MSME & Global Commerce · emerging evidence

AP2, ACP, Beckn: A Reader's Map of the Three Protocols Deciding Whether an Indian MSME Gets Bought by an AI Agent

Last reviewed 2026-08-09. Written by Chandranshu Kumar, Founder, Raveneye Global. · 12 min read

Any AP2 vs Beckn agentic commerce protocol comparison has to begin with a structural fact: three separate protocols now govern whether an AI agent can find, trust, and buy from a business, and they were built by three different camps that do not currently interoperate. Google's Agent Payments Protocol (AP2), announced in September 2025 with more than sixty partners, is a payments trust layer that gives a merchant cryptographic proof that a human authorized an agent to spend. The Agentic Commerce Protocol (ACP), maintained by OpenAI and Stripe and live inside ChatGPT Instant Checkout, is a checkout protocol that lets an agent complete a purchase with a merchant using a delegated payment token. Beckn, stewarded by India's FIDE foundation and the backbone of ONDC, is an open discovery-and-transaction layer that lets buyers and sellers meet without a single platform in the middle. For an Indian MSME the practical point is direct: being listed on ONDC makes the business legible to the Beckn network, but not automatically to an agent shopping through ChatGPT or a Google-orchestrated payment flow. Which surfaces can read a business is now a question that has to be answered deliberately.

Three protocols, three different jobs

The phrase "agentic commerce" hides a design problem that has quietly split into three. Before an AI agent can buy something on a buyer's behalf, three separate things have to happen. The agent has to discover a business and its offer. The agent has to check out, agreeing on the exact cart, price, tax, and shipping with the seller. And the merchant has to trust the payment, meaning it needs some proof that a real human actually authorized this agent to spend this money. Each of the three protocols examined here is built around a different one of those jobs.

That distinction is the first thing to establish, because the headlines treat AP2, ACP, and Beckn as competitors racing for the same finish line. They are not really competing on the same track. AP2 is mostly about the trust and payment step. ACP is mostly about the checkout step inside a specific agent surface. Beckn is mostly about the discovery-and-transaction step across an open network. They overlap at the edges, and each camp would like to own more of the stack over time, but at their core they answer different questions.

The second point is that they were built by three different power centers with different incentives, and as of this writing they do not interoperate. There is no announced bridge that lets a seller implement one and automatically transact through the others. For a large brand with an engineering team, that is an integration cost. For a micro or small enterprise in India, it is something more consequential: a decision about which doors to the agent economy are open to the firm, made largely by which protocol its storefront or channel happens to speak.

AP2: the permission slip for spending

Google announced the Agent Payments Protocol on September 16, 2025, with more than sixty payments and technology partners named at launch, including Mastercard, American Express, PayPal, Adyen, Coinbase, Worldpay, UnionPay International, JCB, Intuit, Salesforce, ServiceNow, and Etsy. Its job is narrow and important: to solve the trust gap that opens the moment a human is no longer clicking the buy button themselves.

AP2 builds that trust with what it calls Mandates, described in Google's announcement as tamper-proof, cryptographically signed digital contracts that serve as verifiable proof of a user's instructions, signed by verifiable credentials. There are two. An Intent Mandate captures what the user asked for and sets the boundaries of the task. A Cart Mandate is signed once the agent has assembled the actual cart, creating what Google calls a secure, unchangeable record of the exact items and price. In plain terms, AP2 is the permission slip and the receipt: the merchant gets a signed guarantee that the human approved this specific purchase, so it can accept an agent-initiated payment with the same confidence it would a human one.

Two design choices matter for the rest of this map. First, AP2 is deliberately payment-method agnostic. Google frames it as spanning credit and debit cards, real-time bank transfers, and stablecoins, and a companion A2A x402 extension, built with Coinbase, the Ethereum Foundation, and MetaMask, adds a crypto rail. Second, AP2 is not a discovery or checkout protocol on its own. It builds on Google's Agent2Agent (A2A) messaging framework and can extend both A2A and the Model Context Protocol (MCP). It assumes some other layer already found the product and built the cart. It governs the moment money moves.

