From Sanskriti Khandelwal | Product & Market Analysis

Agentic Commerce Protocol and B2B SaaS: 15 Purchase Requirements Checked

On this page

The Agentic Commerce Protocol can now carry a purchase order number and net 30 terms, but it still has no way to express a monthly subscription. We checked 15 B2B purchase requirements against the 2026-04-17 version of the spec that OpenAI and Stripe maintain. Six are supported, four are partial and five are missing, and the missing ones are the ones a SaaS checkout depends on.

Key takeaways

  • The Agentic Commerce Protocol has no recurring billing object. The 2026-04-17 checkout schema contains no billing interval, renewal or trial field, and the word "subscription" appears once, in marketing consent.
  • The spec has moved toward business buyers. Company name, tax ID, tax exemption certificates, purchase order numbers, net 15 to net 90 terms and a pending approval status all exist in 2026-04-17. None existed in the 2025-09-29 release.
  • Delegated payment is still a one-time card charge. The payment spec allows exactly one credential type, card, and one allowance reason, one_time, capped by a maximum amount and an expiry.
  • ChatGPT itself no longer runs the checkout. OpenAI said in March 2026 it would let merchants use their own checkout, and its merchant page now says purchases are completed on merchant-owned websites or apps.
4 to 11Fields on the Buyer object between the first ACP release and 2026-04-17, including company and tax ID. Source: ACP spec on GitHub.
0Billing interval, renewal or trial fields in the 2026-04-17 checkout spec. Source: ACP OpenAPI checkout file, Zan Digital field search.
1Credential type allowed by ACP delegated payment: card, with a one_time allowance only. Source: ACP delegate payment spec, 2026-04-17.

What the Agentic Commerce Protocol actually specifies

The Agentic Commerce Protocol, or ACP, is an open specification for how an AI agent, a merchant and a payment provider complete a purchase. Stripe published it on 29 September 2025, the same day OpenAI launched Instant Checkout in ChatGPT. It is licensed under Apache 2.0.

The repository README says the specification is maintained by OpenAI and Stripe and is "currently in beta". Stripe's own documentation lists Meta as a third creator, which no other page we read repeats. Treat the governance as a two-company project with an open licence, not as a standards body.

Your team needs to understand one design choice before reading anything else. ACP keeps the merchant as the merchant of record. The agent collects the buyer's intent and a payment credential, then calls the seller's own endpoints to create, update and complete a checkout session.

Three moving parts

The spec has three pieces that matter for a SaaS seller. A product feed tells the agent what you sell. A checkout API lets the agent create a session, update it and complete it. A delegated payment API turns the buyer's card into a scoped token your payment provider can charge.

Everything in the capability table below comes from those three documents. If a requirement has no field in any of them, an agent following the protocol has no standard way to ask for it, show it or agree to it.

The spec ships as dated versions. The README lists 2025-09-29 as the initial release, then 2025-12-12 for fulfillment, 2026-01-16 for capability negotiation, 2026-01-30 for extensions and discounts, and 2026-04-17 for cart, feed, orders, authentication and MCP. The 2026-04-17 changelog marks the 2026-01-30 version as deprecated.

That pace is the first thing a buyer-side reader should note. Any capability table, including ours, is a snapshot of a beta document. We read the latest stable version on 12 November 2026 and compared it field by field with the first release.

ChatGPT checkout today: the surface moved, the protocol stayed

The question in most founders' heads is narrower than the spec. Can a buyer finish a purchase of your product inside ChatGPT? As of November 2026, the answer from OpenAI's own merchant page is no, for anyone.

The ChatGPT merchants page states that OpenAI is "moving away from a standalone Instant Checkout experience in ChatGPT and prioritizing better shopping discovery and merchant-owned checkout experiences". It adds that users "can discover and evaluate products in ChatGPT, but purchases are completed on merchant-owned websites or apps".

What OpenAI said in March 2026

The retreat was reported first by The Information in early March 2026, according to Modern Retail's 27 March account. OpenAI then confirmed the change in a blog post dated 24 March 2026. Its stated reason, quoted by Modern Retail and Retail Gazette: "We've found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide".

