ChatGPT Agentic Commerce in 2026: The Complete B2B Guide
Summary
Key takeaways
- Agentic commerce uses AI agents to perform commercial tasks on behalf of buyers or merchants, moving beyond recommendations into product discovery, comparison, quoting, ordering, payment, and post-purchase coordination.
- In B2B commerce, autonomous purchasing must operate within approved suppliers, negotiated contracts, customer-specific catalogs, spending limits, payment terms, and internal approval policies.
- Buyer agents and merchant agents serve different roles. Buyer agents represent procurement teams, while merchant agents help sellers qualify demand, prepare quotes, recommend products, enforce pricing rules, and coordinate orders.
- ERP-connected merchants cannot treat the ecommerce platform as the only source of truth. Agents need reliable access to ERP, PIM, CRM, OMS, WMS, CPQ, and payment data.
- Machine-readable product data is foundational. Agents need structured specifications, compatibility information, units of measure, documentation, availability, lead times, and identifiers rather than marketing copy alone.
- Pricing is one of the hardest B2B agentic-commerce challenges because the correct price may depend on the account, contract, quantity, currency, location, channel, effective date, and approved exception.
- Open APIs and headless architecture help, but they do not automatically make a commerce system agent-ready. Reliable identity, persistent cart state, programmatic checkout, tokenized payment, and order-status access are also required.
- Protocols such as MCP, ACP, AP2, and UCP provide emerging ways for agents to access commerce data, tools, checkout, and payment capabilities, but merchants should prepare for a multi-protocol environment rather than depend on one standard.
- Governance determines how far automation can go. Permissions, spending thresholds, approval requirements, audit trails, confidence rules, exception handling, and human overrides must be designed before agents receive transaction authority.
- The safest implementation strategy is phased: begin with discovery and advisory use cases, progress to draft quotes and carts, and permit autonomous transactions only after data quality and control mechanisms have been proven.
When this applies
This applies when a manufacturer, distributor, wholesaler, or enterprise merchant wants AI to participate in real commercial workflows rather than provide a generic storefront chatbot. It is especially relevant for businesses with repeat orders, large technical catalogs, customer-specific pricing, RFQs, procurement approvals, ERP-controlled inventory, negotiated contracts, or account-based payment terms.
It also applies when buyers increasingly use external AI tools to research suppliers and products. In this environment, the merchant must make products discoverable to machines while maintaining accurate pricing, availability, governance, and transaction control across connected systems.
When this does not apply
This does not apply when core ecommerce operations are still unreliable. Businesses with incomplete product data, inconsistent prices, inaccurate inventory, unstable ERP integrations, or unclear system ownership should repair those foundations before delegating decisions to agents.
It is also unnecessary when the proposed use case is only a conventional chatbot, product recommender, or internal content assistant. Those tools may use AI, but they do not constitute agentic commerce unless they can use external tools, follow commercial policies, and execute or coordinate meaningful commerce actions.
Checklist
- Define the commercial tasks agents should perform.
- Separate buyer-agent use cases from merchant-agent use cases.
- Distinguish advisory, semi-autonomous, and autonomous workflows.
- Identify the system of record for products, customers, prices, inventory, orders, and payments.
- Audit product attributes, identifiers, specifications, documentation, and compatibility data.
- Make customer-specific catalogs and entitlement rules machine-readable.
- Document pricing formulas, contract rules, quantity breaks, currencies, and exceptions.
- Confirm that inventory, availability, and lead times are sufficiently current.
- Map ERP, PIM, CRM, OMS, WMS, CPQ, procurement, and payment integrations.
- Define agent identity, authentication, account permissions, and delegated authority.
- Set spending limits, approval thresholds, supplier restrictions, and payment rules.
- Provide APIs for catalog discovery, cart creation, quoting, checkout, and order status.
- Add audit logs, confidence thresholds, exception queues, and human override mechanisms.