ACP: checkout inside the agent

The Agentic Commerce Protocol answers a different question: how does an agent actually complete a purchase with a merchant, end to end, without the buyer ever leaving the chat window? It was co-developed by OpenAI and Stripe, announced on September 29, 2025, and released as an open specification under the Apache 2.0 license. The GitHub repository lists OpenAI and Stripe as the maintainers, and the spec uses date-based versioning, with the first public version dated 2025-09-29 and later revisions through 2026.

ACP is not a whiteboard proposal. It is the technology underneath ChatGPT Instant Checkout, which OpenAI launched the same day, letting United States users buy directly inside ChatGPT from United States Etsy sellers first, with more than a million Shopify merchants such as Glossier, SKIMS, Spanx, and Vuori to follow. At launch it supported single-item purchases and was free to users, though OpenAI has said it takes a fee from completed transactions. The mechanism that makes this work is a delegated payment token. Rather than handing raw card details to the agent, ACP passes a scoped token that the merchant's payment provider can charge. Stripe's documentation describes token properties designed to limit blast radius: single-seller scope, so a token issued to one merchant cannot be replayed at another, a fixed amount cap set at issuance, and a short expiry measured in minutes.

The strategic shape of ACP is worth stating plainly. It is a checkout protocol that currently lives inside one dominant agent surface. A merchant that implements ACP becomes buyable by the agents that speak it, which today means ChatGPT. That is a large and growing front door, but it is a specific one, governed by two private companies. It is not the same thing as being buyable by every agent everywhere, and it is not the same layer as AP2's payment mandates, even though both touch payment. AP2 proves authorization across many rails; ACP runs a checkout and moves a token within a particular commerce experience.

Beckn and ONDC: discovery without a platform

Beckn comes from a different tradition entirely, and it starts one step earlier in the buying journey. It is an open protocol stewarded by FIDE, the Foundation for Interoperability in the Digital Economy, and it was conceived by Nandan Nilekani, Pramod Varma, and Sujith Nair, the same lineage of thinking that produced India's public digital infrastructure. Its specifications are published openly on GitHub. Its purpose is to let discovery and transaction happen across an open, decentralized network, so a buyer's request can be broadcast to many sellers without any single search engine or marketplace sitting in the middle as the gatekeeper.

The Open Network for Digital Commerce (ONDC), a government-backed initiative in India, is built on Beckn. By 2025 ONDC reporting described a network spanning hundreds of cities across many commerce domains, from grocery and food delivery to mobility and fashion, with its chief executive publicly targeting tens of millions of monthly transactions. The point of the architecture is unbundling: a seller can list once on the network and be discoverable by any compliant buyer application, rather than negotiating separately with each walled-garden platform. For a small Indian seller, that is the promise, one connection to a shared network instead of many.

Beckn's advocates now frame it explicitly as agent infrastructure. Sujith Nair of FIDE has described Beckn as a standardized language of transactions, a concise set of specifications that lets two parties, increasingly two software agents, form machine-readable, digitally signed micro-contracts in real time. FIDE has also demonstrated BecknGPT, an agent that manages transactions across the ONDC architecture. The thesis is that when a buyer's agent and a seller's agent need to agree on terms and liability, they need a neutral contracting protocol, and Beckn is positioned to be it. Where ACP is checkout inside one company's agent and AP2 is authorization for payment, Beckn wants to be the open street the agents meet on.

The part nobody advertises: they do not interoperate

Read the three announcements side by side and a gap becomes obvious. Each protocol is described as open, each has an impressive partner list, and each implies it is the foundation of agentic commerce. What none of them claims is interoperability with the other two. As of this writing there is no announced bridge that lets a merchant implement Beckn and thereby transact through ACP's ChatGPT checkout, or wire AP2 mandates into an ONDC flow. They are three overlapping standards governed by three different centers of gravity: a United States platform company (Google) for AP2, two United States companies (OpenAI and Stripe) for ACP, and an India-based foundation with government backing (FIDE and ONDC) for Beckn.

