The Agent Gateway Is Banking's New Front Door
Before customers send AI agents to do their banking, every bank needs a control point that knows who is acting, under which mandate and within which limits.
For most of banking’s history, the bank knew who was on the other side of the counter. A customer in the hall, a teller at the window, a merchant’s card terminal: each request came from a person or a machine the bank had issued. That certainty is about to end.
A growing share of requests will come from software acting for someone else. A customer’s AI assistant will ask for a rate, move savings to a better account, or pay a bill, all under a mandate the customer gave it. The request is real and the money is real, but the actor is a delegate.
Three questions at the door
An agent gateway is where those delegates enter the bank. Before anything reaches the bank’s systems, it answers three questions.
Who is acting? A person, or an agent, and on whose behalf. Identity now has two layers: the principal and the delegate.
Under which mandate? What the principal authorised the agent to do, up to what amount, until when. The mandate is the new unit of consent.
Within which limits? Whether the request fits the bank’s own rules for that customer and that agent: velocity, exposure, sanctions, risk.
Every answer is signed and logged. When something goes wrong, the bank can show exactly who authorised what.
The gateway is the bank boundary: outside agents enter through it before the bank's own autonomous agents, product engines and ledger act.
Why it cannot sit on a batch core
The gateway has to answer in milliseconds. An agent comparing rates across a dozen banks will not wait for an overnight batch to post a balance. That makes the gateway the clearest reason to leave the packaged core: it needs a real-time ledger beneath it, and services that do the right thing when an agent retries.
Banks that cannot verify an agent’s mandate in real time will simply not be chosen by those agents.
The payment industry is already building the standards. Google’s Agent Payments Protocol, published in September 2025 with more than 60 supporters including Mastercard, American Express and PayPal, uses signed mandates; Visa launched its Trusted Agent Protocol the same month.
What boards should ask
Who owns the mandate registry? Who can switch off a misbehaving agent, and how fast? Which decisions may an autonomous agent take alone, and which need a person? These are governance questions before they are technology questions, and the banks that answer them first will be the ones agents prefer.