Can an AI agent move money between company bank accounts?

It depends on how tightly "company accounts" is defined and whether the agent has earned depth 4. Moving money between your own accounts looks safe because the cash stays put. It is safe only when the destination list is fixed from the bank mandates and the amount sits within limits. Most teams keep a treasury approver on every transfer.

Why is an internal transfer still a payment?

An internal transfer moves money, and moving money is depth 4. The risk is not that cash leaves the group by design. It is that it leaves by mistake or by manipulation: an account added to the "internal" list that belongs to someone else, a transfer that drains an account needed for tomorrow's payroll, or an instruction that arrives by email claiming to be from the finance director.

The question of whether a finance agent should touch bank accounts at all is argued in should a finance agent have access to the company's bank accounts. This article assumes it has some access and asks where the limits fall.

What conditions would let an agent transfer on its own?

All of these, at once:

• A fixed destination list. Every permitted account is taken from the bank mandates and maintained by a person. The agent cannot add to it, and no instruction in a message can add to it.
• A standing rule. The transfer follows a written funding rule, such as "top up the payroll account to its required balance two days before pay day".
• Limits per transfer and per day. Set by the customer and lower than any person's limit.
• Earned depth 4. The agent has been certified on your own history of transfers, including the cases where a person declined one.
• Source of the instruction. The trigger is the agent's own reading of sealed balances and scheduled payments, not a request.

If any of these fails, the transfer is drafted and held.

What does the agent do, and where does the approver sign off?

An illustrative example

Take a care home operator with an operating account, a payroll account and a reserve account at the same bank. Its standing rule says the payroll account must hold the full payroll amount two working days before pay day.

Three days before pay day, the agent reads the payroll file total of 640,000 and the payroll account balance of 90,000. It drafts a transfer of 550,000 from the operating account, which holds enough without dropping below its own minimum. The agent is certified to depth 4 for transfers under the standing rule up to 600,000, so it makes the transfer and records it. The treasury analyst sees it on the next morning's list.

The same afternoon, an email arrives from what appears to be the finance director asking for 120,000 to go from the reserve account to "our new escrow account" for a property deal. The destination is not on the list. The agent refuses, records the request as external, and alerts the treasury manager, who confirms with the finance director by phone that no such request was sent.

What controls make it safe?

• List maintained outside the agent. Destination accounts change only through the bank mandate process.
• Limits enforced at the bank and the gateway. The agent's bank access itself carries the per-transfer limit.
• External requests cannot authorise. A message asking for a transfer starts a check, never a payment.
• Daily review. Every transfer the agent made is listed for a named person the next morning.
• Reconciliation of both sides. Each internal transfer should appear as an outflow and an inflow; the daily bank reconciliation confirms it did.

The reconciliation method is set out in how to reconcile bank accounts to the general ledger. Customer-facing transfers carry different rules, covered in whether an AI agent can move money between a customer's accounts. The wider finance controls are in AI agents for finance: books you can audit.