AI Payments Infrastructure: Lessons from 2025

A 50-year anniversary isn’t just a ribbon-cutting moment—it’s a stress test for what a company (and an industry) actually values. At Payments Unleashed 2025 in New York, the headline was ACI’s milestone and 30 years as a Nasdaq-listed company. The real story, though, was what the conversations revealed about where payments infrastructure is headed next: more real-time, more cross-border, more regulated, and far more dependent on AI to keep it trustworthy.

Here’s my take: most payments leaders still treat AI as an “add-on” to fraud tooling or customer support. That framing is already outdated. AI is becoming part of the infrastructure layer—quietly shaping transaction routing, risk decisions, compliance workflows, and uptime outcomes. And as we enter the heaviest volume season of the year (holiday peaks rolling straight into January bill cycles), the gap between “AI as an experiment” and “AI as operational necessity” gets very obvious.

Payments Unleashed 2025 highlighted that shift in a practical way—through discussions about reliability, embedded trust, interoperability, and the next era of rails (including stablecoins and tokenized assets). Below are the lessons I’d carry into 2026 planning if I ran payments, risk, or fintech infrastructure.

Payments infrastructure is now a strategic asset—and AI is part of it

Payments used to be treated as plumbing. That’s no longer true, and it hasn’t been for a while. The keynote perspective shared at the event put it bluntly: payments drive prosperity and economic competitiveness, which means decisions about interoperability, trust, and infrastructure design will echo for decades.

What changes when you accept that?

1. The bar for resilience becomes “always on.” It’s not enough to process most of the time; the expectation is consistent performance through peaks, scheme changes, and incident scenarios.
2. Trust becomes a product feature. Customers don’t separate “payments experience” from “security experience.” They blame you either way.
3. Regulation becomes a design constraint, not a checklist. If compliance is bolted on late, it slows everything down.

AI fits into this strategic shift because it can do something older infrastructure can’t do well at scale: make fast decisions under uncertainty.

The snippet-worthy truth

AI in payments isn’t primarily about automation—it’s about decisioning at real-time speed with auditable controls.

That single sentence explains why AI is showing up everywhere: fraud models, routing optimization, identity risk, AML alert prioritization, customer communication, and dispute prevention.

Real-time payments mean real-time routing decisions (and AI does the heavy lifting)

Real-time payment rails—instant payments, faster settlement, always-on availability—create a new operational reality: you don’t get time to “fix it later.” If a transaction is misrouted, falsely declined, or allowed through when it shouldn’t be, the cost shows up immediately: failed conversion, scam losses, call center load, regulator scrutiny, or all four.

AI-enhanced transaction routing is the practical answer to that.

What “AI routing” actually means in 2026 planning

It’s not one magic model. In mature environments it looks like a set of decisioning services that weigh signals such as:

• Scheme/rail availability and latency (including maintenance windows)
• Historical approval rates by corridor, issuer, and payment method
• Cost-to-serve per route (fees, FX spread, operational overhead)
• Fraud and scam risk per route (including emerging attack patterns)
• Compliance constraints (geo, velocity thresholds, sanctions screening outcomes)

The output is usually one of three actions:

1. Route A (lowest cost, healthy performance, acceptable risk)
2. Route B (fallback due to latency, higher approval odds, or risk controls)
3. Step-up / hold (add authentication, request more info, or pause for review)

Example: holiday peak routing in practice

During late December volume spikes, approval rates can swing because issuer systems and fraud systems become more aggressive. AI routing can respond by:

• Recognizing rising false declines on one path
• Shifting traffic to a better-performing acquirer/rail
• Applying step-up authentication only to the riskiest slice

The point is selective friction. Good AI routing reduces customer pain while protecting the business.

Fraud has changed: scams and social engineering are the real enemy

One theme that keeps surfacing across events like Payments Unleashed is that “fraud” is no longer just stolen cards and bot attacks. The fast-growing problem is authorized push payment scams, account takeover, mule networks, and social engineering.

