How Agentic Payments Actually Work

When the first agentic payment cleared Visa's network in January 2026, most payment operations teams missed it. It looked like any other card-not-present transaction in their dashboards — a tokenized credential, a standard authorization, a T+1 settlement. But the buyer wasn't a person. It was an autonomous AI agent acting on behalf of a procurement manager at a mid-market SaaS company, executing a pre-approved software license renewal at 2:14 AM without a single human click.

Three months later, Stripe reported $14 billion in agent-initiated payment volume for Q1 2026 alone — a 340% increase from the previous quarter. If your operations team isn't already seeing these transactions in your payment mix, you will be by year-end.

We broke down the basics in our carousel — here's the full picture: who built the rails, how money actually moves in an agentic transaction, what the data says about where this is headed, and what payment operations teams need to do right now to prepare.

The Networks and Platforms That Built Agentic Rails in 2026

Agentic payments didn't emerge from a single announcement. Five major players shipped production-ready infrastructure within a six-month window, each solving a different piece of the puzzle. Understanding what each one built — and where each one falls short — matters for how you architect your payment stack.

Visa: Intelligent Commerce and Agent Tokenization

Visa's Intelligent Commerce platform, launched in March 2026, extended their existing token service to issue credentials specifically for AI agents. Each agent receives a unique token tied to a human cardholder's account but governed by a separate set of programmable controls.

The critical detail: Visa's agent tokens support autonomous purchases up to $10,000 without requiring real-time human approval. The cardholder sets the ceiling, merchant category restrictions, and velocity limits at token provisioning. The agent operates freely within those boundaries.

For merchants, Visa's approach means agent transactions arrive looking nearly identical to standard token-based purchases. Your acquirer sees a normal authorization request. But the token metadata includes an agent identifier flag (field 63.2 in Visa's specification), which lets you distinguish agent-initiated purchases from human ones in your reconciliation and analytics.

Operational takeaway: If you accept Visa and aren't parsing the agent identifier flag today, you're flying blind on what percentage of your volume is already agentic.

Mastercard: Agent Pay Credentials

Mastercard took a different architectural approach. Instead of extending existing tokens, they created Agent Pay — a dedicated 16-digit credential class issued specifically for AI agents. Think of it as a virtual card purpose-built for machines.

Agent Pay credentials carry programmable spending rules baked into the credential itself: merchant category codes, single-transaction limits, daily and monthly velocity caps, and even time-of-day restrictions. The issuing bank enforces these rules at authorization, meaning the agent can't override them regardless of how it's programmed.

This design has a meaningful security advantage. With Visa's model, the rules live in the token vault and are enforced at the network level. With Mastercard's model, the rules travel with the credential through every node in the payment chain. If a gateway or processor fails to check the agent's permissions, the issuer still catches it.

Operational takeaway: If you're a merchant processing Agent Pay credentials, these transactions will appear as a distinct BIN range. Update your BIN tables and ensure your payment orchestration layer routes them correctly.

Stripe: Agent SDK

Stripe's Agent SDK, released in February 2026, is the most developer-facing piece of the stack. It provides native API endpoints that let AI agents initiate, authorize, and confirm payments programmatically — with end-to-end latency under 200 milliseconds.

What makes Stripe's implementation noteworthy for payment operations teams is the built-in fraud scoring calibrated specifically for agent behavior. Traditional fraud models flag transactions with no browser fingerprint, no device ID, and no behavioral biometrics. Agent transactions have none of these signals by definition. Stripe retrained their Radar models on agent-specific signals: credential provenance, API call patterns, request signing consistency, and historical agent behavior graphs.

Early data suggests it's working. Stripe reports a 0.04% fraud rate on agent-initiated transactions processed through the Agent SDK — roughly half the rate of standard e-commerce.

Operational takeaway: If you're on Stripe, enabling the Agent SDK isn't just about accepting agentic payments. It's about ensuring those transactions don't get incorrectly flagged by your existing fraud rules. Merchants who haven't updated their Radar configuration are seeing false decline rates as high as 12% on legitimate agent purchases.

PayPal: Autonomous Wallets

PayPal's Autonomous Wallets solve a problem the card networks don't address well: reconciliation and budget isolation. Each AI agent gets a sub-account under the owner's PayPal business account with its own balance, spending limits, and transaction history.

This matters most for mid-market merchants who are also buyers — companies whose procurement agents are making autonomous purchases. The sub-account structure means agent spending shows up in a clean, separate ledger. No more combing through a single corporate card statement trying to figure out which purchases were human and which were machine.