An OpenAI spokesperson told Modern Retail on 12 March that Instant Checkout "is moving to Apps". The same spokesperson said ChatGPT search and product discovery were the priority, "with the Agentic Commerce Protocol serving as the infrastructure". So the protocol survived the product it was launched with. That distinction is the whole frame for this post.

Two other facts from the same merchant page matter for your planning. Shopping "is currently live for ChatGPT users in the U.S." only. And "there are no fees on purchases that start in ChatGPT," which removes the per-transaction cost argument against participating.

The protocol kept shipping after the checkout surface was pulled. ACP spec versions (above the line) and ChatGPT checkout events (below), Sept 2025 to Apr 2026. 29 Sep 2025 ACP 1st release Instant Checkout launches 12 Dec fulfillment 16 and 30 Jan 2026 capabilities, extensions, discounts Early to 24 Mar 2026 OpenAI moves purchases to merchant checkout 17 Apr 2026 cart, orders, B2B fields Sources: ACP GitHub README and changelog; Modern Retail, 12 and 27 March 2026; chatgpt.com/merchants.
The most B2B-capable version of the spec arrived three weeks after ChatGPT stopped completing purchases itself. The protocol and the ChatGPT surface now need to be judged separately.

The B2B capability table: 15 requirements against the 2026-04-17 spec

We listed the things a B2B software purchase usually needs to carry, from first payment to renewal. Then we searched the 2026-04-17 OpenAPI checkout file, its JSON schema and the delegated payment file for a field that expresses each one. A requirement is "supported" when a named field exists for it, "partial" when something exists but cannot carry the full requirement, and "not supported" when nothing does.

This is a reading of the spec, not a test of a live integration. Whether ChatGPT or any other agent renders a given field is a separate question, covered in the limitations section.

B2B purchase requirements in the Agentic Commerce Protocol, version 2026-04-17
RequirementVerdictWhat the spec contains
Company identitySupportedBuyer.company (CompanyInfo) with name, department and cost_center; account_type includes business.
Tax IDSupportedCompanyInfo.tax_id, "Business tax identification number". No separate VAT field.
Tax exemptionSupportedBuyer.tax_exemption with certificate_id, certificate_type (resale, exempt_organization, government) and expires_at.
Purchase order numberSupportedPaymentData.purchase_order_number, which can stand in for a payment instrument.
Net payment termsSupportedPaymentData.payment_terms: immediate, net_15, net_30, net_60, net_90, plus due_date.
Digital delivery and licence keySupportedDigital fulfillment type; order delivery carries access_url, license_key and expires_at.
Seat countPartialLineItem.quantity is an integer, minimum 1. No seat or licence-count concept, and the request Item object has no quantity field.
Quotes and negotiated pricePartialquote_id and quote_expires_at on the session. No object for negotiating or approving the price itself.
Approval before purchasePartialpending_approval status and an approval_required flag. No approver identity or multi-step chain.
InvoicePartialOrderConfirmation.invoice_number, "Invoice number if generated". No invoice object or delivery.
Recurring subscriptionNot supportedNo billing interval, period or renewal field anywhere in the checkout schema.
Free trialNot supportedZero hits for trial in the checkout schema.
Usage-based or metered priceNot supportedPrices are a fixed unit_amount in minor currency units.
Non-card payment (ACH, wire)Not supportedDelegated payment allows one credential type: card.
Renewal or plan change after purchaseNot supportedCancel exists for a checkout session only; nothing addresses a contract after the order.

Source: ACP OpenAPI and JSON schema files for 2026-04-17, read in full on 12 November 2026. The net terms row is the most generous verdict here. The field exists, but the spec does not say who extends credit or how an unpaid invoice is collected.

How to read the split

Six of 15 is 40% supported. Count the partial rows as half and the spec covers about 53% of the list. That sounds respectable until you look at which rows sit where.

The supported rows describe who the buyer is and how a one-off order is paid. The missing rows describe the relationship after the first payment. For a hardware reseller that is a minor gap. For a SaaS company whose revenue is the relationship after the first payment, it is the gap.