- Pilot one narrow workflow with production-like data and controlled users.
- Expand autonomy only after measuring accuracy, adoption, commercial value, and exception rates.
Common pitfalls
- Calling every commerce chatbot an agentic-commerce implementation.
- Giving agents transaction authority before pricing and inventory data are reliable.
- Assuming open APIs alone make a platform agent-ready.
- Exposing public catalog data while hiding the account-specific information B2B buyers actually need.
- Allowing agents to bypass contracts, approved suppliers, credit limits, or internal approvals.
- Using free-text product descriptions where structured specifications are required.
- Choosing one emerging protocol as the permanent standard too early.
- Automating RFQs without preserving versions, approvals, margin rules, and negotiation history.
- Measuring success through agent usage rather than order accuracy, cycle time, conversion, margin, and exception reduction.
- Launching broad autonomous workflows before proving a controlled, reversible use case.
The most useful thing that happened to agentic commerce in 2026 was a failure. Understanding why the first version of AI checkout did not work tells you more about where this is going than any forecast does — particularly if you sell B2B.
What agentic commerce actually means
Agentic commerce is a transaction model in which an AI agent — not a person navigating a storefront — discovers products, evaluates options and initiates a purchase on someone’s behalf, using machine-readable protocols instead of a graphical interface.
The distinction that matters is delegated authority. A chatbot that recommends a product and hands you off to a checkout page is conversational commerce, and it has existed for a decade. An agent that holds a cryptographically signed mandate proving it was authorised to spend up to a defined limit, on defined terms, and then executes against that mandate, is something structurally new. The mandate is the dividing line, not the conversation.
That distinction has a second-order consequence that most coverage skips. If agents transact through structured feeds and APIs rather than rendered pages, then the merchant’s visibility stops being a function of how a page looks and starts being a function of whether the underlying data is exposed, accurate and trustworthy. For a consumer brand that is a marketing problem. For a distributor running account-specific pricing out of an ERP, it is an architecture problem — and a considerably harder one.
In one line: Agentic commerce is what happens when the buyer is software, the interface is an API, and the authorisation is a signed mandate rather than a click.
What changed in 2026
Most guides published this year describe an escalating land grab between OpenAI, Google and the card networks. That was the 2025 story. Two things happened in 2026 that changed the shape of it, and both point in the same direction.
The first version of in-chat checkout was pulled
OpenAI and Stripe launched the Agentic Commerce Protocol on 29 September 2025, and with it Instant Checkout inside ChatGPT — starting with Etsy, then extending to Shopify merchants and Walmart. Roughly thirty merchants shipped it.
By March 2026 OpenAI had retired the standalone Instant Checkout experience, stating on its merchant-facing page that it was moving away from it to prioritise product discovery instead. The reason, reported by Wired on 18 March 2026 and corroborated by CNBC, was conversion. Walmart executive vice president Daniel Danker put numbers to it: in-chat checkout converted at roughly one-third the rate of simply clicking through to Walmart.com, at a 1.18% conversion rate with 77.45% cart abandonment. He called the experience unsatisfying.
Walmart replaced it with its own agent, Sparky, running inside ChatGPT and Gemini, which reportedly converts at around 70% of the rate of Walmart.com itself. The pattern that emerged is not “buy inside the AI.” It is discover inside the AI, transact on the merchant’s own infrastructure.
Reported, not filed: The Walmart conversion figures come from executive commentary reported in the trade press, not from a regulatory filing. Treat the direction as reliable and the decimals as indicative.
Protocol governance consolidated
On 28 April 2026 Google donated the Agent Payments Protocol to the FIDO Alliance, releasing AP2 v0.2 on GitHub the same day. The new version adds “Human Not Present” payments — an agent executing a pre-authorised transaction without the user present, for time-sensitive purchases. Alongside it, Google and Mastercard co-developed and jointly donated Verifiable Intent, a standard that creates a tamper-proof log of user-authorised agent actions.
