On 13 November 2026 OpenAI announced ChatGPT Wallet, built in partnership with American Express, Chase, Mastercard and Visa. Agents pay with a payment token in place of the user's real card details. Before each purchase the user checks the items, the total and the card, then approves it with a security check such as Face ID or a fingerprint. It starts with ChatGPT Pro subscribers in the US with supported US-issued cards.

On 14 November Anne Boden, who founded Starling Bank, set out the thinking behind Lockwood, a UK current account for people and small businesses that "lets your AI agent pay for things safely, within limits you set". Everyday payments inside the limit go through on their own; in the site's words, "Anything bigger waits for a tap from you." Lockwood is not open yet: there is a waitlist, and its site states that Lockwood is not a bank and is not connected with Starling Bank.

How much may an AI agent do with a customer's money without asking first? The two designs answer that differently. ChatGPT Wallet asks every time, for now. OpenAI says it expects that, over time, users will be able to hand purchases to their agent entirely by setting clear parameters, such as spending limits, purchase categories or time periods, and approving such purchases in advance. Lockwood's design starts there: the agent pays on its own within a standing permission, which Boden calls a mandate, and the customer is asked only above the limit.

That model is harder for a bank to support. With no one approving each payment, the bank has to check every payment against what the customer allowed.

In The right time to start again, Boden argues that starting again "is the only way to build a bank equal to the agentic age". She accepts that retrofitting an existing bank for agents "is possible", but calls it "a substantial undertaking". She also names what it involves: a bank must be able to "identify each agent, understand who it represents, enforce a bounded mandate and retain evidence of every decision".

Even for a bank with no plan to start again, that list works as a specification. Her pieces and Lockwood's site add two more items: an approval step above the limit, and one place to see and manage every agent.

A mandate for each agent

ChatGPT Wallet and the card networks behind it deal with how an agent pays without seeing the card number, and with who confirms the payment it makes. That still leaves open what the agent is allowed to do.

In AI agents need a different type of bank, Boden points out that an agent "may remain within its financial limit while acting completely outside the customer's intentions". An amount limit says nothing about what the money is for.

So the first new object in the system is a mandate for each agent, with the customer setting "the purpose, frequency and duration" of its spending. A mandate holds:

  • which agent it belongs to, and which provider runs that agent;
  • what the agent may pay for, and which payees or merchants it may use;
  • an amount per payment, and an amount and a number of payments per period;
  • a start date and an end date;
  • what the agent may never do, such as add a new payee or change its own limit.

The UK Open Banking standard already has something close. Its variable recurring payment consents carry control parameters under which the payments must operate, such as a maximum amount per payment and limits per period, and a consent can be given start and end dates. An agent mandate needs the same kind of controls, and it also names the agent, the provider that runs it and what the agent may never do.

The mandate has to be machine-readable and checked on every request. In Unravelling the new economics of attacking a bank, Boden describes an agent's request that "would be checked against a machine-readable mandate before it could reach the bank's critical systems". She gives an example there: "A shopping agent might, say, be permitted to spend £20 a week with approved supermarkets." For an engineering team, the check means a policy service in front of the payment flow that every agent request must pass through.

Every agent needs its own identity

To enforce a mandate, the bank must know which agent is asking. In Boden's design, an agent has to "establish its identity, the provider behind it, the customer it represents, and the authority it has been given".

That rules out the shortcut of an agent using the customer's login or card details. Each agent needs its own narrowly scoped credential, issued by the bank and tied to one mandate. In our view, this starts in the banking app, where the customer adds the agent and sets its mandate, and the bank issues the credential.

The bank can then tie every request to one agent and act on that agent alone. As Boden points out, a stolen agent credential could be disabled without cancelling the customer's card. The customer's other agents can keep working.

Above the limit: a request the customer can judge

An example screen on Lockwood's site shows the step-up, where a payment above the agent's limit goes to the customer for approval. An agent wants to pay £642 to renew home insurance against a £500 limit. The screen reads "Your AI wants to pay more than its limit", carries the agent's note that the price is £118 less than the renewal quote, shows Lockwood's risk rating, and offers two buttons: "Approve" and "Not this one".

For a bank app, we would make three design choices:

  • The request arrives with context: payee, amount, how far it goes over the mandate, and the agent's reason. A bare request to approve £642 invites either blind approval or reflex refusal.
  • The bank's risk rating sits on the same screen, so the customer can weigh the agent's reason against the bank's view of the payment.
  • The approval is strong authentication bound to that one payment, with a timeout and a defined outcome, such as a decline, when the customer does not answer.

ChatGPT Wallet asks for this kind of approval on every purchase. A bank can offer that in the same flow, because a mandate with a limit of zero sends every payment to the customer.

Designing agent permissions into your app?

Show us your app's current consent and payment flows, and we will map where the mandate check, agent credentials and step-up approval would sit.

