Payment systems are being fundamentally rebuilt right now. Not incrementally, but at a depth last comparable to the introduction of online banking in the 1990s. The driver isn't crypto, nor open banking, nor real-time payments — it's Artificial Intelligence.

If AI agents really do trigger more transactions than humans in the coming years — a forecast now shared by Visa, Mastercard, McKinsey, and a growing number of banks — then the existing rails no longer suffice. This post sorts out what's actually happening, which architectures are taking shape, and what role crypto infrastructure plays in this.

1. Why existing payment systems aren't enough

Credit cards, ACH transfers, SEPA direct debits, Stripe subscriptions — all share one assumption: at the end of the transaction is a human filling in a form and clicking "Pay." ACH transfers take 1–3 days to settle. Credit card payments are chargeback-eligible for months. Stripe subscriptions assume a human has stored a card once.

An AI agent can use none of this. It needs payment systems as programmable and instantaneous as the HTTP calls it's already making. Specifically:

• Sub-cent transactions must be profitable (an API request can cost $0.001)
• Settlement must happen in seconds, not days
• Spending limits must be enforceable programmatically — not via after-the-fact disputes
• Authentication must work machine-readably, not via 3D Secure popups
• Identity must distinguish between legitimate agents and fraud bots

This isn't a nice-to-have. It's the technical precondition for agents to operate autonomously at all. And that's exactly why the biggest card networks, the biggest cloud providers, and the leading crypto-infrastructure companies are pivoting at the same time — each with its own answer.

2. What Visa and Mastercard are building

If you think agentic payments is a crypto topic, you're missing half the story. Visa and Mastercard each launched their own frameworks in 2025 — and they're moving at considerable speed.

Visa Intelligent Commerce (launch: April 2025) rests on four components: Visa-tokenized credentials that can be scoped and issued to AI agents (an agent booking a trip gets a token valid only for that airline and that travel window); authentication via behavioral signals and issuer risk scoring instead of manual confirmation; direct integrations with OpenAI, Anthropic, Microsoft, Perplexity, and Stripe; and the Trusted Agent Protocol for identity verification.

In December 2025, Visa reported "hundreds of agent-initiated transactions in live production." Over 100 partners are currently building on the platform, 30+ are active in the sandbox, and 20+ agents are integrated directly. Visa forecasts that millions of consumers will make AI-agent purchases by the 2026 holiday season. Rubail Birwadker, SVP Growth Products at Visa, put it plainly: "In 2026, AI agents won't just assist your shopping — they will complete your purchases."

Mastercard Agent Pay (launch: April 2025) follows similar logic with "Agentic Tokens" — credentials scoped to AI agents with programmable spending limits, counterparty allowlists, and transaction categories. Mastercard has already expanded the framework to Hong Kong and works with the same LLM platforms as Visa.

Both card networks thus have a clear strategy: the agentic world should run on their rails, not around them. They have 4 billion issued cards, regulatory clarity in 200+ countries, and established fraud detection — advantages crypto rails don't have.

3. What hyperscalers and banks are building

In parallel, cloud providers and traditional banks are building.

AWS Bedrock AgentCore Payments went into preview on May 7, 2026. AWS thereby gives AI agents native payment capability — settlement runs in USDC on the crypto networks Base and Solana. Coinbase and Stripe are the payment-rail partners. A hyperscaler has thus officially embedded programmable crypto payments for AI agents into its standard infrastructure.

Google has established its Agent Payments Protocol (AP2) as an umbrella standard, into which both card rails and crypto rails (via x402) are integrated. Stripe has in parallel announced Shared Payment Tokens (SPTs), interoperable with Visa Intelligent Commerce, Mastercard Agent Pay, Affirm, and Klarna.

Citi launched its own AI-agent platform Arc in May 2026 for banking workflows. More than 80% of Citi's 180,000 employees with access to AI tools use them regularly. In parallel, Citi announced it will offer native Bitcoin and Ether custody in 2026 and is evaluating issuing its own Citi-branded stablecoin. The Citi Token Services platform currently moves close to $1 billion daily.