The FIDO Alliance responded by forming an Agentic Authentication Technical Working Group and directing its Payments Technical Working Group to develop specifications for agent-initiated commerce, drawing on both contributions. This is what it looks like when an emerging transaction class gets absorbed into regulated payment infrastructure: the vendor hands the specification to a neutral standards body and the incumbents take over governance.
The practical read for merchants is that the “protocol wars” framing is now roughly a year out of date. The commerce layer is still competitive. The trust and authorisation layer is consolidating.
The protocol stack, mapped
The single most common error in coverage of this topic is treating every named protocol as a competitor to every other. They are not. They sit at different layers, and a merchant will end up touching several.
| Standard | Owner | Layer and function | Status as of July 2026 |
|---|---|---|---|
| ACP | OpenAI + Stripe | Commerce — agent-initiated checkout, delegated payment token, product feed | Open standard, Apache 2.0, launched 29 Sep 2025. Latest dated spec 2026-04-17. Beta. Product model pivoted toward discovery. |
| UCP | Google + Shopify | Commerce — full journey: discovery, checkout, fulfilment, post-purchase | Launched at NRF, January 2026. Version 2026-04-08. Backed by Visa, Mastercard, Target, Walmart, Etsy, Wayfair. |
| AP2 | FIDO Alliance (donated by Google) | Authorisation — signed Intent, Cart and Payment mandates | v0.2 released and donated to FIDO Alliance 28 Apr 2026. Adds Human Not Present payments. 60+ contributing organisations. |
| Verifiable Intent | Google + Mastercard → FIDO | Accountability — tamper-proof log of authorised agent actions | Co-developed and donated to FIDO Alliance, April 2026. Compatible with AP2. |
| MCP | Anthropic | Connectivity — how agents reach tools and data | Mature and widely adopted. Not commerce-specific. |
| Trusted Agent Protocol | Visa + Cloudflare | Identity — verified agent ID so merchants can tell agent from human | Announced 14 Oct 2025. Folded into Visa Intelligent Commerce Connect, launched April 2026 as a protocol-agnostic gateway. |
| Agent Pay / Agentic Tokens | Mastercard | Payment credential — scoped, tokenised card credentials for agents | Announced 29 Apr 2025. Live pilots. Extends existing tokenisation infrastructure. |
| ACE Developer Kit | American Express | Card + membership — agent registration, intent, credentials | Specifications released 14 Apr 2026, with purchase protection for erroneous agent purchases. |
| x402 | Coinbase + Cloudflare | Machine payments — HTTP 402 stablecoin micropayments | Governance under the Linux Foundation. Relevant to agent-to-service payments, not consumer checkout. |
The honest summary: ACP and UCP are the two commerce protocols and they are not yet interoperable, which means a merchant serious about coverage implements both. AP2 and Verifiable Intent now sit underneath as the neutral authorisation and accountability layer. The card networks are converging on a shared mandate model while diverging on trust architecture — Mastercard through delegated tokenholder credentials, Visa through a merchant-facing gateway. MCP and x402 solve adjacent problems and should not be in the same comparison at all, though they frequently are.
How an agentic transaction actually works
Strip away the branding and every implementation follows the same four stages.
- Intent resolution. The agent interprets what the buyer wants, including constraints the buyer never stated explicitly — budget, delivery window, compatibility with something already owned.
- Catalogue retrieval. The agent queries merchant product data through a structured feed or API: identifiers, descriptions, pricing, real-time availability, fulfilment options. This is the stage where most merchants silently fail, because the data exists but is not exposed, or is exposed but is stale.
- Mandate and authorisation. The agent presents a cryptographically signed mandate proving delegated authority, paired with a tokenised payment credential. The agent does not see raw card details. The merchant generally remains merchant of record.
- Settlement and confirmation. The processor or network clears the transaction and confirms back to both the merchant system and the agent, which reports to the buyer.
