Agentic Commerce Readiness: What ISOs Must Build

TL;DR — Quick Summary

  • The readiness gap is not consumer-facing: on 24 March 2026 OpenAI moved ChatGPT away from its own Instant Checkout and toward product discovery, extending ACP to carry merchant product feeds and promotions. The agent cannot buy what it cannot read, and most ISO merchants’ catalogues were never built to be read by a machine.
  • Four data foundations, in a fixed order: stable item identity, authoritative price and availability, explicit machine-readable terms, and a mandate-to-order trail. None of them is a protocol integration. All four are work on data the ISO already sits on — or, just as often, does not.
  • The strategic question is ownership, not adoption: whoever holds the merchant’s canonical product data decides which channels the merchant can reach. Hold it and the channel becomes a billable software line; leave it with the commerce platform and the channel becomes a bypass. Put that decision into the 2027 roadmap before the merchant’s platform makes it for you.

The Decision, Up Front

By the end of Q1 2027 you need three things written down: an inventory of where each merchant’s canonical product data lives, a four-SKU reconciliation test with named owners, and a position on who publishes and maintains the feed. None of it requires a protocol build. All of it requires deciding whether your platform is a system of record or a reseller of someone else’s.

Data foundations
Four

Pivot date
24 Mar 2026

Positions to choose
Three

Asset that matters
Product data

The Gap Is Not the Consumer

By now the threat framing is familiar: an AI agent sits between a buyer and your merchant, compares options, and completes a purchase without anyone visiting a website. We wrote that argument ourselves in 2025. It is a real distribution risk, and it is not what this briefing is about. This one starts one layer further down, at the unglamorous question of whether the merchant’s products can be described to a machine at all.

The shift that made this concrete happened on 24 March 2026. OpenAI said the first version of Instant Checkout “did not offer the level of flexibility that we aspire to provide”, let merchants keep their own checkout, and redirected its effort to product discovery — extending the Agentic Commerce Protocol so merchants share product feeds and promotions. Seven large retailers had already integrated for discovery. Shopify merchants were connected automatically through Shopify Catalog, with no work required from the individual merchant. In other words, the front door moved from the checkout to the catalogue, and for millions of merchants the catalogue was populated by a platform rather than by their POS provider.

That is an ISO problem, not a merchant problem. A single-location restaurant does not maintain stable variant IDs, a price freshness target, or a machine-readable returns policy, and it never will. The party that already runs the merchant’s system of record is the only party that can. This is a briefing about what to build, in what order, and what it is worth — with the dated facts sourced at the bottom and the arithmetic clearly labelled as arithmetic.

1. The Timeline, And the One Window You Still Control

Unlike a statute, this calendar has no single compliance date. That is exactly why it gets deferred. What has happened since September 2025 is a sequence of architectural decisions that are now settled enough to build against; the only open item is your own roadmap. Note the last node is a commercial decision window, not a regulatory deadline — we are not going to invent a filing date to make it sound more urgent.

SEP 2025 · DONE
ACP ships as an open standard
Stripe and OpenAI release the Agentic Commerce Protocol under Apache 2.0, alongside Instant Checkout in ChatGPT. Businesses remain the merchant of record.

24 MAR 2026 · DONE
OpenAI pivots from checkout to discovery
“The initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide.” Merchants keep their own checkout; ACP is extended to carry product feeds and promotions.

28 APR 2026 · DONE
Standards move to a neutral home
FIDO Alliance forms the Agentic Authentication and Payments Technical Working Groups. Google contributes AP2; Mastercard contributes Verifiable Intent.

MID 2026 · DONE
The stack separates into layers
UCP (ucp.dev) covers the shopping journey with built-in AP2 / A2A / MCP support; its Food group lists Square, Toast, Uber Eats and DoorDash. Card networks supply agent identity and tokens.

HOLIDAY 2026 · LIVE NOW
The first real agent purchases
Visa describes agentic commerce as ranging “from comparing products and initiating checkout to completing transactions with user permission” — and states plainly that the product is still in the process of deployment.

Q1 2027 · YOUR WINDOW
Decide who owns the product data
A commercial decision window, not a regulatory deadline: put data ownership, feed responsibility and reconciliation duties into 2027 roadmaps and merchant agreements.

