Can an AI agent pay a customer's bills?

Within limits, yes. An AI agent can pay a customer's regular bills to billers the customer has already approved, when the amount falls inside a range the customer set. A new biller, an unusual amount, payment details taken from an email, or choosing which bill to pay when money is short stays with the customer or a person.

What makes bills different from other payments?

Bills vary. The rent is the same each month; the energy bill is not, and the credit card bill can swing widely. A rule that says "pay my energy bill" is a mandate to pay an amount nobody knows yet. That is fine within a range, and risky without one.

Bills also arrive from outside. A bill is a document sent by someone else, and a bill can be forged. A fraudster who sends a convincing bill with new payment details is counting on whoever pays it not to check. For the agent, the bill is an external fact: useful for knowing what is due, never enough on its own to send money somewhere new.

What conditions decide it?

The agent may pay a bill when:

• the biller is one the customer added and verified in the bank's authenticated channel;
• the payment goes to the biller's details already on file, not details read from the bill;
• the amount sits within the range the customer set for that biller, such as up to one and a half times the average of recent bills;
• the account will still cover other scheduled payments after this one;
• no fraud or difficulty flag applies.

When the amount is above the range, the agent asks the customer to confirm. When the payment details on the bill differ from those on file, it stops and tells the customer.

What may the agent do, and what stays with the customer or a person?

• Read bills and due dates the customer shares or that arrive through the bank's bill service, at depth 1.
• Pay a verified biller within the customer's range, at depth 4 within the customer's limits.
• Ask for confirmation on amounts above the range, then pay once the customer confirms.
• Stop and alert when the bill shows new payment details for a known biller.
• Remind, not choose, when there is not enough money for every bill. Which to pay first is the customer's decision; the agent can show what each missed payment would mean using approved wording, including which are priority bills such as rent and energy.
• Refer customers who repeatedly cannot cover essential bills to the bank's support team.
• Never add a new biller from a bill or an email.

For bills already in arrears, a promise to pay is a different process; see handling promises to pay with an AI agent.

An example: a winter energy bill

Take a customer who set the agent to pay their energy, water and phone bills from their current account. The energy range is up to 180, based on an average of about 120.

In January the energy bill is 260. The agent does not pay it. It messages the customer showing the bill, the usual range and the balance after payment. The customer confirms, and the agent pays.

In February the phone company "sends" a bill by email saying it has changed banks and showing new details. The agent pays nothing, tells the customer the details differ from the ones on file, and suggests checking with the phone company through its usual contact route. The phone bill is paid to the original details once the customer confirms the real amount.

What controls make it safe?

• Payee details from the system of record. The agent pays the details stored when the customer verified the biller; details on a bill are external and cannot authorise a payment.
• Ranges set by the customer. Each biller has an upper bound the customer chose; above it, the payment is held for confirmation.
• Balance checks first. The agent checks other scheduled payments before paying.
• No new billers. Adding a payee is outside the agent's capability for bill payments.
• Each payment filed. The verdict records the bill, the range, the balance check and the deciding bound.

A business paying its suppliers faces the same problem at larger scale, covered in whether an AI agent can pay supplier invoices. Instalments on buy now pay later plans are bills too, and the decision to take one on is covered in whether an AI agent can approve a buy now pay later purchase. The general limits are in AI agents in banking: what they may and may not do.

Where Pryme fits

In Pryme, a bill reaches the agent with external provenance, and the biller's payment details arrive sealed from the core banking system. When the two disagree, the sealed details win and the payment is refused. The customer's range for each biller is their own bound in the minimum chain; a bill above it is held for the customer, and the verdict names that bound.