How AI routes a card payment, and how to test it before you

The same card payment can often travel more than one route. Behind a checkout there may be several acquirers, the companies that process card payments for a merchant, and each one has its own fees and its own approval habits. Deciding which route a payment takes is called routing. This article explains how AI routing works and how to test it safely.

Why more than one route exists

A merchant may connect to several processing routes for redundancy, for coverage of different countries and currencies, or for cost. For any single payment, more than one route may be able to handle it.

They do not behave identically. A card's issuing bank may say yes more often to one route than another, depending on the card, the country and the amount. Routes also differ in cost, and the cheapest route is not always the one that gets approved. A cheap decline still means no sale.

How fixed routing fails

Older setups send everything down one fixed route, or follow a rule someone wrote months ago: "all euro cards to route A". That works until route A has a bad day. If a route is down, slow or disliked by a particular bank, the payment fails and the customer leaves, even though another route might have worked.

What AI routing does

AI routing tries to answer one question for every payment: which route has the best chance of approval, at a fair cost?

It reads the payment: card type, country, amount, and results of similar past payments.
For each route it estimates the chance the bank says yes. That estimate comes from past outcomes.
It weighs total cost, including fees. A route with slightly lower approval may still cost much less.
It sends the payment down the route with the best balance, in a fraction of a second.

If a route fails for a technical reason, the system can retry on another route. Every retry has a cost, in fees and in the time the customer waits.

What the promises leave out

Providers sometimes quote large approval uplifts from smart routing. Those figures come from the providers themselves, use different baselines, and do not tell you what will happen to your own mix of cards and countries. No independent benchmark we found for 2026 settles the question. Treat any promised number as a hypothesis to test, not a result.

How to test it

Split a small share of traffic. Route, say, a small fraction through the AI-selected path and the rest as before.
Compare like with like. Same period, same countries, same card types.
Look at the whole bill. Approval rate, total fees, number of retries and customer wait time.
Widen only after it holds. If the gain disappears once retries and fees are counted, it was not a gain.

What to keep a person responsible for

Rules about which countries or card types are allowed on which route often have contractual or compliance reasons. Automation can choose among options a person has approved. It should not invent new ones.

Where we land

Routing is a good place to use AI because the decision is frequent, the outcome is measurable and a wrong choice is cheap to detect. It is also a place where hidden costs, such as retries, can quietly eat the gain.

The open question: should a merchant let the system reroute automatically, or approve every routing rule by hand? Automatic rerouting reacts faster when a route is struggling. Hand-approved rules are easier to audit and explain. Which matters more to you?

EPaySe is a payment gateway in development for high-risk and regulated businesses. High-risk and regulated businesses are reviewed and approved case by case.

Sources: