Up to 1.2 billion routing paths, one decision per payment. Open the routing workflow of a merchant with several payment providers and you'll find years of decisions stacked up. German cards go to one provider, Apple Pay to another, and a BIN range that underperformed last spring has a branch of its own.
Each rule made sense the day someone wrote it. Together they decide where millions of payments go, and few teams have the hours to check whether each rule still holds.
That's why we built Autopilot. For every payment, a model trained on your own payment history picks the provider most likely to approve it. If that provider declines, Autopilot moves straight to the next best one.
There are no rules to write for choosing between providers, and none to keep up to date. Each outcome feeds the next retrain of your model, so Autopilot's routing keeps learning and improving as issuers and providers change their behavior.
Key points
- Autopilot gets better with use: every approval and decline becomes evidence for your model's next retrain, so your team doesn't rewrite rules when underlying parameters change.
- After a decline, Autopilot retries on the next best untried provider, up to your attempt limit. By default, hard declines, such as a stolen card, stop the retries.
- Every routing decision is recorded with each provider's estimated approval probability, so you can see why a payment went where it did.
- Payrails A/B tested the approach on live traffic for many months. In one deployment, 22% of payments went to a different provider and total approvals rose by 0.8%, from a starting rate above 90%.
Why routing rules fall behind
Static routing rules send every payment that matches a rule to the same provider. They're written from last quarter's data, while issuers keep changing how they score risk. A rule that sent Spanish debit cards to your second acquirer in March can be costing you approvals by June.
Rules can also only encode the patterns a person has acted on. Payment traffic splits by issuer, card type, currency, network, token type, retry state, even time of the day. Each combination has its own approval rate at each provider.
Across the Payrails platform, those combinations multiply. Start with 100+ integrations with PSPs and acquirers and their merchant accounts. Multiply by card and wallet payments, by every card type and network, by 3DS on or off, by cards issued in 150+ countries and by payments in 30+ currencies, and you reach the 1.2 billion theoretical paths.
Real-world overlap brings the number down, since cards from one country often pay in only 2 or 3 currencies, but it stays above 100 million. No merchant uses all of them. Even a single merchant's share is more than a team can map by hand, and it shifts each time variables change.
Our data team described this in how we use machine learning to lift authorization rates as thousands of needles in a haystack: narrow slices where a different provider would approve meaningfully more payments. An analyst finds some of them. A model trained on every past transaction finds far more, and improves with every retrain.
Picking a provider is a prediction: which one is most likely to approve this payment, right now. Autopilot answers it for each payment as it arrives.
"We assumed enterprise merchants would want to control every routing decision. Most of them told us they'd rather not choose between providers themselves, least of all in markets they don't know well. That's the decision Autopilot makes for them."
– Aristotelis Lefas, Head of Payment Strategy, Payrails
How does ML and AI payment routing pick a provider?
Each merchant has its own model, trained on all past payments. For every payment, it weighs signals about the payment itself, when it happens, and the customer, merchant and payment instrument behind it, including how their earlier payments went. That lets it treat a Turkish debit card at 2 a.m. differently from a UK credit card at lunchtime, even when both match the same routing rule.
From those signals, the model estimates each candidate provider's chance of approving the payment. The scores are calibrated per provider, so they compare as probabilities, and the payment goes to the provider with the highest one. The decision takes a fraction of a second, so it doesn't slow the payment down.
Autopilot also keeps track of how sure it is about that prediction. Next to each score, it records how many similar past payments back it. Where that evidence is thin, it still routes to its best pick, and the outcome becomes evidence for next time. Payrails retrains your model regularly, and each retrain learns from those results too, including the less common payments.
When a provider declines
After a decline, Autopilot picks again from the providers this payment hasn't tried, so a retry never goes back to the provider that just declined it. It's a cascading retry, with the model choosing the order for each payment. By default, a hard decline, such as a card reported stolen, ends the sequence straight away.
Retrying a hard decline on another provider rarely succeeds, and the card schemes charge for it. Visa doesn't allow retries on declines in its "issuer will never approve" category, and Mastercard charges a fee when a payment declined with its "do not try again" merchant advice code is tried again and declined within 30 days (PayPal, 2026).
Simple to configure in Workflow Studio
Autopilot sits in your payment workflow where you'd otherwise authorize with a fixed provider. It ends in one of 2 outcomes, Success or Declined, and you connect each to the rest of your workflow like any authorization.
Here's what you configure:
Your own rules still run before Autopilot. You can keep sending Amex to a direct Amex connection and let Autopilot choose between your acquirers for Visa and Mastercard.
To start small, put Autopilot in a new workflow version and use controlled rollout to send it part of your traffic. Compare its approval rate with your current routing before you move more traffic over.
