> Smart routing isn't mysticism. What dimensions to route by, how to start with rule-based routing, and whether AI routing is worth chasing.

[← Blog](/en/blog)

🧠

Orchestration

# Configuring Smart Routing: By Region, Card Type, Cost, Success Rate

2026-06-01

“Smart routing” gets dressed up as mysticism by orchestration vendors. Unpacked, it’s one thing: **for each transaction, automatically pick the most suitable channel**. Here’s what to route by, how to start simple, and whether AI routing is worth chasing.

## What dimensions to route by

Routing is fundamentally “pick a channel by transaction characteristics.” Common dimensions:

-   **Region**: local transactions authorize better than cross-border, so route main markets through local acquiring;
-   **Card type / network**: some channels approve Visa/Mastercard/Amex differently;
-   **Cost**: channels differ on fees, cross-border charges, FX loss — among channels that would all approve, prefer the cheaper one;
-   **Historical success rate**: give more volume to the route approving best lately, downweight one that wobbles;
-   **Risk**: send high-risk transactions through channels with stronger 3DS / risk controls.

## Start with rule-based routing — don’t chase AI on day one

The pragmatic order:

1.  **Rule-based routing first** — “Brazil → local acquiring,” “large amounts → 3DS channel,” “A wobbled → switch to B.” Transparent, controllable, instantly effective;
2.  **Layer in success-rate feedback** — let routing order adjust to each channel’s real-time performance;
3.  **Then consider whether to add a predictive model.**

For many teams, the first two steps capture most of the gain.

## The honest truth about “AI routing”

Everyone touts “AI smart routing.” Unpacked, Primer’s UpliftAI is essentially a binary classifier predicting whether a channel will authorize, with an official line of “**up to** 5%” authorization lift (a pilot ceiling, not a guarantee). The conclusion is clear: AI routing does help, but **the industry baseline isn’t high**, and marketing often exceeds proven effect.

So don’t put the cart before the horse: nailing the fundamentals (local acquiring, cascade, code-based retries, card updates, rule routing) beats chasing a flashy AI label. AI as icing is fine; as a lifeline, no.

## How to do it

1.  Run 2–3 channels (including local) in main markets so routing has options;
2.  Get rule routing running, then layer in success-rate feedback;
3.  Dashboard each route’s approval rate, cost, and latency so routing “follows the data”;
4.  Treat AI routing as optional enhancement — validate real lift with A/B before scaling.

> Smart routing and success-rate feedback are handled by Flow orchestration (coming soon), with AI as optional enhancement; what’s live today is Vault — card tokens in your name, one token to any PSP. [Book a demo](/en/) and we’ll configure a routing strategy for your channel mix.