Traditional rules engines struggle here because the transaction can look “legitimate” in isolation. AI helps because it can evaluate:

• Behavioral baselines (device, session, payee history)
• Graph relationships (shared identifiers across accounts and recipients)
• Sequence patterns (how the customer got to the payment screen)
• Real-time anomaly detection (unusual timing, amounts, payee changes)

The operational stance I recommend

If you’re a bank, PSP, merchant, or biller building your 2026 roadmap, treat fraud as a multi-layer system, not a tool:

• Pre-transaction: identity proofing, device intelligence, session risk scoring
• At authorization: AI risk decisioning + step-up authentication where needed
• Post-transaction: rapid detection, recovery workflows, and dispute prevention

And don’t ignore the last mile: customer messaging.

A well-timed warning at the moment of payment prevents more scam losses than a thousand after-the-fact investigations.

AI can decide when to warn, how strong the warning should be, and which wording reduces abandonment without reducing protection.

Compliance isn’t slowing down—AI is how you keep up without adding headcount

A senior-leader concern highlighted in the event discussions is familiar: regulatory complexity grows, expectations rise, and nobody gets infinite budget.

Here’s the problem: many compliance processes still assume that humans can review most of the workload. That falls apart in real-time environments.

Where AI delivers measurable compliance ROI

AI doesn’t replace compliance; it prioritizes and structures work so teams can be effective:

• AML alert triage: reduce noise by ranking alerts by probability and impact
• Case summarization: auto-generate investigation narratives from event trails
• Sanctions screening tuning: reduce false positives while preserving coverage
• Audit readiness: maintain model governance artifacts and decision logs

The hard requirement is governance. If you can’t explain why the model made the decision—or you can’t show controls—you’ll end up back in manual mode.

Practical governance checklist (the one teams actually use)

• Define which decisions are fully automated vs. human-in-the-loop
• Keep immutable logs of inputs, outputs, and model versions
• Set performance KPIs beyond accuracy (false declines, scam loss rate, bias checks)
• Run regular drift monitoring (weekly in peak season, monthly otherwise)

Stablecoins and tokenized assets: the infrastructure question comes first

Stablecoins and tokenized assets came up in the context of cross-border evolution and regulatory frameworks. Whether your organization is excited or skeptical, the infrastructure question is the same:

Can you run new value rails with the same reliability, controls, and observability as existing rails?

If the answer is “not yet,” then experimentation should be limited to controlled corridors and well-defined use cases (for example, treasury movements, B2B settlement, or specific remittance flows).

AI’s role here is mostly risk and monitoring:

• Detecting anomalous flows across corridors and counterparties
• Enforcing policy (limits, travel-rule-like requirements where applicable)
• Monitoring liquidity and settlement conditions

The biggest mistake I see is treating stablecoins as purely a product decision. They’re an infrastructure decision.

A practical 90-day plan to make AI usable in payments infrastructure

Most teams don’t fail at AI because they lack ambition. They fail because they try to “install AI” without fixing the decisioning plumbing.

If you want near-term progress (and you want to avoid a never-ending pilot), focus on this sequence.

1) Pick one decision that matters

Good first targets are:

• Real-time fraud scoring for a single channel
• Scam detection warnings for high-risk payees
• Routing optimization for one corridor or one payment method

2) Instrument the journey end-to-end

You need:

• Clean event data (login → checkout → payment → confirmation)
• Outcomes (approved/declined, chargebacks, scam reports, disputes)
• Feedback loops (human review labels, customer confirmations)

3) Define “better” with numbers

Avoid vague goals. Use metrics like:

• Reduce false declines by 15–25% on a target segment
• Cut average fraud loss per 10,000 transactions by X
• Improve authorization rate by 1–2 percentage points without raising loss rate

4) Make it operable

If a model can’t be monitored, rolled back, and explained, it’s not production-ready. Full stop.

The bigger lesson from Payments Unleashed 2025: trust is engineered

The CEO reflection from Payments Unleashed 2025 emphasized something that matters to every payments leader: the best client relationships endure because the infrastructure works—through Y2K, crises, rail shifts, and now AI.

That continuity is a reminder that trust isn’t branding. It’s engineered outcomes: uptime, risk controls, clear audit trails, and predictable performance.

As part of the AI in Payments & Fintech Infrastructure series, this is the thread that ties everything together: AI only matters when it improves the core infrastructure outcomes—security, routing performance, resilience, and compliance velocity.

If you’re mapping your 2026 priorities, here’s the question to carry into every meeting: Which decisions are still slow, manual, or inconsistent—and what would it look like to make them real-time and auditable?