JPMorgan is piloting tokenized deposits and stablecoin settlement via Kinexys. PNC, Wells Fargo, and Citi are jointly exploring a stablecoin initiative via Early Warning Services (the company behind Zelle).

What's happening here is a rare convergence: card networks, hyperscalers, and large banks are simultaneously building agentic infrastructure — and all three are using crypto rails in parts of it.

4. The parallel crypto layer

While Visa/Mastercard extend the established rails, crypto is building a fundamentally different architecture: payment embedded directly into HTTP.

Two protocols have established themselves as standards, both leveraging the HTTP status code 402 ("Payment Required") — reserved since 1996 but barely used.

x402 (Coinbase + Cloudflare, live since 2025, now with Google and Visa in the x402 Foundation). How it works: an agent calls an API, the server responds with HTTP 402 containing price and wallet address. The agent signs a USDC authorization, the request is resent, the resource is delivered. A single roundtrip, no accounts, no API keys. Settlement on Base and Solana, protocol fee zero, only network gas applies.

In April 2026, the x402 Foundation became part of the Linux Foundation — 22 founding members, open governance. Google explicitly integrated x402 into its AP2 protocol.

L402 (Lightning Labs, Bitcoin-centric). Same principle, but settlement via Bitcoin Lightning. Lightning Labs released the LN Agent Toolkit in February 2026. The structural advantage: no issuer risk (USDC depends on Circle, Lightning doesn't), real sub-cent fees, and a single neutral asset instead of stablecoin fragmentation.

Both standards are not a replacement for Visa/Mastercard — they're a parallel layer for exactly the transactions that card networks structurally cannot process profitably: micropayments under $1, machine-to-machine calls, API access.

5. The numbers

The volumes and forecasts are impressive — and at the same time very uncertain.

Stablecoin transaction volume 2025: $33 trillion, +72% over 2024. Supply over $300B. Forecast 2026: roughly $420B in supply. Agentic payments are explicitly named in analyst reports as a growth driver. At current levels, around 99% of AI-agent payments run through USDC. Circle is the biggest quiet beneficiary so far.

Analyst forecasts:

• McKinsey & Company: AI agents account for $1 trillion in US transactions by 2030
• Visa: Millions of consumers will use AI agents for purchases by Holiday Season 2026
• Juniper Research (April 2026): Agentic commerce $1.5 trillion by 2030
• Galaxy Digital: $3–5 trillion in B2C revenue through agentic commerce by 2030
• Goldman Sachs: $50 billion/year just for AI-agent spending on digital services by 2028
• Citi: Global AI market >$4.2 trillion by 2030, of which $1.9 trillion enterprise AI
• MarketsandMarkets: AI-agent market from $7.84 billion (2025) to $52.62 billion (2030), CAGR 46.3%

The spread is wide, the direction consistent. Important context: actual on-chain volumes of pure agent-payment protocols are still in the low six-figure USD range per day. The forecasts describe an expected catch-up path, not the status quo.

6. Bitcoin in the new payment system

The question of whether Bitcoin plays a role in this picture has a surprisingly data-based answer.

The Bitcoin Policy Institute tested 36 large AI models across more than 9,000 simulated monetary-decision scenarios. The models had to choose between Bitcoin, stablecoins, and fiat currencies — for savings scenarios, payment scenarios, and transfer scenarios.

The result was consistent across all 36 tested models:

• In 79% of savings scenarios, the models picked Bitcoin
• In transactional scenarios, stablecoins dominated
• Fiat currencies were rarely picked

The models converged — from first principles — on a structure familiar to the crypto community: Bitcoin as a reserve asset, stablecoins as transactional currency. This is relevant because future AI treasury decisions will be based on exactly this kind of reasoning. A company equipping an AI agent with autonomous spending decisions will implicitly adopt this same logic.

Lightning Labs therefore argues that Bitcoin/Lightning is structurally superior for agent payments: a single asset, no issuer risk, sub-cent fees, globally available. The counterpoint is operational: Lightning channels need to be kept active and liquid, and the tooling is a higher hurdle for Python/Node developers today than USDC on Base.

Realistically, both worlds will coexist. Tether is already building USDt on Lightning via Taproot Assets — stablecoin stability on Bitcoin-native rails.

