Intelligent Payment Routing: How ML Algorithms Can Cut Cross-Border Processing Costs by 30 to 40%

You have 5 to 15 provider relationships. Your routing logic picks the same one every time. Global cross-border costs still average 6.49%, with bank-initiated transfers at 12.66%.¹ Fewer than 34% of retail cross-border payments are credited within one hour.² ML-powered routing solves this by treating every transaction as a fresh optimization decision, but the production system behind it is where things get hard.

How the routing engine works

The system runs five steps for every transaction:

1. Collect transaction features. Amount, corridor, time of day, customer tier, provider status. A $500K B2B settlement at 3pm EST has a completely different optimal route than a $200 consumer remittance at 11pm.

2. Score every available route in real time. For each provider that covers the corridor, the model scores four dimensions: processing cost, expected settlement speed, success probability, and current provider health. These scores are not static. They update continuously based on the last 24 to 48 hours of transaction outcomes.

3. Select the optimal path based on weighted priorities. Not every transaction optimizes for the same thing. A corporate treasury payment might weight speed at 60% and cost at 40%. A consumer remittance might flip those weights. The model applies customer-level preferences to the scoring output.

4. Execute with automatic failover. If the primary route degrades mid-batch, the system reroutes without manual intervention. No one needs to notice, open a ticket, or update a routing table. The model detects degradation from rising failure rates or latency spikes and shifts traffic in real time.

5. Feed every outcome back into the model. Success, failure, actual cost, actual settlement time. Every transaction makes the next routing decision better. The model learns which providers perform well in which corridors, at which times, for which transaction types. This is what static rules cannot do.

The mistakes that keep routing expensive

Static if-then rules that never get revisited. "Send USD to MXN through Bank A, send USD to PHP through Bank B." These rules were set during integration. Provider costs change quarterly. Clearing speeds shift. Success rates fluctuate. Your routing logic optimizes for conditions that existed a year ago. As we covered in how static routing leaks margin, most teams never revisit these decisions after launch.

No cost versus speed tradeoff per transaction. A $500K B2B settlement and a $200 consumer payout hit the same routing rule. The enterprise client needs speed and is willing to pay for it. The consumer payout should optimize for cost. Same corridor, different optimal rails. Static rules treat them identically.

Manual intervention on provider outages. Someone on ops notices the failure rate spiking. They open a ticket. Someone else updates the routing table. By the time traffic is rerouted, you have lost hours of failed or queued transactions. The FSB's 2024 progress report found that the share of cross-border payments settling within one hour actually declined from 34.2% to 33.5%.² When the system is already this fragile, manual failover makes it worse.

What the numbers say

The optimization surface for cross-border routing is enormous. The gap between average cost (6.49%) and the cheapest qualifying providers per corridor (3.21%) is over 300 basis points, nearly half the total cost.¹ On high-spread corridors, ML routing that consistently selects the lowest-cost qualifying provider can close most of that gap, reducing per-corridor processing costs by 30 to 40%. Deloitte projects AI tools will reduce banking software investments by 20 to 40% by 2028, with institutions pairing AI with specialist engineering teams capturing the largest gains.³

Correspondent banks have been retreating globally for over a decade, increasing concentration and driving costs higher in countries with limited correspondent access.⁴ That means fewer routing options per corridor, which makes optimizing across those options even more critical. As we explored in how correspondent fees quietly erode margin, the hidden costs of suboptimal routing go beyond the quoted fee.

The U.S. Treasury offers a useful parallel. Their ML-enhanced systems prevented and recovered over $4 billion in fraud and improper payments in FY2024, up from $652.7 million the prior year, a six-fold increase.⁵ The lesson is not about fraud. The same principle, replacing static rules with models that learn from outcomes, applies directly to payment routing.

For a company processing $500M in annual cross-border volume, even a conservative blended improvement of 30 to 60 basis points through dynamic route selection yields $1.5M to $3M in annual savings. Not from renegotiating provider terms. From routing smarter with the partners you already have. Settlement acceleration adds further value by unlocking trapped float that sits idle during slow clearing cycles.

Why most teams cannot build this internally

The hard part is not the model. It is the system behind it.

Data normalization across 10+ providers. Every provider has different API formats, rate structures, and reporting cadences. Just getting pricing data into a comparable format across your entire network is a project in itself.

Corridor-specific model training. What works for USD to MXN does not apply to USD to PHP. Each corridor has different provider dynamics, regulatory constraints, and volume patterns. A single global model does not capture the nuances that drive routing decisions.

Sub-second decision latency. The routing decision cannot add meaningful latency to the payment flow. That means the entire score, rank, select cycle needs to execute in milliseconds, not seconds.

Continuous retraining and drift detection. Provider performance shifts weekly. New corridors open. Pricing changes without notice. The model needs to retrain continuously and detect when its predictions start drifting from reality. Then there is failover logic that avoids double payments when a primary provider times out. And reconciliation across providers with different settlement reporting.

Most cross-border payments teams have the domain expertise. They do not have the ML engineering capacity, data pipeline infrastructure, or monitoring tooling to maintain this in production.

How to start in the next 90 days

You do not need a full ML system to begin capturing value.

Weeks 1 to 2: Audit your current routing logic. Map every provider per corridor and document why each was chosen. Flag any routing decisions older than six months. Most teams discover rules set during initial integration that no one has revisited.

Weeks 3 to 4: Build a cost and success rate dashboard per provider per corridor using the last 90 days of transaction data. Include actual landed cost, fees plus FX spread, not just quoted fees. This is the baseline that reveals where the gaps are.

Month 2: Identify the five corridors with the highest cost variance between providers. Quantify the dollar gap. These are your quick wins and the strongest candidates for ML optimization.

Month 3: Implement rule-based routing improvements for those five corridors, steering to the lowest-cost provider that meets your speed and reliability SLA. Measure the impact. This gives you the data foundation and the internal proof point for full ML optimization.

How Devbrew builds this

At Devbrew, we build ML-powered routing engines end to end. Data pipelines across your provider network, custom models trained on your transaction data, real-time decision APIs that select the optimal route per transaction, automatic failover, and monitoring that catches drift before it costs you money. The system learns from every transaction outcome, so cost savings compound as volume grows. Typical implementations reach production in weeks, not quarters.

For cross-border payments companies where routing efficiency directly determines margins, this is the highest-ROI AI investment you can make.

Next step

If your routing logic has not been revisited since it was first set up, it is worth a conversation. We will walk through the challenges you are facing, what is at stake, and where ML-powered routing could create meaningful impact in your payments stack. Share a brief description of your problem and what is at stake when you book. It helps us make the most of our time together.

Book 30 minutes at cal.com/joekariuki/devbrew.

Or reach me directly at joe@devbrew.ai.