Citi Says You Own Your AI Agent's Mistakes. Does Your Startup?
Citi, PNC and Citizens just joined JPMorgan and Bank of America in telling customers that anything an AI agent does with their account is on them, even if the agent gets hacked. If your team connects agents to company money, or you build agents that touch money, here is what changed and a one-page policy to adopt this week.

Picture this. Your ops lead connects an AI assistant to the company checking account so it can pay vendor invoices. It works great for a month. Then the assistant's provider has a breach, someone gets the stored login, and $40,000 leaves the account overnight.
Who pays?
A growing list of big U.S. banks just answered that question in writing: you do.
TechCrunch reported on November 18 that Citi updated its online user agreement this week to say customers who give third-party AI agents access to their accounts are responsible for what those agents do. That includes cases where the AI tool leaks credentials or gets hacked. PNC changed its terms the same day. Citizens emailed customers with a similar warning.
None of this is a reason to panic or ban AI from your finance stack. It is a reason to know exactly which agents can touch money at your company, and how.
What the banks changed
Here is the list so far, based on TechCrunch's reporting:
- JPMorgan and Bank of America, the two largest U.S. banks, announced similar policies in October.
- Bank of America emailed customers a message titled "Protecting your finances in the AI World," advising against handing login credentials to anyone, including AI tools. Per TechCrunch's October 27 story, the bank said a customer who shares credentials is responsible for activity resulting from that access.
- Citi, the third-largest U.S. commercial bank, changed its terms this week. Third-party agents are the customer's responsibility. Agents Citi provides itself can be used without that extra liability.
- PNC updated its terms on November 18.
- Citizens sent customers an email saying they are liable for anything their AI assistant does.
Across these policies, the banks say they will not reverse transactions made by automated assistants, whether the customer meant for them to happen or not.
A Citi spokesperson told TechCrunch the bank wants customers "to understand the risks" as third-party AI tools spread.
Why banks are drawing the line now
There are two reasons, and both matter to you.
Risk. An AI agent holding your username and password is one more place those credentials can leak. If the agent's company gets breached, or the agent is tricked by a malicious prompt, the bank can't easily tell a legitimate transfer from a theft. Bank of America's October note also warned that agent activity could trigger a lockout if the bank flags it as an intrusion.
Competition. TechCrunch notes that banks see AI agents as a threat to their position. An agent that shops for better savings rates or helps customers avoid fees could pull money away from the accounts that make banks money. Markets dipped briefly on that possibility. Banks want to own the channel, which is why Citi carves out its own agents.
The practical result: if you use a bank's own tools, you may keep the usual protections. If you use someone else's agent with your password, you probably don't.
Your business account already had thinner protection
Many founders assume the fraud protections they know from personal banking apply to the company account. Often they don't.
The federal rule most people have in mind is Regulation E, which governs electronic transfers and limits a consumer's liability for unauthorized ones. But Regulation E defines an "account" as one "established primarily for personal, family, or household purposes." A business checking account usually falls outside that definition. Your protection there comes mostly from your account agreement with the bank.
There is a second catch even for personal accounts. Under the same rule, a transfer is not "unauthorized" if it was started by a person the consumer gave access to, until the consumer tells the bank that access is revoked. Handing a tool your login looks a lot like giving it access.
So the new bank policies mostly put in writing a risk that was already there. (This is general information, not legal advice. If real money is at stake, ask your banker and your lawyer what your specific agreement says.)
Your AI co-founder is ready when you are.
Foundra turns everything in this article into an actual plan. Validation, customers, pricing, launch. In one place, in your voice, in an afternoon.
Get started→$39/month. Cancel anytime.
Where agents touch money at a startup
Make a list. Most early teams have more of these than they think:
- Bookkeeping and spend tools connected to the bank through an aggregator.
- Accounts payable automation that can schedule or release payments.
- AI assistants with browser access that log in to the bank site on someone's behalf.
- Agentic shopping for software, ads or supplies on a company card.
- Coding agents with access to payment keys in your environment variables.
- A founder's personal account that quietly pays company bills because the business account isn't open yet.