For merchants receiving PayPal Autonomous Wallet payments, the transaction flow is standard PayPal — the funds arrive the same way. The difference is in dispute handling: PayPal introduced a separate dispute resolution track for agent-initiated transactions, with shorter response windows (5 business days instead of 10) and different evidence requirements.

Operational takeaway: Update your dispute response playbooks. If you receive a chargeback from an Autonomous Wallet credential, PayPal expects you to provide the agent's credential verification log and the API-level transaction confirmation — not just a shipping receipt.

Adyen: Agentic Gateway Mode

Adyen's contribution targets B2B transactions specifically. Their Agentic Gateway Mode replaces the 3D Secure authentication flow — which requires a human to interact with a browser or mobile device — with cryptographic agent verification.

Here's how it works: instead of redirecting to a 3DS challenge page, the agent presents a signed cryptographic credential (based on the FIDO2/WebAuthn standard, adapted for machine identity) directly to Adyen's gateway. Adyen verifies the signature against the agent's registered public key, confirms it matches the cardholder's authorized agent list, and processes the authorization. The entire verification step adds roughly 15 milliseconds to the transaction.

For B2B merchants, this eliminates the single biggest friction point in authenticated payments: the assumption that a human is present to complete a challenge. Adyen reports that merchants in their B2B pilot saw authorization rates improve by 8 percentage points after enabling Agentic Gateway Mode — from 87% to 95%.

Operational takeaway: If you sell B2B and require 3DS on high-value transactions, you need an agentic authentication path. Without one, every AI-initiated purchase from your buyers will fail at the 3DS challenge step.

How an Agentic Payment Actually Flows

Understanding the architecture is one thing. Watching the money move is another. Here's the real-time flow of an agentic transaction, with timing at each step.

Step 1: Agent Initiates Purchase Request — 12ms

An AI agent (in this example, running on OpenAI's GPT-5 infrastructure) determines that a purchase is needed — say, an $847.50 restock of office supplies. The agent constructs a payment request containing the cryptographic agent credential, the order details, and a signed hash of the transaction parameters. Elapsed time from decision to request transmission: 12 milliseconds.

This is the moment that matters most from a liability perspective. The agent is acting within pre-approved rules set by the human account owner. Those rules — the $10,000 spending limit, the approved merchant categories, the velocity caps — were established at credential provisioning. The agent doesn't "decide" to spend money in the way a human does. It executes within boundaries.

Step 2: Merchant Validates and Creates Payment Intent — 45ms

The merchant's API (Amazon, in this example) receives the request, validates the agent token against its registered credential database, creates the order, and sends a payment intent to its payment processor. Total time from request receipt to payment intent creation: 45 milliseconds.

For merchants, this step is where your existing order management system intersects with agentic commerce. Your API needs to accept agent credentials as a valid buyer identity and create orders without waiting for a human session. If your checkout flow assumes a browser session with cookies and a user account login, agent transactions will fail here.

Step 3: Gateway Processes and Routes — 89ms

The payment gateway (Stripe, in this example) receives the payment intent, processes the agent credential, runs fraud scoring against agent-specific models, and routes the authorization request to the card network (Visa). Cumulative elapsed time: 89 milliseconds.

At this stage, the fraud check is the critical gate. Stripe's agent-specific Radar model evaluates the transaction against the agent's historical behavior, the credential's signing integrity, and the transaction parameters relative to the agent's programmed rules. A score is generated in under 30ms.

Step 4: Issuer Approves and Settles — 134ms Total

The issuing bank (Chase) receives the authorization request, verifies that the $847.50 purchase falls within the agent's $10,000 spending limit and approved merchant categories, and returns an approval. Total time from agent initiation to issuer approval: 134 milliseconds. Settlement follows standard rails at T+1.

That's the full journey: 134 milliseconds from agent decision to bank approval. No human touched it. No browser opened. No 3DS page loaded. No CAPTCHA was solved.

The Data: $2.7 Trillion and Accelerating

Gartner projects that AI agents will autonomously handle $2.7 trillion in transaction volume by 2028. That number sounds large until you put it in context: global e-commerce hit roughly $6.3 trillion in 2025. Agentic commerce could represent over 40% of digital transaction volume within two years.

The trajectory is steep. Stripe's $14 billion in Q1 2026 agent-initiated volume, up 340% quarter-over-quarter, suggests we're in the exponential growth phase. Mastercard reported that Agent Pay credentials issued in the first 90 days exceeded total contactless card issuance in the platform's first full year. PayPal disclosed 2.1 million Autonomous Wallets created in their first quarter of availability.