See the exact path each payment takes
Once a model makes routing decisions, auditors will want each one explained, and so will merchants and card schemes. Autopilot records each decision on the workflow execution, so you can audit every routing choice. For each payment, you can see:
- the provider the payment was sent to, and the one the model scored highest
- each candidate provider's estimated approval probability
- how many past payments back each estimate, which shows how confident the model was
Across thousands of payments, the record shows where Autopilot sends your traffic and how each provider performs on it.
Tested on live traffic at scale
Autopilot grew out of hundreds of experiments that our data science and payment strategy teams have been running with merchants from early 2025. They trained routing models on each participating merchant's traffic and tested the model's routing against workflows our payment strategy team had built, on the same live traffic.
For one merchant, training drew on roughly 12 million payments across 6 providers, with more than 100 inputs and over 5 months of history.
The tests were strict A/B comparisons. A control group kept the merchant's existing routing, and a treatment group ran the model's recommendations, ramping from 5% to 50% of traffic over about a week, with alerts set to fire if approval rates dropped.
In the deployment we published in August, the model re-routed 22% of about 1 million monthly payments to a provider the existing rules wouldn't have chosen, and lifted total approvals by 0.8%, on a starting approval rate above 90%. Per market, the uplift ranged from 0.5% to 1.5%.
In another deployment, for a merchant in the Middle East that also approved more than 90% of its payments, the uplift reached 4.5%.
Above a 90% approval rate, each extra approval is hard to find, and a gain of a fraction of a percent is invisible on any single payment. Over a year of volume, it adds up. At an average payment of $100, each percentage point of approval on 1 million monthly payments is about $1M a month in sales that would otherwise be declined. To size it for your own volumes, use our guide to the ROI of payment orchestration.
Those tests had a bottleneck: a person applied each routing change, so the loop took days. Autopilot makes the decision inside the workflow, one payment at a time.
Routing that learns and self improves
Running payments on Autopilot means your team sets the providers and the attempt limit once, and Autopilot handles every payment after that. Each one goes to the provider most likely to approve it, and each outcome adds evidence to your model. The longer Autopilot runs, the more of its decisions rest on a deep history of similar payments, and each retrain folds in the latest issuer and provider behavior.
That loop works because payment routing has what automated decisions need: decisions are frequent, outcomes come back fast, mistakes can be reversed and someone clearly carries the liability.
Our data science team is already testing the next step: adding anonymized patterns from payments across the Payrails network to each merchant's own history, so a merchant entering a new market can start from what works there on day one. Early results in this direction already show significant potential: payments optimized from day one in any market you launch.
Approval rate is the first goal Autopilot optimizes for, and cost is next. Cost-aware routing is already in development, so Autopilot will weigh each provider's chance of approval against what it charges.
Who Autopilot is for
Any business, really. Autopilot pays off when you have at least 2 providers that can process the same payment, for example 2 acquirers for the same card brands in a market. To get started, Payrails trains a model on your payment history with those providers and switches Autopilot on in your Workflow Studio.
If you're new to payment routing, our guide on what payment routing is covers the basics. For setup steps, see the Autopilot guide in our docs.
If your team still updates routing rules by hand, start with one market and part of your traffic. Talk to our team to see how Autopilot would route your payments.
FAQs
What is ML and AI payment routing?
Autopilot uses proprietary ML models and AI capabilities to find the path most likely to get each transaction approved. The model is trained on past payments and their outcomes, and it estimates each provider's chance of approving the payment in front of it, so the route is decided for each transaction in real time.
Does Autopilot replace my routing rules?
Only where you decide to allow it. Rules that run before Autopilot keep working, so you can keep sending certain card brands or countries to a specific provider and let Autopilot choose for the rest. You can also start on part of your traffic with controlled rollout and compare approval rates against your current routing.
How does Autopilot handle declined payments?
Soft declines move on to the next provider in the model's order until one approves or the attempt limit is reached. Hard declines stop the sequence by default, because a retry elsewhere rarely succeeds and can bring card scheme fees.
Can I see why Autopilot chose a certain provider?
Yes. Each payment's workflow execution shows the provider chosen and each candidate's estimated approval probability, with the number of past payments behind each estimate.
What data does Autopilot learn from?
Your own payment history. Each merchant has its own model, so its picks reflect your customers, your providers and your business specifics. It also keeps learning from every payment it routes.
Does Autopilot get better over time?
Yes. Autopilot learns from the result of each routing decision, so the longer it runs, the more evidence sits behind each choice. Today that evidence is your own payment history. Soon it will also include anonymized patterns from the Payrails network, which helps most in a market you haven't sold in before.
How many providers do I need?
At least 2 that can process the same payment, and enough payment history with each for the model to learn from.
I am the text that will be copied.
