Skip to main content
    3 min read

    Self-service and AI agents: preparing digital banking for customers who never visit the site

    When an agent acts for the customer, the interface stops being the screen and becomes the data contract. A bank whose operations can only be completed by looking at a screen will be invisible to the assistant managing its customer's money.

    Written for: Leaders in digital banking, product, architecture, risk and compliance within financial institutions.

    Abstract illustration of a node grid with a highlighted route representing operations executed by automated agents.

    When an agent acts for the customer, the interface stops being the screen and becomes the data contract.

    A change of counterpart, not of channel

    For twenty years digital banking has designed for a person who looks. The working assumption was stable: someone opens the app, reads, decides and taps. Everything was built on it — the visual hierarchy, the copy, the warnings, the placement of buttons.

    That assumption is weakening. Forrester estimates that by mid-2026 human visits to bank websites will fall by roughly 20% while traffic initiated by personal assistants will grow by close to 40%. The material change is not fewer sessions: it is that a growing share of requests arrives from a counterpart that does not read, does not interpret visual context and does not tolerate ambiguity. The question is no longer whether the interface is clear, but whether the operation is executable without it.

    Four fronts to resolve before the interface

    Operations expressible without a screen

    It is worth inventorying what a customer can do and classifying it by visual dependency. Checking a balance is usually trivial. Changing a limit, disputing a charge or moving a direct debit normally requires steps that exist only inside a graphical flow. Anything that can only be completed by looking sits beyond an agent's reach.

    Delegated identity and permission scope

    The problem is not authenticating the agent: it is representing that it acts on someone's behalf, within which limits and for how long. An institution must be able to answer who authorised it, which operations that authorisation covers, when it expires and how it is revoked. Without that model, the only safe option is to block — and blocking means disappearing from the channel.

    Evidence and traceability of the instruction

    In banking, an operation does not end when it executes, but when it can be explained. If an agent initiates a transfer, the institution must retain what it was asked, under what mandate, what information it received and what it confirmed. That evidence is what allows a later complaint to be resolved without relying on the customer's word or the assistant provider's.

    Deliberate limits

    Not every operation should be automated simply because it technically can be. Contracting an investment product, raising a credit limit or changing contact details are decisions with asymmetric consequences. Explicitly defining what requires human presence is a product decision, not a technical constraint.

    How to approach it in phases

    1. Inventory operations by visual dependency. Separate what is readable, what is executable and what exists only on screen today.

    2. Define the mandate model. Scope, expiry, revocation and a record of the delegated authorisation.

    3. Expose low-risk operations first. Enquiries, statuses, documents and confirmations.

    4. Instrument evidence from the outset. Log instruction, mandate and response before widening scope.

    5. Publish the limits. State clearly what the institution will not do automatically, and why.

    What to measure in a channel without sessions

    • Executable coverage: share of operations completable without visual interaction.

    • Delegated completion rate: agent requests completed against initiated.

    • Mandate revocations: frequency and reason, as a trust signal.

    • Complaints tied to delegated operations: per thousand operations, compared with the human channel.

    • Time to evidence: how long the institution takes to reconstruct a disputed operation.

    Mistakes already being made

    • Treating the matter as a project to build an in-house conversational assistant.

    • Authenticating the agent as if it were the customer, without representing the mandate.

    • Opening sensitive operations before traceability exists.

    • Assuming the assistant provider will carry responsibility for the instruction.

    When the internal conversation revolves around which language model to use, the institution has not yet reached the real problem, which is one of permissions, evidence and scope.

    Conclusion

    Digital banking prepared itself for a customer who looks at a screen. The customer arriving now, increasingly often, delegates. Adapting does not mean adding an assistant to the existing channel, but ensuring every material operation can be initiated, authorised, executed and audited without anyone seeing it.

    Consumer Services Hub helps financial institutions redesign operations, identity and control for channels where the customer is not always the one tapping.

    Consumer Services Hub - Strategic ecommerce consultancy for B2C service companies

    consumerserviceshub.com

    If you enjoyed, share this article:

    Optimise digital acquisition

    Application funnels, onboarding and compliance without losing conversion.

    Talk to us

    Whether you're launching a new digital channel, optimizing an existing one, or unlocking the next stage of growth — we're ready to help.

    Financiado por la Unión Europea - NextGenerationEUGobierno de España - Ministerio para la Transformación DigitalRed.esPlan de Recuperación, Transformación y ResilienciaKit Digital