For each one, write down two things: what credentials it holds, and whether it can only read data or can also move money.
Read-only versus read-write
This distinction does most of the work.
Anthropic's new Personal Finance tool in Claude, launched in early November, connects accounts through Plaid and is described as read-only: it can see balances and up to two years of transactions but "can't move, manage, or invest your money," per TechCrunch. That kind of access can still leak data, which is a real problem, but it can't wire your runway to a stranger.
A tool that holds your actual banking password, or that can approve payments, is a different category. Treat it like an employee with signing authority.
A few rules of thumb:
- Prefer official connections (the bank's own API, OAuth or an approved aggregator) over pasting passwords into a tool.
- Prefer read-only for anything that only needs to analyze.
- Use the bank's built-in controls: separate accounts, payment approval limits, two-person approval and alerts on every outgoing transfer.
- Keep a small operating account for any tool that can pay, funded on a schedule, so a bad day has a ceiling.
A one-page agent money policy
Five people can adopt this in an afternoon:
- No shared bank passwords with any AI tool. Ever. Use official integrations only.
- Read-only by default. Any agent that can move money needs written approval from a founder.
- Hard limits. Payment-capable tools connect only to a separate account with a set maximum balance.
- Two-person approval for any single payment above an amount you choose, such as $2,500.
- Alerts on. Every outgoing transfer triggers a notification to two people.
- Quarterly review. List every tool with bank or card access and remove what you don't use.
- Personal accounts stay personal. No company bills paid from founders' accounts once the business account exists.
Write it down where your team keeps operating rules, whether that is a shared doc or a planning tool like Foundra, and walk new hires through it in their first week.
If you are building an agent that touches money
This week's news lands differently if your product is the agent.
Your users' banks are now telling them they're on their own if your tool misbehaves. That makes trust your main feature. Some ways to earn it:
- Ship read-only first. Many useful jobs, like categorizing spending or spotting duplicate subscriptions, don't need write access at all.
- Never store raw bank passwords. Use tokenized connections.
- Make every payment confirmable. Show the amount, the recipient and the reason before money moves.
- Offer limits users set themselves, per transaction and per day.
- Say clearly in your terms what you cover if your system is breached. Vague language reads as a red flag now.
TechCrunch has also reported that very few consumers are using AI agents to book rides or order food so far. Adoption of agent payments is early. The startups that handle the liability question well may be the ones customers trust when it grows.
Frequently asked questions
Which banks say customers are liable for AI agents? According to TechCrunch, JPMorgan and Bank of America announced policies in October, and Citi, PNC and Citizens followed in November. Each holds the customer responsible for what a third-party agent does with account access.
Does this apply if the AI tool gets hacked? Citi's updated terms say yes: if a third-party AI tool leaks credentials or is hacked, the customer is responsible for what happens with the account.
Are business bank accounts covered by Regulation E? Generally no. Regulation E applies to accounts established primarily for personal, family or household purposes. Business accounts rely mostly on the account agreement.
Is it safe to connect a read-only finance tool? It is lower risk than giving a tool payment access, since it can't move money. It still exposes data, so check the provider's privacy and security terms.
Can I still use AI for bookkeeping? Yes. Use official, read-only connections where possible, keep payment-capable tools on a separate limited account, and turn on approval rules and alerts.
What should an agent startup change? Default to read-only, avoid storing passwords, require confirmation before payments, let users set limits, and state plainly what you cover after a breach.
Sources
- More banks say customers using AI agents are on their own, even if they get hacked, TechCrunch, November 18, 2026
- Bank of America kindly asks that you don't share your banking password with your AI assistant, TechCrunch, October 27, 2026
- Claude can now analyze your spending with its new Personal Finance tool, TechCrunch, November 5, 2026
- 12 CFR 1005.2, Definitions (Regulation E), Legal Information Institute, Cornell Law School
You just read the theory. Ready to build the thing?
Foundra is your AI co-founder. It turns an idea into a validated business plan, a go-to-market, and your first 10 customers. In an afternoon, not a semester.
$39/month. Cancel anytime. Works in 20 languages.