Talk to our team

One place to see every agent, and to switch one off

Boden argues that the customer needs "one control centre showing every agent with access to their finances", and that each agent's authority "should be individually and immediately revocable".

In the app, that is a new section: a list of agents, each with its mandate, recent activity, pending requests and an off switch. The off switch is the harder part to build. Revoking an agent has to reach every place its authority is held (its credential, any tokens issued to it, payments it has scheduled, open banking consents) and take effect at once, without waiting for the next batch run.

An evidence log for disputes

Boden calls for "a reliable record of what the customer originally authorised, the agent requested", the limits applied and when a payment was refused or referred back.

She asks "what happens when an AI agent has genuine permission to use your money but the wrong decision is made?" and calls it "one of the biggest unanswered questions in agentic commerce".

She leaves that question open, and notes that resolving mistakes "becomes extremely difficult without evidence". The log should tie each payment to the mandate version in force, the agent's request and stated reason, the decision and the rule behind it, and any approval by the customer. Customers should be able to read it in the app, and the disputes team should be able to rely on it. The refund side of the question is in who carries the loss when an agent buys the wrong thing.

Retrofit or rebuild

Boden's case for starting again rests on architecture. Existing banks, she writes, "were built on the assumption that a human is the principal user". In her piece on attacks she adds that "the challenge lies in defending every part of an estate that has grown over decades", where employees, suppliers, payment companies and other banks "use a multitude of different entrances". Lockwood's answer is a single "window" through which every agent interaction passes.

Our view, from building banking apps and core integrations, is that all five parts can be built around an existing core banking system, with most of the work in an integration layer in front of it. Boden accepts that existing neobanks can add agent features, "just as traditional banks eventually produced good mobile apps", while arguing that supporting agents properly "means more than adding another interface". The layer we mean plays the part of her window, "one consistent control layer" for every agent interaction: it holds the mandates, issues the agent credentials, runs the policy check and keeps the log. Her word "eventually" is still a fair warning about pace.

For an established bank, the hard part is making every route an agent could take pass through that layer: cards, open banking APIs and the app. If the mandate check covers card payments and misses open banking payments, an agent can go round its limits by paying through open banking.

Where to start

We would do the work in this order:

  1. Model the mandate as versioned data: agent and provider, purpose, payees, amount per payment and per period, dates, prohibited actions.
  2. Give each agent its own credential that the bank can disable at once, and send every agent request through one policy check.
  3. Log every decision from the first payment, in a form the disputes team can use.
  4. Build the step-up approval with context and the bank's risk rating. Until it is live, decline anything above the mandate.
  5. Add the control centre, and make revocation reach every route at once.

None of this has to wait for Lockwood to open or for the liability question to be settled. With these parts in place, a bank can choose which agents to accept and on what terms.

Planning the layer in front of your core?

Tell us which routes an agent could use to reach your customers' money, and we will show where one policy check can cover cards, open banking and the app.

Talk to our team
Frequently asked questions
What does a bank need to build so an AI agent can pay from a customer's account?

Five parts, in the order WislaCode would build them: a mandate for each agent, kept as versioned data; a credential for each agent that the bank can disable at once, with every agent request sent through one policy check; a log of every decision from the first payment; a step-up approval for payments above the mandate; and one control centre where the customer sees every agent and can switch any of them off on every route at once.

What is ChatGPT Wallet?

A wallet in ChatGPT that lets AI agents pay without access to the user's real card details: the card network supplies a payment token in their place, and the user approves each purchase with a security check such as Face ID or a fingerprint. OpenAI announced it on 13 November 2026, built in partnership with American Express, Chase, Mastercard and Visa. It is available first to ChatGPT Pro subscribers in the US with supported US-issued cards.

What is Lockwood?

A UK current account for people and small businesses that Starling Bank founder Anne Boden is building. It is designed to let an AI agent pay within limits the customer sets, and anything bigger waits for the customer's approval. As of 17 November 2026 it is not open: there is a waitlist, and its site states that Lockwood is not a bank and is not connected with Starling Bank.

What is a mandate for an AI agent?

A standing permission the customer gives one agent: what it may pay for, which payees it may use, how much per payment, how much and how often per period, from when until when, and what it may never do. The bank checks every payment the agent requests against it.

How does an agent mandate differ from a variable recurring payment consent?

A variable recurring payment consent under the UK Open Banking standard already carries control parameters under which the payments must operate, such as a maximum amount per payment, limits per period and start and end dates. An agent mandate uses the same kind of controls and is written for one agent: it names the agent and the provider that runs it, and it lists what the agent may never do, such as add a new payee or change its own limit.

Can an existing bank support AI agent payments without replacing its core?

In WislaCode's view, yes. The mandate check, agent credentials, step-up approval, control centre and evidence log can be built around an existing core banking system, mostly in a layer in front of it. The hard part is making every route an agent could use pass through that layer: cards, open banking APIs and the app.