AI Routing for Cross-Border Payments in a Hybrid Era

A lot of cross-border payments “innovation” is really just new plumbing bolted onto old pipes. And that’s not a bad thing.

The future of cross-border payments will be a co-existence of old and new rails—SWIFT messages and ISO 20022, correspondent banking and instant schemes, batch files and APIs, manual investigations and automated decisioning. The mistake is pretending a single network or standard will replace everything fast. The better approach is building systems that operate well in the messy middle.

This matters because cross-border volume is still rising (B2B payouts, creator economy, marketplace settlements, global payroll, cross-border e-commerce), while expectations have shifted: businesses want predictable fees, faster settlement, fewer exceptions, and better traceability. In the “AI in Payments & Fintech Infrastructure” series, I keep coming back to a theme: AI has its biggest impact when it improves infrastructure outcomes—routing, risk, reconciliation—without requiring the world to reboot.

Why cross-border will stay hybrid (and why that’s healthy)

Cross-border payments will remain hybrid because regulation, liquidity, and local clearing realities don’t standardize on your product roadmap.

Legacy rails persist for practical reasons: they’re widely connected, understood by banks and regulators, and embedded in treasury operations. Newer rails (instant payments, real-time FX, API-first networks, tokenized settlement) bring speed and better data, but coverage is uneven across corridors, currencies, and use cases.

A realistic 2025–2028 operating model looks like this:

• Old rails for reach: correspondent banking, SWIFT-based messaging, established local clearing access through partners.
• New rails for performance: faster payments where supported, API-based payment initiation and tracking, improved remittance data with ISO 20022.
• A decision layer on top: selecting the best path per transaction based on cost, speed, risk, and certainty.

A hybrid future isn’t a compromise—it’s an optimization problem.

The winners won’t be the firms with the loudest “new rails” story. They’ll be the firms that can consistently pick the best available route and explain the outcome to finance teams in plain language.

The real problem: routing, exceptions, and uncertainty

If you’re running cross-border flows, the pain isn’t “we lack rails.” It’s that outcomes are too often uncertain.

Where cross-border breaks in practice

Most operational cost sits in the gap between “payment sent” and “payment settled.” Common failure modes:

• Routing ambiguity: multiple possible paths (and intermediaries) with different fees, cut-off times, and failure rates.
• FX and fee opacity: hidden lifting fees, unpredictable spreads, and corridor-specific quirks.
• Compliance friction: sanctions screening, name matching, and local regulatory rules that vary by destination.
• High exception rates: returns, repair messages, missing beneficiary data, or mismatched formats.
• Investigation overhead: status inquiries, trace requests, manual reconciliation, and partner back-and-forth.

The business outcome is measurable: more exceptions mean more working capital tied up, higher support load, and strained customer relationships.

Here’s the stance I’ll take: If your cross-border strategy doesn’t include an intelligence layer for routing and exception prevention, you’re choosing to pay the “manual tax” forever.

Where AI fits: the decision layer that bridges legacy and modern rails

AI adds value when it acts as a real-time decision engine that can work with both legacy and modern payment infrastructure.

Think of it as “smart routing + smart operations,” not a shiny chatbot bolted onto a broken process.

1) AI-powered transaction routing (cost, speed, certainty)

The best route isn’t always the cheapest on paper. It’s the route that delivers the best probable outcome given constraints.

An AI routing model can score available paths using features like:

• Historical success rate by corridor, bank, scheme, and time-of-day
• Average time-to-settle and variance (predictability matters)
• Fee composition patterns (including likelihood of intermediary fees)
• Return/repair probability based on field-level quality (address formats, name order, purpose codes)
• Current liquidity position and prefunding constraints
• Risk signals (sanctions/name-match likelihood, unusual beneficiary patterns)

Output: a ranked set of routes with a confidence score and an explanation finance teams can audit.

A practical routing objective function usually balances:

1. Certainty (probability of success without repair)
2. Speed (expected settlement time)
3. Total cost (fees + FX + operational cost of exceptions)

If you only optimize for fees, you’ll often increase exceptions—and exceptions are expensive.

2) AI for exception prevention (fix data before it breaks)

A surprising share of cross-border failures come from formatting and data completeness.

AI can prevent issues upstream by:

• Validating and normalizing beneficiary data (names, addresses, bank identifiers)
• Predicting missing fields required for a corridor (purpose of payment, local tax IDs)
• Detecting “high repair risk” payments before submission and prompting correction

This is one of the fastest ROI areas because it reduces manual investigations.

3) AI-assisted compliance and fraud detection without killing conversion

Cross-border risk isn’t just fraud; it’s also compliance (sanctions, AML, KYC). The challenge is reducing false positives.

