Every merchant team we talk to has the same first instinct when approval rates slip: add another acquirer. It almost never works on its own. The acquirers aren't the bottleneck — the routing logic on top of them is.
What "smart routing" usually means in 2026
In most stacks "smart routing" is still a set of static rules: if BIN starts with X, send to acquirer A; if currency is Y, send to acquirer B; if amount exceeds Z, fall back. It's deterministic, auditable, and wrong about 18 hours out of every 24, because issuer behaviour, scheme rules, and fraud posture shift constantly.
What it should mean:
- A model trained on the merchant's own historical authorisation outcomes, not generic benchmarks.
- A retry policy that knows the reason a transaction failed and picks a different path — not just "try the next acquirer".
- A control plane the merchant can actually inspect: why this BIN went to this acquirer, what the predicted approval was, what the realised outcome was.
The data the router actually needs
The routers we see win on three signals most stacks don't surface cleanly:
- Issuer-level approval velocity. Approval rates per issuer can swing ±15% in a single day around scheme updates or fraud waves. Stale weekly averages will steer you into the wrong path.
- Scheme-aware soft declines. "Do not honor" from one scheme means something very different than from another. Treating them identically is how you burn an acquirer relationship by retrying into the same wall.
- Cost-to-approval, not cost-to-attempt. The cheapest rail per attempt is rarely the cheapest rail per successful authorisation once retries and interchange are loaded in.
What we changed
We rebuilt the router around three primitives:
- Verdicts, not rules. Every authorisation candidate gets a model-scored verdict with a confidence interval. Operators tune thresholds, not if/else trees.
- Replay, not guesswork. Every routing decision is logged with the features that produced it, and we can replay yesterday's traffic against today's model before promoting it.
- Per-merchant policy, not one-size-fits-all. A subscription business in EU markets and a marketplace in LATAM share zero routing assumptions — and now they don't have to.
The result
Across pilots in Q1 2026 we saw an average +3.4% lift in net authorised volume versus the previous static-rule routing setup, with the largest gains in markets where issuer behaviour was changing fastest. No new acquirer needed.
The lesson is the boring one: the bottleneck is almost always one layer above the place you're tempted to look.