2. What Actually Changed — Three Facts

Strip away the product announcements and three things are different in kind, not just in degree. Every one of them is verifiable against a primary source below.

1
The feed became the front door

On 24 March 2026 OpenAI said it was “allowing merchants to use their own checkout experiences while we focus our efforts on product discovery”, and extended ACP so merchants “share product feeds and promotions so their catalogs are fully represented”. Seven named retailers had already integrated for discovery, and Shopify merchants were connected automatically through Shopify Catalog with “no additional work” required. The ACP site is equally explicit that implementing the spec does not automatically list your products — each AI platform runs its own onboarding process.

2
The merchant stays merchant of record — but the data has to reconcile

Per the ACP site, businesses “maintain their customer relationships as the merchant of record” and keep control of what is sold and how orders are fulfilled; payment still runs through the merchant’s own PSP, with Stripe’s Shared Payment Token the first compatible implementation. In the ACP feed model the feed is informational and the checkout response is authoritative for price, stock, tax and delivery. If the feed’s variant ID is not the ID the checkout accepts, the merchant must supply a mapping.

3
The credential changed shape, and so did the audit trail

The FIDO Alliance’s 28 April 2026 announcement names the problem precisely: current authentication models “were designed for direct human interaction, not delegated, agent-initiated actions”. Its three work areas are verifiable user instructions, agent authentication, and trusted delegation for commerce. Visa’s own agentic commerce page lists spending limits, approval workflows, authentication requirements and trusted identity signals. There is no moment where a person clicks buy — so the record of what was authorised has to survive a dispute.

3. What It Does Not Do — Read This Before You Brief Anyone

Most internal conversations about agentic commerce go wrong in the first five minutes because somebody overstates it. These are the four corrections to make before a roadmap meeting starts.

It is not another warning about agents taking your funnel. We made that argument in Agentic Commerce: How AI Agents Could Bypass Your Entire Merchant Acquisition Funnel. This briefing starts after it: the funnel is a distribution problem, the readiness gap is a data problem, and the two have different owners.

It is not a plugin, and it is not automatic. The ACP FAQ answers directly: “No, each AI platform will manage their own process for how businesses can participate.” Shopify merchants are connected by Shopify; an independent merchant on a white-label stack is not. Platform approval is an input you do not control — the same commitment boundary we set for onboarding SLAs.

It is not your job to become a payment protocol company. Know Your Agent checks, network tokens and Web Bot Auth signatures sit with the acquirer and the PSP. Your job is narrower and less glamorous: make the merchant’s catalogue, price, terms and order record machine-readable, and keep them reconciled.

It is not a volume forecast. There is no citable benchmark for agentic GMV, and we will not invent one. Adobe’s widely quoted +805% figure is year-on-year growth in AI-referred retail traffic over Black Friday 2025, not completed transactions. Where this article uses numbers for economics, it labels them illustrative arithmetic on stated inputs.

4. What It Means for an ISO — Three Impacts

The useful way to read the four foundations is not as an IT project. It is as three changes to the economics of the book you already own.

1. Product data becomes a line on the bill

Catalogue hosting, feed publication and ID mapping are recurring services with a per-merchant cost and a per-merchant price. That makes them a software line, not a processing feature — the same distinction behind software attach rate: only merchants with a software fee on the bill count, and you report recurring gross profit share alongside attach so a rising attach rate cannot hide flat margin. Add the line, and keep the same discipline.

2. Data quality becomes a retention variable

An agent does not guess. It queries, reads a structured response and either transacts or moves to the next merchant. A stale price is not a minor defect; it removes the merchant from the comparison entirely, with no error message and no page view to diagnose. That is a churn mechanism with no support ticket attached, which is why the reconciliation test belongs next to the churn benchmarks and the onboarding SLA work rather than in a future innovation backlog.

3. The dispute surface moves into your support queue

When no human clicks buy, the traditional authorisation moment disappears. What replaces it is a record: verifiable user instruction, agent identity, and an order that can be traced back to both. FIDO’s three work areas are precisely those three records. For an ISO this lands as a support and reconciliation cost — the same place cost per ticket already lives — not as a fraud project you can hand to the processor and forget.