15 B2B purchase requirements, ACP version 2026-04-17. Verdict per requirement, from a field-by-field read of the checkout and payment specs. 15 requirements 6 supported (40%) company, tax ID, exemption, PO, net terms, digital delivery. 4 partial (27%) seats, quotes, approvals, invoice 5 not supported (33%) subscription, trial, metered price, non-card pay, renewal. Source: ACP OpenAPI checkout and delegate payment files, 2026-04-17. Analysis: Zan Digital, 12 Nov 2026.
Every red segment is something that happens after the first payment. The spec describes a business buyer well and a business relationship barely at all.

What the Agentic Commerce Protocol added for B2B buyers since 2025

The first release was a consumer checkout. Its Buyer object had four fields: first name, last name, email and phone number. Its PaymentData object had a token, a provider limited to stripe, and a billing address. Its session status had five values.

The 2026-04-17 version reads like a different document. The Buyer object has 11 fields, including account_type with a business value and a company object. PaymentData gained four fields that only make sense for a business: purchase_order_number, payment_terms, due_date and approval_required. The status enum grew from five values to 11, including pending_approval and authentication_required.

What changed between the first ACP release and the current one Count of schema elements. Light bar: 2025-09-29. Dark bar: 2026-04-17. Buyer fields 4 11 Session status values 5 11 Fulfillment types 2 4 B2B payment fields 0 4 Recurring billing fields 0 0 Source: ACP OpenAPI checkout files, versions 2025-09-29 and 2026-04-17, GitHub. Count: Zan Digital.
Four rows grew and one did not. The spec authors added a procurement vocabulary in seven months and still have not added a billing period.

Buyer identity is now a business identity

The CompanyInfo object requires a name and allows a tax_id, a department and a cost_center. The tax_exemption object takes a certificate type of resale, exempt_organization or government, and an expiry date. That is enough to let a procurement team buy on behalf of an entity rather than a person, which the first release could not do.

This matters more than it looks. A B2B purchase that cannot name the legal buyer cannot produce a valid invoice in most tax regimes. The spec now clears that bar for one-off orders.

Payment terms and approvals arrived together

The payment_terms enum is described in the schema as "Payment terms for B2B transactions". The approval_required flag is described as "Whether this payment requires approval". Together with the pending_approval status, they sketch a flow where an agent drafts an order, a human signs it off, and the seller ships on terms.

We think this is the most interesting change in the spec, and also the least proven. Nothing we read describes who the approver is, how the approval is captured, or how an agent learns that it happened. A status value is a placeholder for a workflow, not a workflow.

The subscription gap decides whether SaaS can sell in ChatGPT

A SaaS subscription needs at least four facts the protocol has no field for. They are a billing interval, a renewal date, a trial period and the terms on which the price can change. We ran a literal search of the 2026-04-17 checkout file and schema. Recurring, interval, trial and seat each return zero hits.

The single hit for "subscription" sits in the marketing consent options, where it means an email list opt-in. The nearest thing to a billing period is a Coupon.duration description that says a repeating coupon applies over "multiple billing periods". The spec references a billing period it never defines.

Delegated payment is built for one charge

The delegated payment spec says exactly one credential type is currently supported: card. Every token carries an Allowance with a reason, a max_amount, a currency, a checkout_session_id, a merchant_id and an expires_at. The reason enum has one value, one_time.

That design is deliberate and, for consumer retail, correct. A token scoped to one session, one merchant, one maximum and one expiry is hard to abuse. It is also the opposite of what a recurring charge needs, which is standing authority to bill the same card next month for an amount that may change.

Why the workaround breaks the protocol's own promise

A seller could treat the first month as a one-time order and then bill renewals through its own payment stack. That may be technically possible depending on what your payment provider does with the token. We did not find anything in the ACP documents that defines it, so it would be your arrangement, not the protocol's.

The problem is disclosure. The agent shows the buyer a one-time total. If the seller then bills monthly, the buyer agreed to a subscription through an interface that never showed them one. That is the kind of recurring-charge consent gap regulators already police, covered in our piece on subscription checkout rules for AI-sold plans. We would not ship it.

Seat counts, quotes and invoices: partial support is not usable support

The four partial rows in the table look like progress. Each one has a field. None has enough around it to carry a real B2B purchase without a human stepping in.