7. The unsolved problems

For all the momentum, three structural problems still lack mature answers:

AML and KYC compliance. If an AI agent makes 10,000 payments per day — who bears AML responsibility? PSD2 in the EU provides no mechanism for AI agents to be recognized as equivalent payment actors. Visa and Mastercard solve this via tokenized credentials remaining bound to a verified human account. Crypto rails have a structurally open problem here, which would primarily need to be solved by wallet providers.

Wallet security. If the private key of an agent wallet is compromised, an attacker has autonomous spending power — no chargeback, no fraud-detection layer, no dispute resolution. This is currently the biggest enterprise-adoption blocker. Trusted Execution Environments (TEEs) and solutions like Coinbase Agentic Wallets address this, but no solution is yet industry-standard.

Identity and trust. How does a merchant distinguish a legitimate consumer agent from a fraud bot? Visa developed the Trusted Agent Protocol for this, OpenAI in parallel developed the Agentic Commerce Protocol, and Google the Universal Commerce Protocol. These frameworks still compete — the final identity standard isn't decided. Whoever sets it will have a significant strategic advantage.

8. Conclusion and outlook

Three statements can be made with high confidence.

First: The payment systems of the next decade won't be one system but a layered architecture. Card rails (Visa Intelligent Commerce, Mastercard Agent Pay) for consumer-facing purchases, crypto rails (x402, L402) for machine-to-machine transactions and micropayments, bank-driven stablecoin initiatives for institutional settlements. No winner, but coexisting layers.

Second: HTTP 402 becomes the universal payment layer. The fact that Google, Visa, Coinbase, and Cloudflare are simultaneously embracing the same 1996 RFC standard is no coincidence — it's the only interface that works both for card tokens and stablecoins.

Third: The next bottleneck isn't payment, it's identity. Whoever sets the identity standard for AI agents — be it Visa Trusted Agent Protocol, OpenAI ACP, Google UCP, or something new — controls a significant part of the value chain.

For companies and builders, this means: whoever equips tools with programmable payment hooks today — whether via Visa tokens, USDC on Base, Lightning, or classic Stripe — is building for a world where machines are a meaningful part of buyers. The interesting question isn't which chain or which card network wins. The interesting question is which products are first to be consistently designed for machine buyers.

FAQ

Will Visa and Mastercard make crypto obsolete for agent payments? No. Card networks and crypto rails address different use cases. Visa and Mastercard are well-suited for consumer purchases at established merchants, where regulatory clarity, fraud detection, and chargeback rights are paramount. Crypto rails (x402, L402) are well-suited for micropayments, machine-to-machine API calls, and global payments without intermediaries. Both will coexist and grow in parallel.

When will I personally interact with AI-agent payments? Most likely without noticing it first. Booking platforms, comparison shopping, B2B procurement tools — in many of these, AI agents already operate behind the scenes. Visa forecasts that millions of consumers will complete AI-agent purchases by Holiday Season 2026. The shift will look unspectacular at first ("Buy now via your AI assistant") and gradually expand to more categories.

What's the biggest unsolved problem right now? Identity. The technical question of how to pay an agent is largely solved — Visa, Mastercard, x402, and Lightning all offer working solutions. The question of how a merchant can trust that an agent is legitimate, isn't a fraud bot, and is acting on behalf of a real consumer — that's still unresolved. The first robust answer to this question will become an industry standard.

Should I integrate x402 into my SaaS or API today? If you sell digital services to other businesses, the answer is increasingly yes. Implementation effort is modest — a handful of developer days — and it future-proofs you for machine-driven demand. At the same time, this isn't urgent for most B2C businesses, where Visa Intelligent Commerce and Mastercard Agent Pay are more likely to bring agent traffic first.

How does this affect existing financial infrastructure providers like Stripe or PayPal? Stripe is positioned at the center of both worlds — Shared Payment Tokens for card-rail integration, parallel AgentCore Payments cooperation with AWS for crypto-rail integration. PayPal launched its own stablecoin (PYUSD) and has integrations with most relevant agent frameworks. Both will likely keep their role — but they're no longer the only relevant payment infrastructure for agentic commerce.