Stage two is where B2B diverges sharply from retail, and it is the reason the retail playbook does not transfer. In B2B, the correct price is not a catalogue price. It is a function of the account, the contract, the volume tier, the currency, the incoterm and sometimes the individual order. That price usually lives in the ERP, not the commerce platform.
Market size, reconciled
Published forecasts for agentic commerce differ by more than an order of magnitude. This is usually presented as analyst disagreement. It is not. It is definitional: the firms are measuring different things, and once you normalise for what each one counts, the range becomes legible.
| Source | Estimate | Horizon | What it actually counts |
|---|---|---|---|
| eMarketer (Dec 2025) | $20.57bn US in 2026, ~1.5% of US online retail; $144bn by 2029 | 2026–2029 | Narrowest. Checkout completed on an AI platform. The most verifiable number here. |
| Morgan Stanley AlphaWise (Nov 2025) | $190bn base case to $385bn bull case, US, 10–20% | By 2030 | Strict. Requires genuine autonomous action by the agent. |
| Bain & Company (Dec 2025) | $300–500bn US, 15–25% of total US ecommerce | By 2030 | Broader. Purchases “influenced” by agentic discovery. |
| McKinsey (Oct 2025) | $3–5tn globally | By 2030 | Broadest. Any AI-orchestrated commerce, global, including assisted discovery. |
| Gartner (Oct 2025) | 90% of B2B buying AI-agent intermediated, >$15tn of B2B spend | By 2028 | B2B only. Intermediation, not autonomous purchase. The most-cited B2B figure. |
The spread between eMarketer and McKinsey is roughly thirty-five fold, and almost all of it is definitional. If you are building a business case, use eMarketer for what will actually clear through AI platforms in the near term and Bain or Gartner for what share of demand will be agent-influenced. Quoting McKinsey’s number as though it describes agent-executed transactions is the most common error in vendor decks on this topic.
Adoption signals worth more than the forecasts
- AI-referred retail traffic grew 393% year on year in Q1 2026 and 138% year on year in May 2026 (Adobe Analytics, based on over a trillion visits).
- The conversion picture reversed. In March 2025 AI-referred traffic converted 38% worse than other traffic. By March 2026 it converted 42% better, and by May 2026, 54% better (Adobe Analytics).
- Shopify reported AI-driven orders growing roughly thirteen-fold year on year in Q1 2026, disclosed in its 8-K filing of 5 May 2026 — the most reliable single datapoint in this list because it is a regulatory filing.
- Salesforce put AI and agent-influenced global sales during Cyber Week 2025 at $67bn, roughly 20% of all global orders, based on data from 1.5 billion shoppers.
Methodology caveat: The Adobe and Salesforce figures are vendor-stated with methodology that is not fully public. The direction is corroborated across independent sources; individual decimals are not independently verifiable. Cite the trend, not the decimal.
The conversion reversal is the number that matters. It means AI-referred visitors stopped being tyre-kickers and started being buyers with pre-formed intent. That happened while in-chat checkout was being withdrawn — which is precisely the point. The value is arriving through discovery, and it is landing on merchant infrastructure.
The B2B reality check
Almost every guide to agentic commerce is written about retail. That is a problem, because the B2B version of this shift is larger by value, slower by nature, and governed by constraints that simply do not exist in consumer commerce.
Digital Commerce 360 reported on 10 March 2026 that agentic commerce was hitting a reality check in B2B. Rather than replacing ecommerce platforms and procurement systems, companies were repositioning AI agents as assistants for discovery and purchasing workflows while transactions stayed inside merchant and enterprise systems. The reasoning was operational, not technological: B2B orders involve negotiated pricing, credit terms, ERP integration and complex logistics, and letting an external AI interface execute transactions directly would bypass financial and operational controls deliberately embedded in distributor and manufacturer systems.
