Visa Rewires Payments for AI Agents and Stablecoins

Think of Visa the way you think of the old Roman road network: nobody actually loves the cobblestones, but every legion, merchant, and tax collector ended up using them because the alternative was a muddy field. At the Visa Payments Forum 2026, the company effectively announced it's repaving the whole thing, with AI agents driving one direction and stablecoins driving the other. The roads are still Visa's. That's the whole point.

What Happened

The headline act is Visa Intelligent Commerce, an architecture aimed at giving AI agents trust, controls, and connectivity inside the Visa network. Alongside it, as Fintech Singapore reported, Visa announced a collaboration with OpenAI to enable secure payments using Visa infrastructure inside agentic commerce environments. So when your ChatGPT instance buys you a flight next year, there's a decent chance Visa is sitting in the auth path.

Two supporting tools landed with it. Agent Score is a readiness evaluation for agentic commerce, basically a credit score for bots. The Agentic Directory is a verified registry of legitimate merchants and agents, a phone book for machines that don't want to get phished by other machines. Behind the scenes, Visa deployed its Large Transaction Model, an AI system trained on billions of data points, designed to improve fraud detection while reducing false declines.

On the back end, the story is stablecoins. Visa has moved billions of dollars in stablecoins across VisaNet and hit an annualised run rate of roughly US$7 billion as of March 2026. Issuing banks can now settle onchain seven days a week. The network is expanding pilots across multiple blockchains, currencies, and regions, and is building infrastructure to help banks convert traditional deposits into programmable, always-on tokenised deposits. More than 160 stablecoin-linked card programs are live or in development globally.

Rounding it off: the cloud-native Pismo core banking platform for modular issuer infrastructure, and Unified Checkout, an orchestration layer for both card and non-card rails inside existing merchant systems. Jack Forestell, Chief Product and Strategy Officer at Visa, summed it up: "AI is transforming the front end of commerce. Stablecoins are reshaping the back end." His pitch is that Visa's job is to make all of it "work securely, reliably and at global scale, for every participant in the ecosystem."

Technical Anatomy

Strip the marketing veneer off and what Visa is actually shipping is an identity and risk layer for non-human buyers. That's the boring bit, and it's the interesting bit.

Consider the agentic flow. An LLM-driven agent decides to buy something on your behalf. The merchant's checkout doesn't know if it's talking to a human, a benign agent, or a jailbroken script trying to drain a card. Agent Score and the Agentic Directory together try to solve that: one gives a risk-weighted assessment of the agent's behaviour, the other verifies the agent and the merchant are who they claim to be. It's PKI thinking applied to autonomous software, with Visa as the de facto certificate authority. Anyone who has wired up OAuth flows for service accounts knows how quickly "trust the token" turns into "trust the issuer of the token". Visa wants to be the issuer.

The Large Transaction Model is the other half. Fraud detection on agent traffic is a different problem than fraud detection on humans. Agents are fast, repetitive, and don't fat-finger CVVs. Conventional rules engines will either nuke legitimate agent volume with false declines or wave through coordinated drains. A model trained on billions of data points, with agent behaviour as a first-class feature, is the only way that math works at network scale.

Now the back-end road. The stablecoin pilots aren't a novelty: a US$7 billion annualised run rate as of March 2026 puts them firmly in production-grade territory. Seven-day onchain settlement for issuing banks is the quiet killer feature. Anyone who has run a treasury reconciliation job on a Sunday night, waiting for Monday's ACH cutoff, understands what removing weekends from finance actually means. Combine that with tokenised deposits (programmable money that still sits on a bank's balance sheet) and you get a settlement layer that looks more like Kafka than SWIFT.

Pismo as the cloud-native core gives Visa a place to host issuer logic next to all of this. Unified Checkout is the merchant-side adapter, sliding non-card rails into existing card integrations without forcing a forklift upgrade. If you've ever run blue/green deploys on a payments service, you'll recognise the pattern: keep the old road open while you pave the new one underneath.