The four data foundations, in build order

Dependency diagram, not a maturity score. Build top to bottom: each layer assumes the one above it is already clean.

L1
Identity
Stable item and variant IDs. One identifier across POS, feed, checkout and order management. GTIN, UPC or MPN where they genuinely exist — never a fabricated substitute. If the feed ID is not the ID the checkout accepts, maintain an explicit mapping.

L2
Authority
Authoritative price and availability. The feed is informational; the checkout response is authoritative. Set a freshness target and test it. Contract and tier pricing must be exposed through the data layer, not trapped in a storefront UI or a rep’s memory.

L3
Terms
Explicit, machine-readable terms. Delivery, tax, returns, privacy and service terms published as structured data, and matching what fulfilment actually does. An agent will not infer a policy from prose, and a mismatch becomes a dispute.

L4
Trail
A mandate-to-order trail. One authorised request produces exactly one order. You can reconstruct what the agent was permitted to do, what it did, and whether the two matched — including duplicate requests, stale inventory and interrupted payment.

An illustrative model, not a forecast

Take the same book used across this series: 5,000 merchants, $68 NMGM, $600 CAC (the margin benchmark). Suppose 20% of the book — 1,000 merchants — takes an agent-ready catalogue service at $25 per merchant per month. That is 1,000 × $25 × 12 = $300,000 a year of recurring revenue; at the 85% software gross margin used in the attach analysis, $255,000 of gross profit. On the other side of the ledger, the churn work put one point of annual attrition at roughly $71,000; holding three points is $213,000 of margin not lost.

Every input above is an assumption, not a survey result, and the numbers are arithmetic on those assumptions. Replace them with your own. The structural point survives the substitution: the merchant-level record that produces NMGM is the same record an agent channel needs. If you own it, the channel is a product. If the commerce platform owns it, the channel is a bypass. That is the same “who holds it” question we asked about funds under the GENIUS Act, applied one layer earlier.

5. The Three Positions

There are only three defensible answers to “who owns the merchant’s product data”. Most ISOs will take the first by default, which is the only position that carries no decision at all. The middle one is where the asset sits; the third is a different business.

Position What you do Revenue effect Compliance load
Pass through Resell POS and processing; leave product data with the merchant or the commerce platform No new line. The platform connects the merchant directly — Shopify Catalog already does Low — you touch no agent credentials
Own the data layer Hold the canonical catalogue in your white-label platform; publish feeds, maintain ID mapping, bill per merchant Recurring software revenue on the book you already have Bounded — data quality and reconciliation, not scheme rules
Become the channel Build agent-facing endpoints, apply to platforms on the merchant’s behalf, operate the integration Highest ceiling, slowest to land High — protocol churn, platform approval, dispute handling

6. Risk Exposure by Role

The exposure is uneven, and it does not fall where the announcements suggest. The party with the most to lose is the one that currently holds no canonical data, not the one that will process the payment.

Role What creates the exposure What it costs Bites on
Single-location restaurant Menu, price and modifiers live only in the POS UI; no stable item IDs An agent reads a stale price or an out-of-stock item and routes to the next merchant — no comparison, no sale Holiday 2026
Multi-location / group Each store carries different prices and availability; variant IDs do not map to one catalogue Duplicate orders, wrong totals, support tickets, and a merchant who blames the platform Q1 2027
ISO / white-label platform Does not hold canonical product data, so the commerce platform connects the merchant directly The merchant relationship thins out and there is no software line to put on the bill 2027 renewals
Acquirer / PSP Bot rules block every non-browser client; no agent credential path enabled A signed, legitimate agent is declined alongside the scrapers Now
B2B / wholesale merchant Contract pricing and catalogue restrictions exist only in the UI or in a rep’s head A procurement agent cannot obtain a price, so the supplier is never shortlisted 2027

7. The Decision Checklist

Three columns, and the third matters as much as the first. Half the cost in this transition comes from work that should never have been started: fake identifiers, a feed treated as the price of record, and promises tied to somebody else’s approval queue.

