AI agents need a different type of bank

Giving an agent the means to pay is becoming possible. Giving it the right authority is much harder.

Experiments with Instinct and Muse mean there are a lot of pairs of socks turning up at my house. My purpose is to check how various existing payment options deal with agent orders. Imagine, though, you weren't looking at this from a building-new-technology point of view. You simply ask your household agent to top up your sock drawer. It duly finds a supplier, places an order, and pays with your stored card. Yet, a few hours later, 20 pairs of socks arrive when you wanted two.

It's an interesting situation. Nothing's been stolen, the card details you shared were valid and the merchant delivered precisely what it was asked to supply. Everything, including the payment system, worked perfectly, except what you intended didn't happen. This apparently trivial mistake highlights one of the biggest unanswered questions in agentic commerce; what happens when an AI agent has genuine permission to use your money but the wrong decision is made? The potential errors might not always be as harmless as leaving you with an overflowing sock drawer. An agent might mistakenly book a non-refundable hotel, renew an insurance policy with inadequate cover, pay an invoice twice, or move money out of an account shortly before the mortgage is due. It may remain within its financial limit while acting completely outside the customer's intentions. Giving an agent the means to pay is becoming possible. Giving it the right authority is much harder.

A payment credential is not permission

Let's look at the first part which is moving at pace, the means to pay. Card networks are developing ways for agents to use secure payment tokens. Digital wallets can issue virtual credentials without revealing the underlying card number. Open banking allows regulated providers to access account information and initiate certain payments with the customer's consent. These developments make it possible for an agent to act without being handed an unprotected card number, or requiring us to approve every individual purchase.

So far, so promising. What these developments do not solve is the issue of authority. A secure payment credential proves an approved mechanism was used, but it doesn't necessarily prove that the resulting decision reflected what the customer wanted.

To understand the issue better, it is helpful to think about it in terms of traditional banking which has long been based on a relatively simple sequence. A person decides to make a payment, instructs the bank, and authenticates the transaction. The bank then checks that the instruction is valid, the money is available and there are no signs of fraud. Agentic commerce changes that sequence. Instead of instructing a bank to make one payment, the customer gives an agent an objective: keep the fridge stocked, find a cheaper energy tariff, renew my insurance, or pay approved invoices when they're due. The agent will then select the merchant it will order from. Along the way, it will make decisions about price versus quality and delivery speed, all based on the customer's previous saved preferences. By the time the instruction reaches the bank, much of the consequential decision-making has already happened elsewhere. The bank may see a perfectly valid £20 transaction from an established retailer. What it may not see is whether or not the customer asked for two pairs of socks or twenty.

One customer, many agents

Where it becomes more complicated is when people begin using multiple agents. You may use one agent for everyday shopping, another for travel, one for bills and subscriptions, and another for savings. An SME might use separate agents for reconciling expenses, buying stock and managing cash flow. Those agents may be run by different companies, using different AI models, requiring very different levels of authority. A shopping agent may be trusted to spend £50 without interruption, whereas a travel agent may occasionally need permission to spend £2,000. An agent managing savings might be allowed to move money between approved accounts, but never to send it to a new payee. Consumers will need somewhere to see and control this activity. There will be decisions to make about which agents have access to their money and the parameters of what each one will be allowed to do. Some decisions will still require approval and some permissions might need to be renewed.

Agentic commerce requires more than an agent-compatible card. There needs to be a banking relationship designed around delegated authority, which establishes the identity of an agent and the provider behind it. With an agentic-enabled current account, customers would set a bounded mandate, permissions limited by account, and be able to dictate the purpose, frequency and duration of their agentic spending. These accounts would permit routine activity while escalating unusual or higher-risk decisions to the customer. The customer, whether a business or individual, also needs one control centre showing every agent with access to their finances. Their authority should be individually and immediately revocable. There should also be a reliable record of what the customer originally authorised, the agent requested, any limits applied and when a transaction is refused or referred back. Resolving any mistakes becomes extremely difficult without evidence.

Keeping the customer in control

An agent-native current account could become the independent control layer between the consumer and the growing number of AI providers. Existing banks will undoubtedly begin to add to some of the capabilities being discussed here. However, present-day banks were designed around people opening apps, examining transactions, and pressing buttons. Persistent machine-readable authority was not one of their founding assumptions. Neither was the need to distinguish a legitimate customer agent from a hostile agent imitating it. Designing around those requirements from the beginning creates the possibility of something materially different: not simply another bank with an AI assistant, but a bank capable of serving customers through whichever agent they choose.

More thinking

All our thinking

How can I join Lockwood?

Lockwood isn't open yet. Join the waitlist to hear about trials and early access.