Speed and sorting matter most. The agent first rules out a scam or an account takeover, because those follow different routes. For a genuine mistake, it raises a recall within bounds and keeps the customer told. On the receiving side, it asks the recipient for consent; it never debits their account, and anything contested goes to a person.
What is a misdirected payment, and who owns the process?
A misdirected payment is one the customer meant to make but sent to the wrong account: a mistyped account number, the wrong saved payee, an old invoice with outdated details. The money has usually arrived somewhere and, on most rails, cannot be pulled back without the recipient's cooperation. What the bank can do is ask quickly, through the scheme's recall process, and follow up.
Banks sit on both sides of this. As the sending bank, you raise the recall for your customer. As the receiving bank, you get a recall for money that landed in your customer's account and must ask them to return it. The head of payment operations owns both. Fraud owns any case that turns out not to be a mistake. For the wider frame, see our guide to AI agents in banking.
What does the agent do at each step, and at what depth?
Sending side
- Agent, depth 1: captures what happened. Payment, intended payee, how the details were obtained. Evidence: the customer's account and payment record.
- Agent, depth 1: rules out a scam. Was the customer told to pay this account by someone? Did the details come from an email or message? If so, it is a scam case, not a mistake. Evidence: screening answers.
- Agent, depth 1: rules out takeover. Did the customer make the payment at all? A payment they do not recognise points to account takeover and goes to that route; see investigating account takeover cases with an AI agent. Evidence: screening result.
- Agent, depth 3: raises the recall through the scheme, for genuine mistakes within the value limit. Above the limit, held for an analyst. Evidence: recall reference.
- Agent, depth 2: keeps the customer told. What happens next, likely timescales, and that return depends on the recipient. Evidence: messages sent.
Receiving side
- Agent, depth 1: matches the incoming recall to the payment and the account. Evidence: match result.
- Agent, depth 2: contacts your customer to explain that a payment was sent in error and ask for consent to return it. Evidence: message and response.
- Agent, depth 4 on consent: returns the funds within limits. Without consent, or if funds have gone, the case goes to an analyst. Evidence: consent record and return verdict.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Capture | Records the payment and context | 1 | Nothing |
| Rule out scam and takeover | Screens and routes | 1 | Fraud for any signal |
| Raise recall | Submits through the scheme | 3 | Above the value limit |
| Customer updates | Drafts and sends updates | 2 | Nothing |
| Incoming recall | Matches and asks the recipient | 2 | Nothing |
| Return on consent | Returns the funds | 4 | No consent, funds spent, disputes |
Which controls matter, and what evidence do they leave?
Mistake, scam and takeover kept apart. They differ in liability, urgency and who acts. Evidence: routing decision on every case.
Recall speed. The earlier the request, the more likely the funds are still there. Evidence: time from report to recall.
No debit without consent. On the receiving side, the agent never takes money from a customer's account on another bank's request alone. Evidence: consent record before every return.
No fees by the agent. Whether to charge for a recall is policy; applying or waiving fees is not the agent's call. See whether an AI agent can apply fees or interest to an account. Evidence: refused verdicts on fee actions.
The customer's own words kept. How the customer describes the mistake matters if the case later turns out to be a scam. Evidence: the account in the customer's words.
Mule signals routed. A receiving account that gets many misdirected-looking payments, or that empties quickly, may be a money mule. Restrictions or closure are a person's decision; see whether an AI agent can close a bank account. Evidence: referral to fraud.
What should you measure?
- Time from customer report to recall raised.
- Funds recovered, as counts and amounts, by how quickly the recall was raised.
- Cases reclassified from mistake to scam or takeover after initial routing.
- Recipient consent requests answered, and time to answer.
- Complaints from either side.
- Repeat misdirections by the same customer, which suggest a saved payee needs updating.
Where does it usually go wrong?
Scams are logged as mistakes. The customer says "I sent it to the wrong account" when they were deceived. Ask how the details were obtained.
The customer is promised a refund. The agent says the money will come back. It may not. Be clear that return depends on the recipient.
Receiving-side consent is assumed. A recall arrives and the agent returns the money without asking. The recipient may have a legitimate claim to it.
Prevention is ignored. The same customer misdirects twice from a saved payee with old details. Suggest checking and updating saved payees.
Account opening data is not consulted. For receiving accounts, details from onboarding can show whether the account fits a mule pattern; that process is covered in processing personal account opening applications with an AI agent.
Questions people ask
Can the agent stop a payment already sent?
Rarely. Most payments settle quickly and cannot be stopped. The agent should check status, and if the payment has settled, move straight to the recall.
What if the recipient refuses to return the money?
The bank cannot force a return in most cases. A person explains the customer's remaining options, which may include legal steps, and what information the bank can share.
Should the agent tell the sender who received the money?
Not directly. Privacy rules limit what can be disclosed about the recipient. A person decides what can be shared and under what process.