DO NOW
Before the next roadmap review — five inputs you actually control

  • Inventory. List, per merchant, where the canonical product data lives: inside your platform, only in the POS UI, or at a third-party commerce platform. Most ISOs cannot answer this today.
  • Reconcile four SKUs. One simple item, one variant family, one out-of-stock item, one discounted item. Compare ID, price, currency, availability and delivery terms across POS, feed and checkout. Record each mismatch as a defect with an owner and a retest — not as a vague note that “the data is messy”.
  • Set a freshness target and test it. Change one price and time how long it takes to appear everywhere an agent would look. A feed that lags the checkout is worse than no feed, because the checkout response is authoritative and the mismatch lands on you.
  • Publish the terms as data. Returns, delivery, tax, privacy and service terms in machine-readable form, matching what fulfilment actually does. An agent does not infer policy from a paragraph of prose.
  • Ask your acquirer one question. Which agent credentials and verification signals will reach your merchants, and will current bot rules drop a signed agent along with the scrapers?
WATCH
Four things that would change the build order

  • FIDO output. Specifications from the Agentic Authentication and Payments Technical Working Groups (AP2 and Verifiable Intent).
  • OpenAI checkout. Whether native Instant Checkout returns, and whether multi-item carts arrive.
  • UCP Food. Which surfaces Square, Toast, Uber Eats and DoorDash actually open for restaurant merchants.
  • Dispute rules. Whether card networks publish formal chargeback treatment for agent-initiated transactions.
EXPLICITLY DON’T
Three shortcuts that only create cleanup work

  • Do not invent identifiers. The ChatGPT feed specification requires a GTIN to carry a valid check digit and tells merchants not to fabricate a part number to stand in for a missing one.
  • Do not treat the feed as the price of record. Under ACP the checkout response is authoritative for price, stock, tax and delivery. Pricing off the feed creates exactly the mismatch the protocol was designed to prevent.
  • Do not promise a go-live date you do not control. Platform approval is somebody else’s input. Commit to the data work you perform, not to “selling through AI in 30 days”.


Why the ISO, Not the Merchant, Owns This

A restaurant owner will never maintain variant IDs or a feed freshness target, and should not be asked to. The party running the merchant’s system of record is the only party that can make the catalogue authoritative. That is an argument about where data sits, not about which vendor has the best AI roadmap.

  • One merchant-level record. The same per-merchant view that produces NMGM is the view that tells you whether a catalogue is worth maintaining at all.
  • Software revenue on the same ledger. A feed service only counts if it is on the bill — the attach rate discipline applies unchanged.
  • An export for diligence. When a platform or acquirer asks which merchants you can represent, a merchant-level export answers it in an hour rather than a quarter — the same payoff as the funds inventory in the last briefing.

Frequently Asked Questions

Is agentic commerce already generating meaningful volume for ISOs?

There is no citable benchmark for agentic GMV, and we will not manufacture one. Adobe’s +805% figure measures year-on-year growth in AI-referred retail traffic over Black Friday 2025, not completed orders. Visa states its Intelligent Commerce product is still in the process of deployment. The honest position is that the architecture is settling faster than the volume, which is why the build order in section 4 is worth doing now and the revenue case is not worth promising.

Do I need to implement ACP or UCP myself?

Almost certainly not as a first step. The ACP FAQ says each platform manages its own merchant onboarding, so implementing the spec does not list your products anywhere. Agent credentials, network tokens and bot verification sit with the acquirer and PSP. The work that is unambiguously yours is the four data foundations: identity, authoritative price and availability, machine-readable terms, and an order trail. Build those first, then decide whether a direct integration earns its cost.

What is the single most common data gap?

Two, usually together: no stable item or variant identifier across systems, and a price or stock change that never reaches the feed. Test it with four SKUs — one simple item, one variant family, one out-of-stock item, one discounted item — and compare ID, price, currency, availability and delivery terms across POS, feed and checkout. Under ACP the checkout response is authoritative for all four commercial facts, so any divergence is a defect with an owner, not a rounding difference.

Will this cannibalise my processing residual?

Not in the way a new rail does. Under ACP the business remains the merchant of record and payment still runs through the merchant’s own PSP, so an agent-initiated card purchase is not automatically a disintermediation event. The risk here is different and quieter: if a commerce platform holds the catalogue, it owns the discovery relationship, and the ISO becomes a processing vendor with no software line. The threat is to attach, not to interchange.

