• Taras Khodnevych, Product Manager at Corefy


• 26.10.2026 11:15 am
#payments

Ask a vendor what AI does for smart payment routing, and you will get a percentage. Ask how they measured that percentage, and the conversation usually gets shorter.

The most instructive public case I know comes from Stripe. After Twilio switched providers, its authorization rate rose 10% against the previous provider, and Twilio credits roughly 1% of that to Stripe's machine-learning retry tool. Stripe's write-up attributes the other gains to less glamorous work: standardized address verification, local acquiring, and automatic updates for stored cards.

That proportion matches what I see in payment routing setups from the product side. Machine learning in payments is real, and it does move approvals, but in most stacks it delivers the last increment rather than the first ten points. Below is how I separate the two, and the questions I would ask before paying for the second.

What is dynamic payment routing and what does it actually decide?

Dynamic payment routing chooses a route for each transaction when it is created, based on the transaction's attributes and current conditions, instead of sending all matching traffic down a fixed path. The inputs are your payment routing data: BIN, issuer country, card type, amount, currency, merchant category, time of day and, crucially, how each provider has responded to similar traffic before.

Two ways turn that data into a decision. A person writes rules, such as 'UK-issued debit over £100 goes to acquirer A, then B'. Or a model estimates the approval probability on each available route and picks the highest. Vendors call both intelligent payment routing, which is part of the confusion.

Either way, routing picks among routes you already have, in an order you can influence. It cannot create a local acquirer you never signed, and it cannot change how an issuer scores the transaction once it arrives. Keep that boundary in mind, and most AI routing claims become easier to read.

How to improve payment approval rates without AI

When an authorization rate optimization project works, the gains usually come from five places, roughly in this order of impact.

Route coverage

The biggest lever is having the right acquirer for the issuer in the first place. Domestic acquiring for a cardholder's home market typically outperforms a cross-border route for the same card, and no algorithm can route to a provider you are not connected to.

One of our clients in a high-risk vertical sat at a 20% payment conversion rate despite having several providers across regions. When our team analyzed the declines, a large share traced back to acquirer restrictions on specific countries and card types that nobody had mapped. Rebuilding the routing scheme around issuer country and user IP, plus a cascade, took conversion above 60%. Every rule in that scheme was written by a person.

Credentials

Network tokens change what the issuer sees, no matter where you route. Visa reports a 4.6% lift in authorization rates on tokenized card-not-present transactions compared with PAN. Account updater services do similar work for stored cards that have expired or been reissued. Both are plumbing, and both often outperform anything sold as AI payment optimization.

Decline handling

Most recoverable revenue sits in how you treat the first decline. A soft decline caused by a timeout or an acquirer's risk engine deserves another route; a closed account or a stolen-card code deserves nothing. As I wrote in the guide to cascading payments, the decline-code classifier comes before the cascade, not after it.

The card schemes also set hard limits. Both Visa and Mastercard cap how often a declined card can be retried, charge for attempts beyond those caps, and return codes that say when not to retry at all. A routing engine that ignores them recovers some approvals and pays fees on the rest.

Authentication

3D Secure and Strong Customer Authentication exemptions decide whether a legitimate payment reaches approval at all. Requesting authentication where the issuer expects it, and skipping it where an exemption applies, often shifts approval rates more than a route change. When a payment cascades, the authentication result must travel with it: forcing a second challenge on a new provider loses more customers than the cascade recovers.

Provider health and observability

Routes go down, merchant accounts hit turnover caps, a configuration gets disabled during a release. A surprising share of routing underperformance turns out to be a route that silently stopped being eligible. That is why end-to-end routing traceability was one of our main investments last year: before you optimize a decision, you need to see why each transaction went where it did.

