The networks are building the rails. The protocols are defining how agents talk. But who manages the merchant side? That gap just got its first serious answer.

The agentic commerce stack has been filling in from the top down. Visa launched Intelligent Commerce Connect. American Express shipped its Agentic Commerce Kit. Stripe published the Machine Payments Protocol. Google, OpenAI, and a growing list of protocol designers have staked out how AI agents should discover products, authenticate identity, and settle transactions.

All of that matters. None of it answers a straightforward question: once an AI agent shows up at a merchant's front door, who routes the transaction?

Gr4vy, a cloud-native payment orchestration platform, thinks the answer is an orchestration layer built specifically for agent-initiated commerce. The company just launched its Agentic Development Kit, and the timing tells us something about where the stack is heading.

The agentic commerce conversation has focused on rails and protocols. The merchant-side orchestration layer, the middleware that actually routes, retries, and manages agent transactions across multiple payment methods, has been largely absent.

What Gr4vy Actually Built

Gr4vy's Agentic Development Kit (ADK) is designed to help merchants build and deploy AI-native storefronts inside conversational AI platforms, including ChatGPT. The kit runs on the Model Context Protocol (MCP), which means it plugs into the same infrastructure that AI agents already use to interact with external tools and services.

The architecture works like this. Each merchant gets a single-tenant Gr4vy instance with an MCP server deployed on top. That server acts as the front end for AI agents, handling interactions between the agent, the merchant's inventory, and Gr4vy's orchestration layer underneath. Once connected, merchants can process transactions through their existing payment stack or tap into more than 400 payment methods and payment service providers through Gr4vy's network.

"AI is quickly becoming part of the checkout journey," said John Lunn, Founder and CEO of Gr4vy. "We're already enabling agentic payments inside ChatGPT today. The Agentic Development Kit is the next step, providing merchants with a structured way to adopt this model."

This is not Gr4vy's first move in the space. The company launched an Alpha MVP for agentic payment orchestration last year in collaboration with Google's Agent Payments Protocol (AP2). That earlier release included three components: an agentic shopping layer for product discovery and transactions, an inventory management bridge for real-time stock syncing, and a wallet layer built on Google's AP2 Credentials Provider framework using Gr4vy's vaulting infrastructure.

The ADK builds on that foundation. Where the Alpha MVP proved the concept, the ADK packages it for merchant adoption.

The Orchestration Gap

To understand why this matters, consider the layers of the agentic commerce stack as we have mapped them over the past quarter.

At the top, you have the agent layer. OpenAI's Operator, Google's Gemini agents, and dozens of vertical shopping agents that browse, compare, and purchase on behalf of consumers. Below that sit the protocols: Visa's Trusted Agent Protocol, Stripe's Machine Payments Protocol, Google's Universal Commerce Protocol, Coinbase's x402, and others defining how agents identify themselves, request payment authorization, and settle transactions.

Below the protocols sit the networks and processors. Visa, Mastercard, and Amex are extending their rails to support agent-initiated transactions. Processors like Fiserv and FIS are building the interfaces that connect those rails to acquiring infrastructure.

But between the protocols and the merchant, there is a layer that nobody was talking about until recently. The orchestration layer. The merchant-side middleware that takes an inbound agent transaction and decides which payment method to use, which processor to route through, which fraud rules to apply, and what to do when the first attempt fails.

For traditional e-commerce, payment orchestration platforms have handled this for years. Companies like Gr4vy, Spreedly, and Primer built businesses around giving merchants a single integration point for managing multiple payment providers. The pitch was straightforward: one API, many processors, smart routing, and no vendor lock-in.

Agentic commerce introduces new complexity. Agent transactions do not look like a consumer clicking "pay now." They arrive programmatically, often without a browser session, sometimes in rapid succession, and frequently across multiple merchants in a single shopping flow. The orchestration layer needs to handle tokenized credentials passed by an agent, apply merchant-specific spend controls, route based on agent identity and trust level, and retry across different payment rails when transactions fail.

That is a different problem than routing a credit card payment to the cheapest processor.

Why Merchants Cannot Just Wait for the Networks

Visa's Intelligent Commerce Connect is impressive. It provides a unified system for businesses to connect with AI agents through a single integration, handling secure payments, tokenization, spend controls, and authentication. Visa says it supports both Visa and non-Visa cards and is building compatibility with every major agentic protocol.

Amex is leaning into its closed-loop advantage. Because it is simultaneously a card issuer, payment network, and merchant acquirer, Amex argues it can verify agent identity end to end without relying on external trust frameworks. The company says it has already processed thousands of test agentic transactions.

Both of these are network-level capabilities. They define how transactions travel through the rails. They do not solve the merchant's problem.

A mid-market retailer running Shopify with three payment processors, two fraud screening tools, and a buy-now-pay-later provider does not need another protocol. It needs a way to plug agent transactions into its existing stack without rebuilding everything. It needs routing logic that understands which processor handles which markets, which fraud rules apply to agent-initiated versus human-initiated transactions, and how to fall back when a provider is down.

