For banks and credit unions
AI Agents for Banks
AI agents for banks are software that can take an action on banking systems, under rules the bank sets, with a confirmed step before any money moves. A chatbot answers and hands off. An agent finishes the task the authenticated customer asked for, such as a transfer, an ACH payment or a domestic wire, and records every step.
Read the agentic banking guide
What an AI agent for a bank is
An AI agent for a bank is software that can take an action on banking systems when an authenticated customer asks. Rules the bank sets decide whether that action can proceed. Before any money moves, the customer confirms the exact details.
That is different from software that only drafts an answer. The category name for the money-moving version is agentic banking. How the payment itself should be checked is covered under agentic payments.
A chatbot and an agent
Most banking AI still stops at an explanation. An agent is the system that carries the authenticated request through to a finished task, inside limits the bank controls.
| Chatbot | Banking agent |
|---|---|
| Answers, then hands the customer to a form, a page or a person | Finishes the task the customer asked for |
| Explains how to send a wire | Prepares the transfer and asks the customer to confirm the exact amount, accounts and payee |
| Keeps its rules in the prompt | Keeps rules in code the model cannot change |
| Speaks with a generic script | Speaks in the institution's voice |
A test you can apply to any product
Three questions sort a demo quickly. If any one of them is missing, treat the product as a chatbot.
It finishes the transaction
It can complete a transfer, an ACH payment or a domestic wire, and it asks the customer to confirm the exact details.
The bank sets the rules
Limits and approvals live outside the model, in code the model cannot rewrite.
Every step is logged
The request, the checks and the action can be reconstructed later by compliance or an examiner.
What the customer-facing agent can do, and what stays with a person
Lead with the external agent: the one a signed-in customer uses to check an account or move money. That is the agent this page is about.
What the agent can do
- Check balances, accounts and holds.
- Search and explain transactions, such as how much was spent on dining last month. The math is done by a calculator. The model does not guess the amount.
- Move money by internal transfer, ACH or domestic wire. The customer confirms the exact amount, accounts and payee before anything is sent.
- Manage payees and saved wire templates.
- Link an external account and set up automatic savings.
- Show spending trends and simple forecasts, always labeled as estimates.
What stays with a person
These requests are handed to a person.
- Loans, mortgages and credit cards.
- Disputes, lost or stolen cards, and fraud freezes.
- Opening or closing an account, and changing contact information.
- International wires.
- Investment, tax or legal advice.
Internal agents for staff
Internal agents for staff are also available. They help an authenticated employee find a procedure the institution already wrote, including for teller work and lending operations. They assist the employee. They do not replace the employee, and they do not make a credit decision.
- The employee is authenticated, and access follows that role.
- The same ideas apply: rules outside the model, a log of what was asked and answered, and a way to switch the agent off.
- The customer-facing money-moving agent remains the primary one. Segment notes are on AI for community banks and agentic AI for credit unions.
The risk question
The risk is real because the agent can request a payment. These are the controls that keep that request inside rules the bank sets. The agent acts only when an authenticated customer asks and confirms.
Requests, then a rules layer
The agent can request an action. It never touches payment rails directly. A separate rules layer decides whether that action can proceed.
Hard limits
The institution sets per-transaction, daily and monthly caps. A payment over a hard limit is declined. Nothing overrides that limit.
Confirmation, then a step-up
Every payment, at any amount, needs explicit confirmation of the exact details. Larger payments also need a one-time code.
Payee screening
A new payee is screened for sanctions (OFAC) and account risk before that payee can be paid. If screening is unavailable, the payee is not added.
Only the tools that request needs
The agent receives only the tools that request needs. A tool it was not given cannot be called, no matter what anyone types.
Tricky instructions
Hidden or tricky instructions in a message or a document are blocked or treated as plain data, and flagged for the institution.
Separate data, full log
Each institution's data is kept separate. Every conversation and every action is logged step by step.
The institution stays in control
A dashboard shows activity and security alerts. Staff can edit spending policy, block a user and switch an agent off immediately. Staff can also try the agent in a sandbox with realistic data before it goes live.
Preparation for an exam conversation is on the AI agent exam readiness checklist. It is preparation, not legal advice.
Related guides
Common questions
What are AI agents for banks?
AI agents for banks are software that can take an action on banking systems when an authenticated customer asks, under rules the bank sets, with confirmation before money moves. The primary agent is customer-facing and finishes tasks such as balances, internal transfers, ACH and domestic wires. A wider definition is on agentic banking.
What is the difference between a banking chatbot and an AI agent?
A chatbot answers and then hands the customer to a form, a page or a person. An AI agent finishes the task, such as a transfer or a domestic wire, and the customer confirms the exact details before money moves. If it cannot complete the transaction, if the bank cannot set policy outside the model, or if there is no full audit trail, treat it as a chatbot.
Is it safe to let AI move money? What stops it from sending the wrong amount?
The agent can request a payment. It never touches payment rails directly. A separate rules layer checks hard limits, and the customer confirms the exact amount, accounts and payee. A payment over a hard limit is declined. Nothing overrides that limit. Math is done by a calculator. See How to let AI agents move money safely and agentic payments.
How do AI agents handle OFAC and sanctions screening?
A new payee is screened for sanctions (OFAC) and account risk before that payee can be paid. If screening is unavailable, the payee is not added.
Do we need to replace our core system?
The agent works through the digital banking and the core the bank already runs. Customers stay signed in there, under the bank name. Replacing the core is not part of the setup. The same is true for credit unions, covered in agentic AI for credit unions.
What happens if the AI makes a mistake or someone tries to trick it?
The customer still has to confirm the exact payment, and a hard limit still declines an amount over the cap. Hidden or tricky instructions in a message or a document are blocked or treated as plain data, and flagged for the bank. Every step is logged. Examples are on AI banking safety.
What tasks should still go to a human?
Loans, mortgages, credit cards, disputes, lost or stolen cards, fraud freezes, opening or closing accounts, contact changes and international wires go to a person. The agent does not give investment, tax or legal advice. Internal agents for staff can help an employee find a procedure. They do not replace that employee, and they do not make a credit decision.
Can a small bank or credit union afford and run this?
A smaller bank or credit union can run an agent without a large AI engineering team. The agent sits in the digital banking the institution already has. The institution keeps the limits, the off switch and the audit trail, and staff can try it in a sandbox first. This page does not quote a price. See AI for community banks.