AI advancements and offering guidance on harnessing these tools to elevate one's life.

Recent Posts