Seat counts depend on which part of the schema you read

In 2026-04-17, LineItem.quantity is an integer with a minimum of 1, which is what a seat count needs. But the request-side Item object has only an id, a name and a unit_amount, with additional properties disallowed. The spec's own examples still send quantity on items. The 2025-09-29 version put quantity on Item as a decimal greater than zero, and the changelog describes a move to decimal.

So the schema contradicts itself on where quantity lives and what type it is. A seller building to it today would have to pick one reading and hope the agent picked the same. For seat-priced software, where quantity is the price, that ambiguity is not cosmetic. Our analysis of seat compression in SaaS pricing covers why quantity is already the most contested number in a renewal.

Quotes and invoices are identifiers, not documents

The session can reference a quote_id and a quote_expires_at. The order confirmation can carry an invoice_number. Each field points at a document that lives somewhere else, in your CPQ tool or your billing system.

That is a sensible boundary for a checkout protocol. It also means the negotiation, redlines and approvals that produce a B2B quote stay outside the agent. The agent can carry the quote's reference into checkout. It cannot create the quote.

Where this argument is weakest

Three things could make this post wrong, and you should weigh them before acting on it.

First, the spec moves every few months. Five versions shipped in seven months, and the B2B fields went from absent to present inside that window. A recurring billing extension could land in the next release. The extension mechanism exists precisely so that capabilities can be added without breaking older clients. If it lands, the five red rows shrink fast.

Second, we read schemas, not live traffic. A field in a beta spec is not evidence that ChatGPT, or any agent, renders it, asks for it or passes it through. We found no public example of an agent sending payment_terms or a purchase_order_number. The supported column describes what the protocol permits, which may be more than any surface does today.

Third, the protocol may not be the path that matters. OpenAI's spokesperson said Instant Checkout was moving to Apps. A SaaS company's own app inside ChatGPT can run its own checkout, including a subscription, without ACP carrying the billing terms. That is a real counter-case. It routes the purchase through your own interface, which is the argument in our analysis of MCP apps as the SaaS front end. It also means the capability table is the wrong test for that route.

What we are confident about is narrower. Today, an agent following ACP cannot express a recurring B2B subscription in a standard way, and ChatGPT does not complete purchases itself. Both statements are checkable in documents you can open.

Where AI discovery should end for your product

If ChatGPT is where buyers are forming a shortlist, the real decision is what happens on the click after the shortlist. Our piece on the AI shortlist in the B2B buyer journey covers how the list forms. This section covers where it should land, by deal type.

Where an AI-referred B2B buyer should land, by deal type
Deal typeBest destination todayWhy
One-off digital product, licence key or template packACP checkout, if you are an approved partner.Digital delivery and one_time card payment both fit the spec.
Self-serve monthly or annual plan on a cardYour own checkout page, deep-linked by plan.No recurring billing in ACP; your page shows the renewal terms the agent cannot.
Seat-based team plan with an invoiceYour own checkout or a trial signup.Quantity is ambiguous in the schema and invoices are references only.
Contracted deal with procurementA sales call with the context the buyer already gathered.Quotes, redlines and approvals live outside the protocol.

The practical consequence for most SaaS teams is unglamorous. AI discovery ends in a link, so the page behind the link does the converting. A buyer who arrives from an AI answer already holds a comparison and a short list. Sending them to a generic homepage throws away the context that made them click.

Most seat-priced products need three things more than an ACP integration. They need plan-specific pricing pages with the renewal terms stated, a checkout that accepts a company name and tax ID, and a sales form that asks what the buyer already compared. Our work on how AI referral traffic converts sets out why that visitor deserves a different landing page.

Frequently asked questions

What is the Agentic Commerce Protocol?

The Agentic Commerce Protocol is an open specification, maintained by OpenAI and Stripe under the Apache 2.0 licence, that defines how an AI agent creates and completes a checkout with a merchant. It covers a product feed, a checkout session API and a delegated payment API that turns a buyer's card into a scoped token. It launched on 29 September 2025 and is still labelled beta.

Can you sell a SaaS subscription inside ChatGPT?