Read that alongside the Instant Checkout withdrawal and the same conclusion arrives from two independent directions. Retail found that in-chat checkout converted badly. B2B found that in-chat checkout would break the controls. Both landed on: discover in the agent, transact in your own system.
The strategic implication: The B2B question is not “how do we let agents buy from us.” It is “is our product, pricing and availability data good enough, and reachable enough, for an agent to recommend us at all.” Those are very different projects, and the second one is mostly an integration project.
The constraint nobody wants to name is the ERP
Deloitte Digital research reported in February 2026 found roughly two-thirds of B2B suppliers not currently using agentic AI intended to adopt it, with only 9% ruling it out. The gap between intent and execution was attributed largely to ERP modernisation cycles. One characterisation from that research is worth quoting for how precisely it captures the problem: a half-billion-dollar chemical company with a thin IT function does an S/4HANA upgrade roughly once every twenty years, and may well be in the middle of one right now.
This is the structural reason B2B agentic commerce will not follow the retail curve. The bottleneck is not model capability or protocol maturity. It is that the pricing logic an agent needs to query lives inside a system that gets touched once a decade, behind an integration layer that was built for nightly batch synchronisation rather than sub-second lookup.
Elogic Commerce’s own analysis of B2B readiness found roughly 20% of sellers considered themselves ready and around 18% rated their AI-commerce maturity as advanced — figures consistent with the Deloitte picture. For deeper treatment of the underlying platform and architecture questions, see our guides to B2B ecommerce for distributors and B2B ecommerce trends for 2026.
ERP readiness assessment
The question a distributor or manufacturer should be asking is narrow and answerable: can the systems we already run expose accurate, account-specific pricing and real-time availability to an external query, fast enough and safely enough to be useful to an agent?
The grid below is drawn from Elogic Commerce’s integration practice across these systems. It is deliberately framed as constraints and questions rather than vendor capability claims, because the honest answer for every one of these platforms is “it depends on your version, your deployment model and how your integration was built.”
| ERP | The constraint that usually bites | What to verify before committing |
|---|---|---|
| SAP S/4HANA | Pricing logic often lives in heavily customised condition tables; real-time exposure can conflict with performance governance. | Whether account-specific pricing can be resolved through an API call within acceptable latency, and who owns the pricing master. |
| SAP Business One | Lighter deployment, but integration is frequently middleware-dependent with batch-oriented sync. | Whether inventory and pricing can move to event-driven sync rather than scheduled batch. |
| Microsoft Dynamics 365 | Strong API surface, but trade agreements and pricing hierarchies can be complex to resolve externally. | Whether Dataverse and Finance & Operations pricing entities can be exposed consistently. |
| Oracle NetSuite | Capable API layer; governance limits and concurrency ceilings are the usual practical constraint. | Request throughput headroom under realistic agent query volume. |
| Infor | Industry-specific variants differ substantially; integration patterns are rarely portable between them. | Which variant is deployed and what the supported integration path actually is. |
| Epicor | Distribution-oriented and generally well suited to this, but customisation depth varies widely by install. | How much pricing logic sits in customisation versus standard configuration. |
| Visma | Regional variation across products; API maturity is not uniform across the portfolio. | Which specific product is in use and its current API capability. |
| Odoo | Flexible and API-accessible; data quality and discipline are more often the limitation than the platform. | Whether product data is complete and consistent enough to be machine-readable without cleanup. |
How to read this grid: These are field observations from integration engagements, not vendor-certified capability statements. Every one of them varies by version, deployment model and how the original integration was built. Verify against your own instance before making architectural commitments.
Not sure whether your ERP can support agent-initiated orders? Elogic Commerce runs a readiness assessment against your actual integration architecture. Contact Elogic Commerce.
The seven-point readiness audit
Run this before spending anything on agentic commerce tooling. Most organisations discover that five of the seven are data and integration problems they already needed to fix.
- Catalogue completeness. Does every sellable SKU carry the identifiers, attributes and descriptions an agent needs to match it to a stated need? Agents cannot infer what your sales team knows by heart.