This is a familiar pattern in the history of standards. When a valuable new layer of the economy is up for grabs, several credible bodies each publish an open specification, and open does not mean unified. Openness lowers the cost of adopting any one of them; it does nothing, by itself, to make them speak to each other. The most likely near-term outcome is not a single winner but a period of parallel networks, with adapters and translation layers eventually stitched between them by third parties who profit from the seams. That stitching is not here yet.

For a business, the consequence is that reach in the agent economy is currently partitioned by protocol. Being fully legible to one network does not carry over to the others. A brand that is buyable inside ChatGPT via ACP is not, by that fact, buyable through a Beckn-native Indian app, and a seller thriving on ONDC is not, by that fact, discoverable to a Google-orchestrated AP2 payment flow. The doors are separate, and each one has to be opened on purpose.

What this means for an ONDC-only Indian MSME

Now put an actual small Indian business in the middle of this map: a kirana store, a regional apparel label, a services firm that has done the responsible thing and joined ONDC. On the Beckn network, that business is legible. Its catalog is structured, its offers are discoverable, and an agent operating inside the ONDC world can find and transact with it. That is real progress, and it is more open than being trapped inside a single marketplace.

But the agent economy that Western press writes about, and that a growing share of cross-border and premium demand flows through, largely runs on the other two protocols. The ChatGPT shopper using Instant Checkout is transacting over ACP. The Google-orchestrated agent settling a payment is leaning on AP2 mandates. An ONDC-only seller is, for those surfaces, effectively invisible unless it also does the separate work of becoming legible to them. The uncomfortable version of the sentence is this: joining ONDC answers the Indian open-network question and leaves the global-agent question untouched.

None of this argues against ONDC. The open-network model is strategically sound, and for domestic reach it is a genuine advantage. The argument is narrower and more practical. Protocol coverage is now a distinct dimension of whether a business gets bought, separate from having a website and separate from ranking in search. An owner who assumes that one integration covers the agent economy is making the same mistake that owners made when they assumed a website covered being found online. It covers one surface. The others have to be measured and worked deliberately.

The layer beneath all three protocols

There is a foundation that every one of these protocols quietly depends on, and it is one an MSME can actually control. Before an agent can check out with a business, authorize a payment to it, or contract with it over an open network, the agent has to be able to read the business: to parse who it is, what it sells, at what price, with what credibility, in a form a machine can trust. Structured product data, consistent identity across the places the business appears, clear and corroborated pricing and reputation. That machine-readable substrate is upstream of AP2, ACP, and Beckn alike.

This is the more durable investment, because it does not require betting on which protocol wins. A business that is genuinely legible to machines is easier to onboard to whichever networks matter for its market, and easier for any future adapter to bridge. The protocols are moving fast and may consolidate in ways no one can predict. The requirement to be readable will outlast the specific standards, which is exactly why it deserves attention before the standards war resolves.

How to read the evidence

The temptation with a topic this new is to overclaim in either direction, to announce that agents have already changed everything or to dismiss the whole category as hype. The measured reading is in between. The protocols are real, backed by serious partners, and already handling live transactions in the case of ACP. The market forecasts around them, including McKinsey's October 2025 estimate that agentic commerce could redirect three to five trillion dollars of global retail spend by 2030, are genuine projections from credible firms, and they are still projections, not measured outcomes. Treat the direction as real and the magnitudes as uncertain.

For an Indian MSME, the action items are modest and concrete. The first is to understand that ONDC and the global agent surfaces are governed by different, non-interoperating protocols, so coverage on one is not coverage on all. The second is to invest first in the machine-readable foundation that sits beneath every protocol, because it is the part the business controls and the part that survives whichever standard wins. The third is to measure where the business actually stands, rather than assuming, because the surfaces that now decide whether an agent finds and buys from a firm are not the surfaces its normal analytics were built to show.

That measurement discipline is the practical payoff of this map. It asks the useful question: which agent-buying surfaces can currently read the business, which cannot, and what it would take, in order of return, to open the doors that matter for its market.