Not through a standard protocol today. The Agentic Commerce Protocol version 2026-04-17 has no billing interval, renewal or trial field, and its delegated payment token only allows one-time card charges. ChatGPT also no longer completes purchases itself; OpenAI's merchant page says purchases finish on merchant-owned websites or apps. A SaaS company can still link buyers to its own subscription checkout.

Does ChatGPT Instant Checkout still exist?

Not as a standalone feature. OpenAI said in March 2026 it would let merchants use their own checkout experiences and focus on product discovery, citing a lack of flexibility in the first version. Its merchant page now says users discover products in ChatGPT and complete purchases on merchant sites or apps. The underlying Agentic Commerce Protocol continues to ship new versions.

Does the Agentic Commerce Protocol support B2B purchases?

Partly. Version 2026-04-17 supports company identity, tax ID, tax exemption certificates, purchase order numbers and net 15 to net 90 payment terms. It has partial support for quotes, approvals, invoices and quantity. It has no support for recurring subscriptions, trials, metered pricing, non-card payment methods or renewals. Of 15 B2B requirements we checked, 6 were supported, 4 partial and 5 missing.

What payment methods does the Agentic Commerce Protocol support?

The delegated payment specification supports exactly one credential type, card, which can be a card number or a network token. Each token carries an allowance limited to one_time use, with a maximum amount, currency, merchant and expiry. The 2026-04-17 changelog adds seller-side handlers such as saved cards, gift cards, points and store credit. ACH and wire transfer are not defined.

Should B2B SaaS companies implement ACP now?

Most should not prioritise it yet. ChatGPT does not complete purchases itself, onboarding is limited to approved partners, and the protocol cannot express a subscription. Sellers of one-off digital products are the exception. For everyone else, the better investment is a landing page and checkout built for buyers arriving from AI answers, while watching the spec for a recurring billing extension.

Where to start

Open your pricing page and ask one question: could a buyer who arrived from a ChatGPT answer finish the purchase on this page without talking to anyone? If the plan, seat count, renewal term and invoice details are not all visible and settable there, that is your actual AI commerce gap, and no protocol closes it for you.

Then put a quarterly reminder against the ACP changelog. Look for one thing only: a billing interval or recurring allowance in the checkout or delegated payment spec. When it appears, rerun the 15-row table above against your own checkout, and decide then.

Related analysis

For how agent-led procurement is changing who reviews a deal, read the analysis of AI reviewers in the enterprise sales cycle. For the distribution question behind agent stores, see what agent marketplaces are worth as a channel.

References

  1. Stripe, Developing an open standard for agentic commerce, 29 September 2025. Used for release date and stated goals.
  2. Agentic Commerce Protocol, GitHub repository and README, read 12 November 2026. Used for maintainers, licence, beta status and version list.
  3. Agentic Commerce Protocol, 2026-04-17 changelog. Used for deprecation, quantity change and seller-side payment handlers.
  4. Agentic Commerce Protocol, OpenAPI checkout specification, 2026-04-17, compared with the 2025-09-29 file in the same repository. Used for every field in the capability table and the version comparison.
  5. Agentic Commerce Protocol, OpenAPI delegated payment specification, 2026-04-17. Used for card-only credentials and the one_time allowance.
  6. OpenAI, ChatGPT merchants page, undated, read 12 November 2026. Used for the move to merchant-owned checkout, US availability and fees.
  7. OpenAI, Commerce product feed specification, read 12 November 2026. Used for checkout eligibility flags and the absence of recurring price fields.
  8. Modern Retail, What went wrong with ChatGPT's Instant Checkout, 27 March 2026, and its 12 March 2026 report on Shopify and the OpenAI spokesperson statement. Used for the March 2026 timeline and OpenAI's quoted reasoning.

The weakest part of this source base is that the capability verdicts rest on a beta specification read by us, not on observed agent behaviour. OpenAI's own 24 March 2026 post could not be opened directly, so its wording is taken from Modern Retail and Retail Gazette. Last verified 12 November 2026.

MJ
Sanskriti Khandelwal
Writes for Zan Digital about AI product economics, B2B software markets and what specifications and filings actually say behind vendor announcements.

Related reading