- Price resolution. Can account-specific or contract pricing be resolved through an API call rather than a quote request? If pricing requires a human, agents will route around you.
- Availability accuracy. Is inventory real-time or nightly? An agent that recommends an out-of-stock item once will deprioritise the source.
- Order API integrity. Is order creation idempotent, and does it enforce credit limits, tax exemptions and approval thresholds server-side rather than in the storefront?
- Agent identification. Can you distinguish an agent request from a human one at the edge, and apply different rate limits, logging and authorisation to each?
- Machine-readable exposure. Is your product data available as structured data and a clean feed, not only as rendered HTML? This is where AEO and agentic readiness converge.
- Governance and audit. If an agent places an order that should not have been placed, can you reconstruct what it was authorised to do, by whom, and when?
Point seven is the one procurement and finance will ask about first, and it is the one most technical readiness assessments omit entirely.
Trust, fraud and liability
Delegating spending authority to software breaks assumptions that card-not-present rules were built on. Every serious protocol now treats verifiable trust as a first-class feature rather than a later addition, which is why the FIDO Alliance consolidation matters more than any individual product launch.
Three developments define the current liability picture. Visa’s Trusted Agent Protocol, built with Cloudflare, gives issuers and merchants a way to identify whether a transaction was initiated by a human or an agent and route liability accordingly. American Express went further on 14 April 2026, releasing its ACE Developer Kit alongside purchase protection covering erroneous agent purchases — the first meaningful signal that a network was prepared to absorb some of the risk rather than push it to merchants. And Verifiable Intent, now under FIDO governance, addresses the audit question directly by creating a tamper-proof record of what an agent was instructed to do.
The precedent nobody planned for
The most consequential development in agentic commerce during 2026 was not a protocol. It was a court ruling. Amazon obtained a preliminary injunction against Perplexity in the Northern District of California, under the Computer Fraud and Abuse Act, restricting Perplexity’s shopping agent from transacting on Amazon. Perplexity appealed to the Ninth Circuit in April 2026.
The unresolved question is whether a platform can lawfully exclude a third-party agent acting on behalf of a legitimate, authenticated customer. If the injunction survives appeal, large platforms gain a defensible basis for blocking agents they did not authorise — and the open agentic web becomes considerably less open. Amazon is meanwhile building its own assistant, Rufus, which it has credited with roughly $12bn in incremental annualised sales in 2025.
Status: This litigation is unresolved as of July 2026. Any strategy that assumes universal agent access to third-party marketplaces is making a legal bet, not a technical one.
For B2B specifically, the fraud surface looks different again. The risk is less stolen credentials and more an agent with legitimate credentials making a materially wrong decision at volume — ordering against the wrong contract, exceeding a credit limit, or triggering a replenishment loop. Server-side enforcement of every commercial control is not a best practice here; it is the entire safety model.
What to do in the next 90 days
Sequenced so that every stage produces value even if agentic commerce underdelivers against the forecasts. This is deliberate: none of the first three stages is a bet on agents.
| Window | Action | Why it pays off regardless |
|---|---|---|
| Days 1–30 | Audit product data completeness and structured data coverage across the catalogue. | Improves conventional search, marketplace syndication and PIM quality independently of agents. |
| Days 1–30 | Instrument AI-referred traffic separately in analytics. | You cannot manage a channel you cannot see. Most merchants still bucket this as direct or referral. |
| Days 31–60 | Map where pricing and availability truth actually lives, and measure real latency from ERP to storefront. | Exposes the integration debt that already degrades quoting, sales-assisted ordering and portal accuracy. |
| Days 31–60 | Run the seven-point readiness audit and rank gaps by cost to close. | Produces a defensible roadmap rather than a vendor-driven one. |
| Days 61–90 | Implement agent identification at the edge and decide your access policy deliberately. | You are already receiving agent traffic. Right now you are probably neither measuring nor governing it. |
| Days 61–90 | Pilot one narrow agent-assisted workflow — reorder, quote or availability lookup — behind existing controls. | Generates real operational data instead of vendor projections, at contained risk. |
Notice what is absent. There is no recommendation to implement ACP or UCP in the first ninety days. For most B2B merchants that would be premature: the protocols are still moving, the commerce layer has not converged, and the binding constraint is almost always data and integration quality rather than protocol support. Fix the foundation while the standards settle.