How does this connect to software attach rate and valuation?

Directly, if you bill for it. Catalogue hosting and feed maintenance are recurring software services, so they belong in the attach numerator only when there is a software fee on the merchant’s bill — free trials and enabled-but-unbilled merchants are excluded. Report recurring gross profit share alongside attach, exactly as in the attach analysis, so a higher percentage cannot mask flat dollars.

Are the economics in section 4 forecasts?

No. They are illustrative arithmetic on stated inputs — 5,000 merchants, 20% coverage, $25 per merchant per month, 85% gross margin, $71,000 per point of churn — and every one of those inputs should be replaced with your own numbers. The dated facts come from the primary sources listed below; the model is there to show the structure of the decision, not to predict a result.

Bottom Line

The protocols will keep changing; the data will not. On 24 March 2026 the front door moved from the checkout to the catalogue, the merchant stayed merchant of record, and the payment credential became a verifiable instruction instead of a click. Four foundations follow in a fixed order — stable identity, authoritative price and availability, explicit terms, and a mandate-to-order trail — and none of them is a protocol integration. The decision is not whether to adopt agentic commerce. It is whether your platform holds the merchant’s canonical product data, because the party that holds it decides which channels the merchant can reach. Put that answer in the 2027 roadmap, run the four-SKU reconciliation test, and name the position before a commerce platform names it for you. OrderPin is an ISV focused on restaurant POS software.

Sources, Dates & What Would Change This

Primary sources — official and dated: agenticcommerce.dev (ACP: Apache 2.0, merchant of record, REST/MCP, Stripe Shared Payment Token, platform-run onboarding); OpenAI, “Powering Product Discovery in ChatGPT”, 24 March 2026 (Instant Checkout pivot, ACP extended to feeds and promotions, named retailers, Shopify Catalog); ucp.dev (UCP co-developers, built-in AP2 / A2A / MCP support, public specification on GitHub); FIDO Alliance, “FIDO Alliance to Develop Standards for Trusted AI Agent Interactions”, 28 April 2026 (two working groups, AP2 and Verifiable Intent contributions); Visa, “Agentic commerce at scale” (spending limits, approval workflows, trusted identity signals; product stated to be in deployment); Stripe Docs, “Agentic commerce” (Shared Payment Tokens, Catalog feed, UCP).

Treated as secondary, not as binding dates: ACP Feed API version 2026-04-17, UCP 2026 specification versions, the April 2026 edition of the Visa Core Rules, Mastercard’s Agent Pay Acceptance Framework (October 2025), and Adobe’s +805% Black Friday 2025 AI-referral figure all come from vendor announcements or industry summaries. Where a version or effective date matters to a decision, verify it against the official specification site or your acquirer before acting.

Status as of 18 November 2026: No single protocol has won. ACP (checkout and feeds), UCP (shopping journey) and AP2 (authorisation mandates) sit at different layers and are explicitly designed to interoperate. FIDO’s two working groups are still developing specifications. Visa states its Intelligent Commerce product is in the process of deployment, and broad consumer agent checkout is not yet a default path. Implementing ACP still does not automatically list a merchant in ChatGPT.

What would change this analysis: Final FIDO specifications; OpenAI restoring native Instant Checkout or shipping multi-item carts; a major platform opening self-serve merchant onboarding instead of an application gate; card networks publishing formal dispute rules for agent-initiated transactions; or the UCP Food group making restaurant POS data a default integration path. If one lands, re-run the section 7 checklist rather than rewriting the strategy.

Not a forecast, not legal advice: There is no citable agentic GMV benchmark, so this article does not offer one. The economics in section 4 are illustrative arithmetic on inputs stated in the text and should be replaced with your own. Scheme-role and compliance questions depend on facts and on how acquirers apply unsettled rules — validate them with your processor and counsel.

About OrderPin

OrderPin is a white-label POS platform built for ISO and MSP partners. We offer full data ownership, flexible pricing, and seamless API integrations to help you build a recurring revenue business under your own brand. Learn more about OrderPin’s white-label solution

Scroll to Top