AI in Payments After Sibos 2025: What Actually Changes

Sibos 2025 didn’t feel like a victory lap for any single technology. It felt like an industry admitting something out loud: payments modernization is now an integration problem, not an invention problem.

That’s why the most useful signal from Sibos wasn’t “stablecoins won” or “blockchain is back.” It was the quieter convergence of three ideas that, together, define the next phase of fintech infrastructure:

• Digital money types will coexist (stablecoins, tokenized deposits, and CBDCs), because no one model fits every regulatory and operational reality.
• Core networks are adding shared-ledger capabilities to make cross-border settlement less brittle.
• Agentic AI is moving from dashboards to decisions, but only if it can operate inside tight guardrails.

For this AI in Payments & Fintech Infrastructure series, Sibos 2025 is a clean case study: the future isn’t a single rail or a single model. It’s a portfolio of rails, coordinated by AI-driven payment orchestration, governed like critical infrastructure, and measured by outcomes (availability, fraud loss, exception rates, and cost per transaction).

Sibos 2025’s real message: interoperability beats “winner-takes-all”

Answer first: The industry is choosing interoperability because it reduces risk, shortens timelines, and avoids betting the franchise on one architecture.

The original Sibos framing—“showcasing ideas, not declaring change”—lands because most financial institutions can’t afford ideology. They need systems that can:

• Settle across borders 24/7
• Meet evolving compliance requirements
• Handle both legacy and emerging rails
• Stay resilient under fraud pressure

Stablecoins vs. CBDCs vs. tokenized deposits: it’s not a cage match

Answer first: These instruments solve different problems, so the practical path forward is controlled coexistence.

Here’s the simplest way to think about the three (and why AI matters across all of them):

• CBDCs: strongest sovereign backing, slowest rollout cadence. Design choices (privacy, programmability, access model) make them politically and operationally cautious.
• Tokenized deposits: the “bank-friendly” route—deposit guarantees and relationships stay intact, while tokens add programmability and interoperability.
• Stablecoins: fastest iteration cycles and strong cross-border appeal when regulation, reserves, and redemption mechanics are credible.

The operational punchline: once you support multiple forms of money, routing becomes a software problem. You’re not just choosing a corridor—you’re choosing:

• Settlement finality model
• Cutoff times (or lack of them)
• FX path and pricing
• Compliance screening workflow
• Liquidity source and cost

That’s where AI in payments stops being a nice-to-have. If your organization is heading toward “many monies,” you need intelligence that can choose the best path transaction by transaction—and explain why.

Swift’s shared ledger: not “crypto,” but infrastructure modernization

Answer first: Swift’s move signals that the incumbents are absorbing shared-ledger primitives to improve cross-border settlement without forcing everyone to abandon existing rails.

At Sibos 2025, Swift announced it will add a blockchain-based shared ledger to its infrastructure stack, collaborating with more than 30 financial institutions and working with a technology partner. The initial target use case: real-time, 24/7 cross-border payments, with prototype work already underway.

I read this less as “Swift becomes a blockchain company” and more as “Swift is modernizing the parts that break at scale.” Cross-border payments don’t fail because messaging is impossible; they fail because settlement, compliance, and reconciliation don’t line up cleanly.

Why a shared ledger matters (even before it’s mainstream)

Answer first: A shared ledger changes the coordination costs—sequencing, validation, rule enforcement—especially across institutions.

A shared ledger can:

• Record and sequence transactions consistently across participants
• Validate state changes (who owes what, when finality happens)
• Enforce rules via smart-contract-like logic

But the most realistic near-term value is not “everything goes on-chain.” It’s that on-chain primitives can reduce exceptions—the expensive human work of handling investigations, recalls, mismatched references, and ambiguous statuses.

The hidden dependency: ISO 20022-quality data

Answer first: Shared ledgers won’t fix bad data; they’ll amplify it.

Swift’s strength is governance and standards, especially around ISO 20022. That matters because AI systems in payments are only as good as the structured data they can learn from and act on. If payment messages don’t carry consistent party data, purpose codes, and remittance details, you get:

• Higher false positives in sanctions/AML screening
• More manual review
• Poor explainability for any automated decision

A practical takeaway for payments leaders: data modernization is an AI prerequisite. If your ISO 20022 program is treated like a compliance checklist rather than an operational upgrade, agentic AI will underperform—and regulators will notice.

Agentic AI in payments: autonomy is the easy part; governance is the hard part

Answer first: Agentic AI will be adopted first where decisions are bounded, measurable, and reversible.

Sibos discussions highlighted agentic AI—systems that don’t just recommend actions but can execute them inside defined constraints. In payments infrastructure, the early “safe wins” tend to cluster around decisions with:

• Clear policies
• Strong audit trails
• Quantifiable outcomes
• Human override paths

Where agentic AI can act first (without scaring your risk team)

Answer first: Start with constrained operational domains: routing, liquidity, and fraud operations.