According to a report commissioned by Visa Acceptance Solutions, over 50 percent of acquirers cite integration cost as the primary obstacle to merchant deployment of agentic commerce. The recommended approach is not rip-and-replace. It is layering APIs, middleware, and orchestration platforms on top of existing systems to modernize around legacy infrastructure.

That is exactly where payment orchestration lives.

The Demand Signal Is Real

The question of whether agentic commerce will generate meaningful transaction volume is moving past theoretical debate.

Morgan Stanley's AlphaWise consumer survey found that 23 percent of American consumers have made at least one AI-assisted purchase in the past month. A separate Boston Consulting Group report projects that AI-driven shopping agents could influence more than $1 trillion in e-commerce spending, representing roughly half of today's total online commerce. BCG also found that 81 percent of consumers expect to shop using agentic AI, with adoption highest among households with children at 93 percent.

Lunn expects agentic payments to account for a high-single-digit or low-double-digit percentage of purchases this year. That is a pragmatic forecast. Even at the low end, it represents billions of dollars in transaction volume that needs to flow through merchant payment infrastructure designed for human-initiated checkout.

The early adoption curve looks like what we saw with mobile commerce a decade ago. The first use cases are routine: groceries, household supplies, repeat purchases where brand preference is established and risk is low. As we explored in our analysis of how AI agents actually shop and pay, the more complex transactions, those involving price negotiation, multi-vendor comparison, or credit decisions, require infrastructure that barely exists yet.

Orchestration is part of that infrastructure. Without it, every merchant builds its own bespoke integration with every protocol and every network. That does not scale.

Where Gr4vy Fits in the Stack

Gr4vy's positioning is specific. It is not building a protocol. It is not extending network rails. It is building the merchant-side middleware that connects protocols and rails to the merchant's existing payment infrastructure.

The ADK's MCP-based architecture is a deliberate choice. MCP has become the de facto standard for how AI agents interact with external tools, and by building on MCP, Gr4vy ensures that any agent already using the protocol can interact with a merchant's payment stack without custom integration work.

The single-tenant deployment model matters too. Each merchant gets its own Gr4vy instance, which means transaction data and payment credentials stay within the merchant's control. In an environment where AI agents pass tokenized credentials between merchants and payment providers, data isolation is not a nice-to-have. It is a compliance requirement.

Gr4vy has raised $27 million to date and is backed by investors including March Capital, Nyca Partners, and Activant Capital. The company's founder, John Lunn, previously held leadership roles at CyberSource (acquired by Visa for $2 billion in 2010) and PayPal, where he built the Developer Relations team and contributed to the Braintree acquisition. That payments infrastructure pedigree is relevant here. Orchestration is plumbing work, and Lunn has spent two decades building plumbing.

The Broader Pattern

Gr4vy is not the only company recognizing this gap. Shopify has been building agentic storefronts that connect directly to ChatGPT and Gemini. Stripe launched both the Machine Payments Protocol and its own agent-ready payment infrastructure. Adyen has been quietly adding agent transaction support to its unified commerce platform.

The pattern is clear. The stack is filling in from both ends. Networks and protocol designers are working down from the transaction rails. Commerce platforms and orchestrators are working up from the merchant's existing infrastructure. The two will meet in the middle, and the companies that control the orchestration layer will have significant influence over how agent transactions actually flow.

This resembles what happened in traditional e-commerce during the 2010s. Payment gateways commoditized. The value shifted to orchestration, routing intelligence, and merchant control. Companies like Adyen and Stripe won by giving merchants a single integration point with smart routing underneath.

Agentic commerce is setting up the same dynamic, but with higher stakes. When an AI agent initiates a transaction, the orchestration layer does not just route a payment. It determines whether the agent's credentials are valid, which fraud model to apply, whether to retry on a different rail, and how to report the transaction back to the agent in a format it can process. The orchestration layer becomes, as one industry analysis put it, the new operational perimeter for payments security.

What to Watch

Three things will determine whether payment orchestration becomes the critical middleware layer for agentic commerce or gets absorbed into the networks themselves.

First, protocol fragmentation. Right now, we have Visa's Trusted Agent Protocol, Stripe's MPP, Google's AP2 and UCP, OpenAI's ACP, and x402 from Coinbase. If these converge on shared standards, the orchestration layer becomes simpler. If they fragment, merchants will need orchestration just to manage protocol translation, exactly the kind of complexity that favors independent platforms like Gr4vy.

Second, network ambition. Visa's Intelligent Commerce Connect already supports non-Visa cards and multiple protocols. If the networks decide to own the full merchant integration stack, independent orchestrators face an existential challenge. If the networks stay focused on rails, the middleware market stays open.

Third, merchant demand velocity. The Morgan Stanley and BCG numbers are bullish. But agentic commerce still depends on consumers trusting AI agents with their money. If adoption follows a gradual curve, merchants will have time to retrofit. If it accelerates, the merchants without an orchestration layer will scramble.

Sources

The networks own the rails. The protocols define the language. But the merchant still needs someone to manage the conversation. As agentic commerce scales, will orchestration become the most valuable layer in the stack, or will the networks swallow it whole?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.