Modern approaches combine:

• Graph features (relationships among entities, accounts, devices, beneficiaries)
• Behavioral patterns (new corridors, unusual payment timing, payee velocity)
• Adaptive thresholds by corridor and customer segment

The goal is fewer unnecessary holds while improving true-positive detection.

4) AI for reconciliation and payment status intelligence

Hybrid networks mean hybrid tracking. Some rails support rich status updates; others don’t.

AI can infer status and expected settlement windows by correlating:

• Message events (acknowledgements, repair requests)
• Partner response patterns
• Similar past transactions

This isn’t about “guessing.” It’s about giving operations a probability-weighted view: likely settled, likely pending, likely to require repair, with suggested next actions.

A workable architecture for “old + new” cross-border

You don’t need to replace your core banking system to get smarter outcomes. You need a layered approach.

The three-layer model

1) Connectivity layer (rails and partners)

• SWIFT/correspondents
• Local clearing access via partners
• Instant payment schemes where available
• API-based payout networks

2) Orchestration layer (rules + workflow)

• Payment initiation and formatting
• Routing constraints (currency, cut-offs, limits)
• Exception workflows and case management
• Audit trails and reporting

3) Intelligence layer (AI decisioning)

• Route scoring and selection
• Risk scoring and compliance triage
• Data quality prediction and enrichment
• Settlement-time prediction and reconciliation support

The orchestration layer makes things run. The intelligence layer makes them run well.

What this looks like in a real scenario

Consider a mid-market marketplace paying out sellers in three regions.

• Some destinations support fast local transfers.
• Some require correspondent banking.
• Some corridors have volatile failure rates due to beneficiary data requirements.

A rules-only approach quickly turns into a brittle maze.

An AI-assisted approach can:

• Route repeat sellers through the fastest low-exception path
• Send first-time sellers through a path optimized for certainty and compliance
• Auto-prompt for additional data only when the model predicts repair risk
• Dynamically reroute during holidays and cut-off windows (December is notorious)

That’s hybrid infrastructure behaving like a modern product.

What to measure: outcomes that matter to finance and ops

If you want AI in payments to generate leads (and budgets), talk in metrics your CFO and Head of Ops care about.

Track these five:

1. Straight-through processing (STP) rate: % of payments settled without manual touch
2. Exception rate: returns + repairs per 1,000 transactions
3. End-to-end settlement time: median and 90th percentile (P90)
4. Total cost per payment: fees + FX + internal ops cost
5. Traceability SLA: time to provide a reliable status update to customers

A healthy AI routing program should improve P90 settlement time and exception rate first. Cost savings typically follow because fewer exceptions reduce human workload.

If you can’t explain your cross-border outcomes with numbers, you can’t improve them with AI.

Implementation: how to introduce AI without breaking controls

AI in financial infrastructure needs guardrails. Here’s what works in practice.

Start with “decision support,” then graduate to automation

Phase 1 (2–6 weeks):

• Model scores routes and flags risky payments
• Humans approve routing changes
• You collect feedback and outcomes

Phase 2:

• Auto-route within approved constraints (corridors, amounts, customer tiers)
• Auto-block only when confidence is high

Phase 3:

• Continuous optimization with A/B tests across corridors
• Automated exception handling for predictable repairs

Keep an audit trail that survives scrutiny

For each routing decision, store:

• Candidate routes considered
• Model score and top features
• Final route chosen and why
• Post-transaction outcome (settled/returned/repaired, time, fees)

This is how you keep compliance comfortable and improve the model.

Don’t ignore data contracts

Hybrid payments fail when data isn’t consistent. Invest in:

• Normalized data models (ISO 20022 mapping where relevant)
• Clear field validation rules per corridor
• Feedback loops from exceptions back into enrichment

If your data is messy, your AI will be confidently messy.

The near-term future: smarter interoperability, not a single winner

The next few years won’t crown one universal rail for cross-border payments. What will change is how intelligently firms combine rails.

I expect three trends to dominate 2026 planning:

• Routing as a competitive advantage: not just “send a payment,” but “send it the best way right now.”
• Operational AI becomes mandatory: exception prevention and reconciliation will decide unit economics.
• Compliance moves from static rules to adaptive risk: fewer false positives, better escalation.

For teams building in the “AI in Payments & Fintech Infrastructure” space, this is the real opportunity: AI that makes old and new infrastructure behave like one coherent system.

If you’re modernizing cross-border payouts or treasury flows, the question to ask your team is simple: Where are we still relying on manual judgment that could be encoded, measured, and improved?