The evidence

Key findings, with their sources

Google announced the Agent Payments Protocol (AP2) on September 16, 2025, with more than sixty payments and technology partners including Mastercard, American Express, PayPal, Adyen, Coinbase, and Worldpay.

established Google Cloud Blog, "Announcing Agent Payments Protocol (AP2)", September 2025, cloud.google.com.
AP2 uses cryptographically signed Mandates (an Intent Mandate and a Cart Mandate), backed by verifiable credentials, to prove a human authorized an agent's purchase; it is payment-method agnostic across cards, bank transfers, and stablecoins.

established Google Cloud Blog, "Announcing Agent Payments Protocol (AP2)", September 2025, cloud.google.com.
The Agentic Commerce Protocol (ACP) is maintained by OpenAI and Stripe, licensed under Apache 2.0, uses date-based versioning, and was first published on 2025-09-29.

established Agentic Commerce Protocol GitHub repository (agentic-commerce-protocol), maintained by OpenAI and Stripe.
ChatGPT Instant Checkout launched September 29, 2025, powered by ACP, letting United States users buy from United States Etsy sellers first, with more than a million Shopify merchants (Glossier, SKIMS, Spanx, Vuori) to follow; it supports single-item purchases and OpenAI takes a fee on completed transactions.

established OpenAI, "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol", September 2025, openai.com; Stripe blog, "Developing an open standard for agentic commerce".
ACP passes a delegated payment token with single-seller scope, a fixed amount cap set at issuance, and a short expiry measured in minutes, so the token cannot be replayed at another merchant or grown.

established Stripe Documentation, "Agentic Commerce Protocol (ACP)", docs.stripe.com.
Beckn is an open protocol stewarded by FIDE (Foundation for Interoperability in the Digital Economy), conceived by Nandan Nilekani, Pramod Varma, and Sujith Nair; ONDC is built on Beckn.

established Analytics India Magazine interview with Sujith Nair (FIDE), 2025; Beckn protocol-specifications GitHub.
By May 2025 ONDC reporting described a network spanning hundreds of cities across many commerce domains, with its chief executive publicly targeting 30 to 40 million monthly transactions.

contested Outlook Business, "Expect 30-40 Million Monthly Transactions by March 2025: ONDC CEO", 2025.
McKinsey estimated in October 2025 that agentic commerce could redirect three to five trillion dollars of global retail spend by 2030, and up to one trillion dollars of United States orchestrated retail revenue.

contested McKinsey & Company, October 2025, reported via Digital Commerce 360, "McKinsey forecasts up to $5 trillion in agentic commerce sales by 2030".
The three protocols do not currently interoperate: there is no announced bridge that lets a merchant implement one and transact through the others.

emerging Analysis by Raveneye Global of the published AP2, ACP, and Beckn specifications and announcements, 2025-2026.

Calibration

What is proven, what is promising, what is unproven

Evidence tier	Tactics	What the evidence says
established	The existence, ownership, launch dates, licensing, and core mechanics of AP2 (Google, signed mandates), ACP (OpenAI and Stripe, delegated token, live in ChatGPT Instant Checkout), and Beckn (FIDE, the backbone of ONDC).	Google Cloud Blog (2025); Agentic Commerce Protocol and Beckn GitHub repositories; Stripe and OpenAI documentation; FIDE interviews.
emerging	The reading that the three protocols occupy different layers (discovery, checkout, payment trust) and do not interoperate today, so agent-economy reach is partitioned by protocol, and that a machine-readable foundation sits upstream of all three.	Comparative analysis of the published specifications; no interoperability bridge has been announced by any of the three governing bodies as of this writing.
contested	The market-size forecasts (McKinsey trillions by 2030; industry AI-spend projections) and the specific ONDC transaction targets used to size the opportunity.	Named-firm projections and executive statements; genuine but inherently uncertain forecasts, not measured outcomes.