For mid-market merchants, the volume shift creates both opportunity and risk. On the opportunity side: agent-initiated transactions convert at dramatically higher rates because there's no checkout friction, no session timeout, no payment form to fat-finger. On the risk side: if your payment stack can't process agent credentials, you're invisible to a rapidly growing buyer population.

Traditional Checkout vs. Agentic Payments: The Numbers

The performance gap isn't subtle.

Metric	Traditional Checkout	Agentic Payment
End-to-end completion time	47 seconds average	380ms
Failure/abandonment at payment	23% cart abandonment	0.3% transaction failure
Authentication method	Manual 3DS (human required)	Cryptographic agent verification
Human review threshold	$500+ purchases flagged	Pre-approved rules up to $10K
Fraud rate	~0.08% (CNP average)	~0.04% (Stripe Agent SDK)

The 47 seconds to 380ms gap is striking, but the cart abandonment comparison is where the revenue impact lives. A 23% abandonment rate at payment is the industry average for standard e-commerce checkout (Baymard Institute, 2025). Agent-initiated transactions see a 0.3% failure rate — and those failures are almost entirely technical (network timeouts, credential expiry), not behavioral.

For a mid-market merchant processing $50 million annually with a 23% payment-stage abandonment rate, recovering even a fraction of that lost revenue through agentic channels represents millions in recaptured sales.

The human review threshold shift is equally significant for operations teams. Most fraud teams manually review transactions above $500 — some above $250. Agent credentials with pre-approved limits up to $10,000 mean those transactions can be auto-approved with confidence, freeing fraud analysts to focus on genuinely anomalous activity. Platforms like Cellix that provide transaction-level visibility across payment channels can help operations teams distinguish agent-initiated volume from human purchases and calibrate review thresholds accordingly.

What Payment Operations Teams Should Do Now

This isn't a 2028 problem. Agent-initiated transactions are hitting your payment stack today, and the volume curve is steep. Here's what to prioritize.

1. Audit Your Current Agent Transaction Volume

Check whether your processor or gateway provides agent identifier flags in transaction metadata. If you're on Stripe, check your Agent SDK dashboard. If you're on Adyen, look for Agentic Gateway Mode transaction markers. If you're on a legacy gateway, ask your provider — you may be processing agent transactions today without knowing it.

2. Update Your Fraud Rules

Default fraud models will over-decline agent transactions. No device fingerprint, no behavioral biometrics, no browser data — these all trigger traditional fraud signals. Work with your gateway to enable agent-specific fraud models or, at minimum, whitelist verified agent credential types from your most aggressive rules.

3. Prepare Your API for Headless Purchases

If your checkout flow requires a browser session, cookies, or JavaScript rendering, agent transactions cannot complete. Build or enable API-first purchase endpoints that accept tokenized credentials and return order confirmations without a front-end session.

4. Revise Dispute and Chargeback Playbooks

Agent-initiated transactions have different evidence requirements in disputes. Maintain logs of agent credential verification, API-level transaction confirmations, and the agent's programmed spending rules. These are your primary evidence artifacts — not delivery confirmation or IP address logs.

5. Model the Revenue Impact

Calculate what a 380ms, near-zero-abandonment payment channel means for your conversion rate. If 5% of your traffic shifts to agent-initiated purchases within 12 months — a conservative estimate for B2B merchants — what does that do to your top line? Build the business case now.

Key Takeaways

• Five major payment platforms shipped production agentic rails in 2026 — Visa, Mastercard, Stripe, PayPal, and Adyen each solve different parts of the stack. Most mid-market merchants will need to support at least two of these to capture agent-initiated volume across B2C and B2B channels.

• An agentic payment clears in 134 milliseconds end-to-end, replacing a 47-second human checkout flow. The operational impact isn't just speed — it's a structural reduction in cart abandonment from 23% to a 0.3% transaction failure rate.

• Your existing fraud rules are likely declining legitimate agent transactions today. Agent purchases lack every signal traditional fraud models rely on. Updating your fraud configuration for agent-specific scoring is the single highest-ROI action you can take this quarter.

• Dispute handling for agent transactions requires different evidence artifacts. Credential verification logs and API transaction confirmations replace shipping receipts and IP geolocation as your primary chargeback defense. Update your playbooks before your first agent-initiated dispute lands.

• The volume curve is exponential, not linear. Stripe's 340% quarter-over-quarter growth in agent payment volume means the merchants who build agentic payment acceptance now will capture disproportionate share of a channel projected to reach $2.7 trillion by 2028.