Elogic Commerce builds and integrates B2B commerce systems for manufacturers, distributors and wholesalers with real ERP complexity. Explore our ecommerce consulting services.
FAQ
Frequently asked questions
What is agentic commerce?
Agentic commerce is a model in which an AI agent discovers, evaluates and completes a purchase on a buyer’s behalf through machine-readable protocols rather than a graphical storefront. What separates it from a chatbot is delegated authority: the agent carries a cryptographically signed mandate proving it was authorised to transact within defined limits.
What is the Agentic Commerce Protocol?
ACP is an open standard co-developed by OpenAI and Stripe, launched on 29 September 2025 under an Apache 2.0 licence. It defines how an agent completes checkout with a merchant using a shared payment token and a structured product feed. The merchant remains merchant of record and the agent never sees raw card details.
What is the difference between ACP and UCP?
ACP, from OpenAI and Stripe, covers agent-initiated checkout. UCP, from Google and Shopify and launched at NRF in January 2026, covers the full journey including discovery, fulfilment and post-purchase. They are not currently interoperable, so merchants seeking broad coverage generally implement both.
Did ChatGPT Instant Checkout shut down?
The standalone Instant Checkout experience was retired around March 2026. OpenAI stated it was moving away from it to prioritise product discovery with merchant-owned checkout. Reported conversion data was the driver: in-chat checkout converted at roughly a third the rate of clicking through to the merchant site.
Who controls AP2 now?
The FIDO Alliance. Google donated the Agent Payments Protocol on 28 April 2026 and released v0.2 the same day, adding Human Not Present payments. Verifiable Intent, co-developed with Mastercard, was donated at the same time. FIDO formed working groups to develop both into industry specifications.
How large is the agentic commerce market?
Forecasts range from roughly $20.6bn in US AI-platform checkout for 2026 (eMarketer) to $3–5tn globally by 2030 (McKinsey). The thirty-five-fold spread is definitional rather than a disagreement about growth — the firms count different things, from agent-executed transactions to any AI-influenced purchase.
Does agentic commerce apply to B2B?
Yes, and Gartner projects that by 2028 around 90% of B2B buying will be AI-agent intermediated, representing over $15tn of spend. But intermediation is not autonomous purchasing. Reporting through 2026 indicates B2B is settling on agents assisting discovery while transactions remain inside ERP-governed merchant systems.
Why is B2B agentic commerce slower than retail?
Because the pricing an agent needs is not a catalogue price. It depends on account, contract, volume tier, currency and incoterm, and typically lives in the ERP. Deloitte Digital research reported in February 2026 attributed the gap between adoption intent and execution largely to ERP modernisation cycles.
What ERP changes does agentic commerce require?
Usually three: account-specific pricing resolvable through an API rather than a quote request, real-time rather than batch inventory, and server-side enforcement of credit limits, tax exemptions and approval thresholds. Most of this is integration work that improves quoting and portal accuracy regardless of agents.
Do AI-referred shoppers convert better?
They do now, which is a reversal. Adobe Analytics recorded AI-referred traffic converting 38% worse than other traffic in March 2025, then 42% better by March 2026 and 54% better by May 2026. These are vendor-stated figures, but the direction is corroborated by Shopify’s reported thirteen-fold growth in AI-driven orders.
Can platforms block third-party shopping agents?