Reference

Glossary

AP2 (Agent Payments Protocol): Google's open protocol, announced September 2025, that gives a merchant cryptographic proof (signed Mandates backed by verifiable credentials) that a human authorized an agent to make a specific purchase. It is payment-method agnostic and governs the trust-and-payment step, not discovery or checkout.
ACP (Agentic Commerce Protocol): An open checkout standard maintained by OpenAI and Stripe under Apache 2.0, live inside ChatGPT Instant Checkout. It lets an agent complete a purchase with a merchant using a scoped, short-lived delegated payment token, without the buyer leaving the chat.
Beckn: An open, decentralized protocol stewarded by India's FIDE foundation that lets discovery and transaction happen across a shared network rather than inside one platform. It is the backbone of ONDC and is now positioned as neutral contracting infrastructure for AI agents.
ONDC: The Open Network for Digital Commerce, a government-backed Indian initiative built on Beckn, where a seller can list once and be discoverable by any compliant buyer application instead of negotiating with each marketplace separately.
Delegated payment token: A scoped credential, used by ACP, that lets a merchant charge a payment without the agent ever holding raw card details. It is limited to one seller, capped at a fixed amount, and expires within minutes.
Machine legibility: The degree to which a business's identity, offering, price, and credibility are structured and corroborated in forms a machine can read and trust. It sits upstream of AP2, ACP, and Beckn: an agent must be able to read you before any protocol can transact with you.

Straight answers

Frequently asked questions

What is the difference between AP2, ACP, and Beckn?

They solve different steps of agent buying. AP2, from Google, is a payments trust layer that proves a human authorized an agent to spend. ACP, from OpenAI and Stripe, is a checkout protocol that completes a purchase inside an agent such as ChatGPT using a scoped payment token. Beckn, stewarded by India's FIDE and the basis of ONDC, is an open discovery-and-transaction network that lets buyers and sellers meet without a single platform in the middle.

Do AP2, ACP, and Beckn work together?

Not today. Each is published as an open specification, but open does not mean unified. As of this writing there is no announced bridge that lets a business implement one and automatically transact through the other two. They are governed by three different centers of gravity, so reach in the agent economy is currently partitioned by protocol.

If a business is on ONDC, can an AI agent in ChatGPT buy from it?

Not automatically. Being on ONDC makes a business legible to the Beckn network, which is a real advantage for domestic open-network reach. ChatGPT Instant Checkout, however, runs on ACP, a separate protocol, so becoming buyable there is separate work. One integration does not cover the whole agent economy.

Which protocol should an Indian MSME prioritize?

The evidence suggests investing first in the layer beneath all of them: being machine-readable, with structured product data, consistent identity, and corroborated pricing and reputation. That foundation is what every protocol depends on, it is the part a business controls, and it survives whichever standard wins. Protocol coverage itself should follow where the market's demand actually flows.

Is agentic commerce actually large enough to plan for?

The direction is credible and the magnitudes are uncertain. ACP is already handling live transactions inside ChatGPT, and McKinsey projected in October 2025 that agentic commerce could redirect three to five trillion dollars of global retail spend by 2030. Treat that as a serious forecast rather than a measured fact, and plan for the trend without betting the business on a specific number.

Provenance

Sources

Google Cloud Blog, "Announcing Agent Payments Protocol (AP2)", September 2025 (established)cloud.google.com
Digital Commerce 360, "Google launches payments protocol for AI commerce, names dozens of partners", September 2025 (established)digitalcommerce360.com
Agentic Commerce Protocol, official GitHub repository, maintained by OpenAI and Stripe (established)github.com
Stripe, "Developing an open standard for agentic commerce", September 2025 (established)stripe.com
Stripe Documentation, "Agentic Commerce Protocol (ACP)" (established)docs.stripe.com
OpenAI, "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol", September 2025 (established)openai.com
Beckn Protocol, official protocol-specifications repository, FIDE (established)github.com
Analytics India Magazine, "When Two AI Agents Communicate, Beckn Can be the Contracting Infrastructure" (interview with Sujith Nair, FIDE) (established)analyticsindiamag.com
Outlook Business, "Expect 30-40 Million Monthly Transactions by March 2025: ONDC CEO", 2025 (contested, executive projection)outlookbusiness.com
Digital Commerce 360, "McKinsey forecasts up to $5 trillion in agentic commerce sales by 2030", October 2025 (contested, forecast)digitalcommerce360.com

Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.