Starling founder Anne Boden has set out plans for Lockwood, a UK current account designed to let authorised artificial intelligence agents make payments within limits chosen by their customers.
The proposed service is aimed at individuals and small businesses. Lockwood is not open, and joining its waiting list does not create an account. The venture is not a bank, and the permissions and protections that would apply to any eventual service remain to be determined.
Boden founded Starling Bank in 2014. Lockwood is a separate venture, built around the idea that customers may increasingly want software to carry out financial tasks rather than simply ask a person to complete each payment.
What an agent could be allowed to do
An AI agent is software that can pursue a task on a user’s behalf. In the proposed account, a customer could choose which assistants to trust and give them authority to perform jobs such as paying bills or buying goods. Everyday transactions within the permissions could proceed without a separate approval for each payment; requests outside them could be referred back to the customer.
The permissions are intended to be more specific than handing over an account password. They could restrict an agent by the account it may use, the purpose of its spending and the frequency or duration of its authority. A customer would be able to withdraw an individual agent’s access without necessarily turning off every other assistant.
For a small business, that might mean allowing one assistant to handle a defined set of purchasing tasks while another works on bills. The point of the boundary would be to give each agent only the authority needed for its role, rather than treating all software connected to an account as equally trusted.
Lockwood’s design also envisages a record of what a customer authorised, what an agent requested and whether the request was allowed, refused or sent back for review. Such a record could help explain a disputed transaction, although the published product description does not settle how every mistake or dispute would be handled.
Permission is different from intention
A payment can satisfy a spending limit and still be the wrong purchase. An assistant asked to order a small quantity of an item might misunderstand the instruction and buy too much, for example. The seller could fulfil the order and the payment could be valid even though the result was not what the customer wanted.
That distinction makes the proposed controls more complicated than setting a maximum amount. A system must identify which agent is acting and for whom, but an account provider may not see all the choices made before a payment request reaches it. How those decisions are recorded, challenged and corrected would matter as much as the payment limit itself.
Lockwood also plans to use AI in building and operating its own technology, with people overseeing the work. Proposed uses include helping to write and test software and monitoring account activity. These are design aims, not proof that a finished system has been independently tested or can prevent every error or fraud.
Why the account’s status matters
A current account’s name does not by itself establish how money held in it is protected. Eligible deposits with an authorised UK bank may qualify for Financial Services Compensation Scheme protection if the bank fails. Authorised electronic-money and payment firms generally protect customer funds through safeguarding arrangements instead of direct deposit protection against the failure of the payment firm.
The distinction is relevant to any future product, but it does not establish what structure Lockwood will ultimately use. Its authorisation, provider arrangements and customer terms will need to be clear before the protections attached to an actual account can be assessed.
For now, the venture presents a prospective way to delegate routine payments while retaining boundaries around access. Whether it works as intended will depend on the service that is eventually authorised, the controls customers can use and the arrangements for resolving problems when an agent acts within its technical permissions but outside a customer’s wishes.