That is being litigated. Amazon obtained a preliminary injunction against Perplexity in the Northern District of California under the Computer Fraud and Abuse Act; Perplexity appealed to the Ninth Circuit in April 2026. The outcome determines whether large platforms can lawfully exclude agents acting for authenticated customers.
How do I make my catalogue readable by AI agents?
Expose complete structured product data through a clean feed and schema markup, ensure availability reflects real stock rather than a nightly snapshot, and make pricing resolvable through an API. Agents query data, not rendered pages, so anything only visible in HTML is effectively invisible to them.
Should we implement ACP or UCP now?
For most B2B merchants, not yet. The commerce layer has not converged, the specifications are still revising on a roughly quarterly cadence, and the binding constraint is nearly always data and integration quality. Fix the foundation first — it is prerequisite work for either protocol.
What is the difference between agentic commerce and conversational commerce?
Conversational commerce uses a chat interface to recommend products and then hands the buyer to a checkout. Agentic commerce gives the software authority to transact, backed by a signed mandate and a tokenised credential. The mandate is the dividing line, not the presence of a conversation.
How should we measure agentic commerce performance?
Separate AI-referred traffic in analytics as its own channel, then track citation share — how often your products and brand appear in AI-generated answers for commercially relevant queries. As transactions move to merchant infrastructure, visibility inside the agent becomes the leading indicator, not checkout volume.
Changelog
| Date | Change |
|---|---|
| 29 Jul 2026 | Initial publication. All protocol statuses, market figures and litigation status verified as of this date. |
Sources
- OpenAI and Stripe — Agentic Commerce Protocol specification and repository, agenticcommerce.dev and github.com/agentic-commerce-protocol (launched 29 Sep 2025; spec 2026-04-17)
- Stripe — “Developing an open standard for agentic commerce,” stripe.com/blog
- OpenAI — merchant guidance on product discovery, chatgpt.com/merchants
- Wired — reporting on the withdrawal of Instant Checkout and Walmart conversion data, 18 Mar 2026; corroborated by CNBC
- Google — “Google donates Agent Payments Protocol to FIDO Alliance,” blog.google, 28 Apr 2026
- FIDO Alliance — announcement of AP2 donation and Agentic Authentication Technical Working Group, fidoalliance.org, Apr 2026
- PYMNTS — “Google and Mastercard Contribute Agentic Commerce Standards to FIDO Alliance,” 28 Apr 2026
- Google Cloud — “Announcing Agent Payments Protocol (AP2),” 16 Sep 2025
- Visa — Intelligent Commerce Connect and Trusted Agent Protocol announcements, Oct 2025 and Apr 2026
- Mastercard — Agent Pay and Agentic Tokens, developer.mastercard.com
- Digital Commerce 360 — “American Express launches developer kit, purchase protection for agentic commerce,” 14 Apr 2026
- Digital Commerce 360 — “Agentic commerce faces reality check in B2B ecommerce,” 10 Mar 2026
- Digital Commerce 360 — “Deloitte Digital: B2B suppliers lag on agentic AI as ERP upgrades slow adoption,” 25 Feb 2026
- Bain & Company — “2030 Forecast: How Agentic AI Will Reshape US Retail”
- McKinsey — “The Agentic Commerce Opportunity,” Oct 2025
- Morgan Stanley AlphaWise — agentic commerce forecast, Nov 2025
- eMarketer — US AI-platform checkout forecast, Dec 2025
- Gartner — “Top Predictions for IT Organizations and Users in 2026 and Beyond,” 21 Oct 2025
- Adobe Analytics — AI-referred traffic and conversion data, Q1 and May 2026 releases
- Shopify — Form 8-K, 5 May 2026
- Salesforce — Cyber Week 2025 shopping data, 5 Dec 2025
- Amazon.com Inc. v. Perplexity AI Inc., N.D. Cal., preliminary injunction Mar 2026; appeal to the Ninth Circuit Apr 2026
Get in Touch
Looking for a partner to grow your business? We are the right company to bring your webstore to success.