Here are three high-value areas where I’ve seen the strongest business case:

1. Intelligent transaction routing

   • Choose rail based on SLA, cost, risk score, and customer preference
   • Re-route when a rail degrades (latency spikes, timeouts, scheme incident)
   • Predict failure likelihood using historical outcomes and real-time signals

2. Liquidity and prefunding management

   • Forecast intraday liquidity needs by corridor and currency
   • Recommend prefunding moves under policy constraints
   • Trigger alerts when abnormal drawdowns occur

3. Fraud and scam response orchestration

   • Prioritize cases based on loss expectancy and customer harm
   • Auto-request additional verification on high-risk patterns
   • Route alerts to the right channel/team to reduce response time

If you want one sentence to align stakeholders: agentic AI earns trust when it reduces exceptions, not when it adds surprises.

“Explain why it did it” isn’t a feature—it’s the product

Answer first: In regulated payments, the business requirement is decision traceability, not model sophistication.

Many institutions got burned by early “black box” risk models that couldn’t explain outcomes to auditors, operations, or customers. Agentic AI raises the bar even further because it’s not just scoring; it’s acting.

A workable standard for agentic AI decisions in payments is:

• Policy-linked reasoning: “This transaction routed via Rail B because Rail A failed SLA threshold for the last 10 minutes and policy allows rerouting under $X.”
• Evidence capture: store features, thresholds, and external signals used
• Replayability: the institution can reproduce the decision later
• Human controls: approval workflows for high-impact actions

If your AI vendor can’t describe their audit approach in plain language, you’re not buying AI—you’re buying future operational debt.

What “adaptable payments architecture” looks like in 2026 planning

Answer first: The winning architecture is modular: orchestration, observability, and controls sit above rails so you can change rails without rewriting everything.

Sibos 2025 reinforced a strategic reality: the rails will keep changing. Stablecoin regulations evolve, CBDC pilots expand unevenly, tokenized deposit networks form consortium-by-consortium, and cross-border infrastructure upgrades arrive in stages.

So the enterprise question becomes: how do you modernize without building a new one-off integration every time the market shifts?

The stack I’d want if I owned payments uptime

Answer first: Put intelligence and controls at the orchestration layer, not inside each rail integration.

A pragmatic target architecture for AI-driven payments modernization typically includes:

• Orchestration layer

  • Normalizes APIs/events across rails
  • Supports rule-based + AI-based routing
  • Handles retries, fallbacks, and idempotency

• Real-time risk and fraud decisioning

  • Shared signals across channels (A2A, cards, wallet payouts)
  • Step-up authentication orchestration
  • Scam pattern detection with feedback loops

• Compliance-by-design

  • Screening and monitoring integrated into flows
  • Consistent case management and audit trails

• Observability and SRE-grade monitoring

  • Rail health metrics (latency, error rates, timeouts)
  • Business metrics (STP rates, exception queues, investigation volumes)

• Data foundation

  • ISO 20022 enriched data
  • Event streaming for payment lifecycle visibility
  • Strong data governance (lineage, retention, access controls)

This matters because the industry is heading toward 24/7 expectations—and we’re publishing this in mid-December, when payment volumes and fraud attempts are both elevated. Holiday peaks expose weak exception handling fast. AI helps most when it’s embedded in the flow, not bolted on after incidents.

Practical next steps: how to move from “Sibos ideas” to operational impact

Answer first: Pick one corridor, one workflow, and one measurable outcome—then expand.

If you’re leading payments, fintech infrastructure, or risk in 2026 planning cycles, here’s a grounded way to start:

1. Choose a narrow use case with clear metrics

   • Example metrics: routing success rate, fraud loss per 10,000 transactions, AML false positives, average investigation time

2. Instrument your payment lifecycle end-to-end

   • You can’t optimize what you can’t see
   • Capture failure reasons, not just “failed”

3. Introduce AI where policy is already defined

   • Let the agent operate inside rules you’d be comfortable explaining to an examiner

4. Build an “explainability packet” for every automated action

   • This reduces internal resistance and accelerates model approvals

5. Design for coexistence from day one

   • Assume multiple rails and multiple digital money types
   • Make switching paths a configuration problem, not a code release

A useful litmus test: if adding a new rail requires rewriting your fraud logic and compliance workflow, your architecture is too tightly coupled.

The next 12 months won’t bring one sweeping change across payments. But teams that build AI-ready, rail-agnostic infrastructure will accumulate advantages quickly: fewer exceptions, lower operating cost, faster product launches, and better resilience when the market shifts.

If you’re mapping your roadmap for AI in payments infrastructure—routing, fraud, compliance, and cross-border settlement—what’s the one decision you’d trust an agent to take next quarter if it could explain itself perfectly?