What machine learning adds to smart payment routing

Once those five are in place, machine-learning payment routing optimization earns its keep in three specific ways:

• Granularity. Real-time routing decisions based on PSP approval history only work if the history is sliced finely enough: by BIN, amount band, time of day, and route. That produces thousands of segments whose performance drifts as issuers adjust their risk models. People cannot maintain rules at that resolution; a model can relearn it continuously. When Worldline launched AI-powered routing, it reported an uplift of over 2% for pilot merchants on top of the 3% achieved with rule-based routing. The size of the increment matters less than its framing: the model's contribution is measured against an optimized rule set, not against a static baseline.

• Risk-based payment routing. A model that scores transaction risk can send low-risk payments down frictionless routes and steer riskier ones to providers with stronger fraud screening or mandatory 3DS. The result is fewer false declines for legitimate cardholders without loosening controls across the board.

• Cost. Least-cost routing that loses approvals is a false saving. When vendors talk about AI optimizing payment routing for cost efficiency, it should mean weighing each route's approval probability against its fees, so the engine maximizes expected net revenue rather than either metric alone.

All three depend on conditions many stacks do not meet. The model needs enough volume per segment to learn something other than noise, and at least two viable routes for the traffic in question. It also needs to know why each payment failed, and that is harder than it sounds. Issuers often return a generic decline code such as 'do not honor' instead of a specific reason, so the model learns that a payment failed but not why. For a merchant with two acquirers and a few thousand transactions a day, a well-maintained rule set will usually perform as well as a model and is far easier to explain.

How to evaluate AI payment routing claims: 7 questions to ask vendors

All these make vendor numbers easy to misread. Wherever the pitch comes from – a PSP, an orchestration platform, or one of the intelligent payments network providers, the same questions separate a measured result from a marketing one:

1. Uplift against what? A concurrent control group on the same traffic is evidence. A before-and-after comparison, or a comparison with your previous provider, measures everything that changed.

2. Percent or percentage points? A 5% relative lift on an 80% approval rate is four points, not five.

3. Attempts or unique payments? Counting each retry as a new attempt changes the denominator. Measure approval on the first attempt of each payment and report recoveries separately.

4. 'Up to' or median? A best-case client may be a ceiling, not a forecast.

5. What else changed? Tokens, a new acquirer, or a 3DS update launched in the same quarter will show up in the 'AI' number.

6. What happens when the model is wrong? Ask for the fallback logic and whether you can see why a specific transaction went to a specific route.

7. Does it respect scheme retry rules? Approvals bought with excessive retries come back as fees.

A vendor who answers these clearly is worth testing. One who cannot is selling the percentage, not the result.

How to build a payment routing strategy before adding AI

The order of work matters more than the choice of algorithm. A payment routing strategy that sets a model up to succeed usually follows the steps below.

• Normalize payment routing data across providers: decline codes, advice codes, and outcomes for every attempt, in one format.

• Segment approval rates by provider, issuer country, BIN, and payment method, and track the first-attempt rate separately from recoveries.

• Fix coverage, credentials, decline handling, and authentication with explicit rules and cascading.

• Test routing strategies against each other on live traffic before rolling them out.

• Introduce a model on the segments where rules have plateaued, and keep a permanent holdout so you always know what it adds.

Most teams are still working on the first two steps. Our research found that only 8.3% of businesses operate responsive payment systems that can adapt routing and iterate on performance without heavy engineering effort. For everyone else, the fastest way to improve payment approval rates is the ability to change and test rules in hours rather than release cycles. That is the principle behind our routing & cascading engine: 100+ routing attributes, priority, weight, and balance-based strategies, and A/B testing between routing schemes before a change goes live.

Key takeaways

Smart payment routing is one of the most effective levers a payment team has, and machine learning can sharpen it. But a model inherits the routes, credentials, and data you give it. If your decline codes are unclassified, your tokens unprovisioned, and your coverage thin, AI will optimize around the gaps instead of closing them.

Fix the fundamentals first: coverage, tokens, decline handling, and authentication. Then, if you add a model, keep a small share of traffic on your existing rules so you can see exactly what the model adds on top of them. Expect that gain to be a few points at most, not a step change. And when a vendor's AI uplift figure holds up against the questions above, it is worth paying for.