Beyond the app: designing agentic financial experiences

An agentic experience is one where software does not just show you something — it does something for you. It moves money between your accounts to avoid an overdraft. It cancels a subscription you stopped using. It negotiates a payment date with a biller.

The technology to do this largely exists. The design problem is the part nobody has solved, because it is not really a design problem about screens. It is a design problem about trust.

Consent is not a checkbox

The instinct is to gate agency behind a one-time permission at setup. That fails in both directions: it asks customers to authorise something abstract before they understand it, and it gives the institution a mandate far broader than anyone intended.

Consent for an agent needs to be granular, legible and revocable in the moment. A customer should be able to say yes to moving up to a certain amount between their own accounts, and no to anything that touches a third party — and change their mind on a Tuesday without hunting through settings.

Design for the moment it goes wrong

Every agentic feature will occasionally do the wrong thing. The measure of the design is not how it behaves when it is right; it is how quickly and calmly a customer can understand and undo a mistake.

• Show the reasoning, not just the result. "Moved £200 from Savings because Rent of £950 was due tomorrow and Current held £780" is auditable in a way that "We helped you avoid a fee" is not.
• Make reversal a first-class action, sitting beside the record of what happened rather than buried in support.
• Keep a plain-language ledger of every action taken on the customer's behalf, in one place, exportable.

Customers do not need an agent that is never wrong. They need one whose mistakes are visible, explicable and cheap to reverse.

The interface shrinks, the stakes grow

As agency increases, the visible surface tends to get smaller — fewer screens, less navigation, more happening quietly. That is the point, and it is also the risk. Less visible software is harder for a customer to form an accurate mental model of, and a wrong mental model is where trust breaks.

The teams doing this well spend proportionally more of their design effort on the explanation layer than on the action layer. The action is a transfer. The explanation is the product.

A practical starting point

Begin with reversible, self-contained actions inside a customer's own money — sweeps between their accounts, rounding up, moving a buffer. No third parties, no irreversible outcomes, low regulatory surface. Get the consent, explanation and reversal patterns right where the stakes are manageable, then extend them outward.