Who Gets Burned

The first group sweating is the agentic-payments startup cohort. Anyone building a "Stripe for AI agents" thesis just watched the incumbent show up with an OpenAI partnership, a scoring system, a directory, and a fraud model. That doesn't kill the category, but it makes the pitch harder. Differentiation now has to be in vertical depth or in non-Visa rails, not in the generic plumbing.

Second: traditional fraud and risk vendors. If Visa's Large Transaction Model becomes the default decisioning layer for agent traffic, third-party fraud scoring on Visa rails turns into a feature, not a product. Teams that sell rules engines and ML-based scoring to issuers will need a story for why their model adds signal on top of Visa's, not instead of it.

Third: crypto-native payment processors. Visa pushing US$7 billion annualised through stablecoin settlement, with 160+ card programs live or in development, eats directly into the wedge that on/off-ramp specialists were carving out. The gambling and remittance verticals especially, where seven-day settlement is a real pain point, now have an incumbent answer.

iGaming operators have a more nuanced exposure. Faster settlement and tokenised deposits are a gift for cash-flow management and player payouts. But Agent Score adds a new gating function: if your platform integrates AI assistants that place bets or manage bankrolls (a category that's quietly growing), those agents will need to clear Visa's readiness bar or get declined. Compliance teams should be reading the Agentic Directory spec the same week it drops.

Ad-tech and fintech infra teams get pulled in too. Unified Checkout means non-card rails (stablecoins, account-to-account, wallets) start showing up inside flows that were card-only. Anyone running PCI scope analysis, tokenisation services, or chargeback tooling needs to re-scope. The next 90 days for affected platform teams look like architecture reviews, not feature work.

Playbook for Engineering Teams

If you're a platform lead in fintech, iGaming, or commerce infrastructure, here's where I'd put cycles this quarter.

First, treat agent identity as a real domain. Stand up a registry of which agents your system trusts, with revocation, scopes, and audit logs. Even if you never integrate Visa's Agentic Directory directly, the design pattern (signed agent identity, behavioural scoring, merchant-side allowlists) is going to be table stakes. Your existing service-to-service auth on something like Kubernetes service accounts is a reasonable mental model, just with stricter blast-radius controls.

Second, instrument everything. Agent traffic will distort every dashboard you have: conversion rates, fraud rates, AOV, session length. If you don't have proper distributed tracing on your checkout path, with semantic conventions you can extend for "actor type", you'll be flying blind. OpenTelemetry is the obvious starting point; add a custom attribute for actor identity early, before you need it.

Third, model stablecoin settlement in your treasury and reconciliation systems now, even if you don't intend to use it for six months. Seven-day onchain settlement breaks assumptions baked into ledger jobs, batch windows, and reporting cuts. The teams that get burned will be the ones who discovered on a Sunday night that their daily close job assumes a Friday cutoff.

Fourth, audit your checkout for non-card readiness. Unified Checkout is a sign of where the merchant side is going. If your payment service can't cleanly add a non-card rail without a six-month project, that's tech debt with a deadline now attached.

Key Takeaways

• Visa Intelligent Commerce, the OpenAI partnership, Agent Score, and the Agentic Directory together position Visa as the identity and risk authority for agentic commerce.
• The Large Transaction Model, trained on billions of data points, makes Visa's fraud layer the default for agent traffic, squeezing third-party fraud vendors.
• US$7 billion annualised stablecoin run rate as of March 2026, plus seven-day onchain settlement for issuing banks, means stablecoins are now production infrastructure, not a pilot.
• 160+ stablecoin-linked card programs and tokenised deposit infrastructure put pressure on crypto-native processors and traditional correspondent banking flows.
• Engineering priorities: agent identity, actor-aware observability, treasury systems that handle seven-day settlement, and checkout layers ready for non-card rails.

Back to the Roman road for a second. The empire didn't fall because the roads failed; they were still being used a thousand years later by people who'd forgotten who built them. Visa is betting that whether the next decade of commerce is run by humans, agents, stablecoins, or all three at once, the cobblestones underneath will still say VisaNet. On the evidence of this week, that bet is looking annoyingly well-placed.