On September 2, Anthropic published a blueprint for building commerce agents on Claude, including two open-source agents. One helps shoppers search catalogs, compare products and assemble carts. The other supports merchant operations by analyzing performance, identifying inventory issues and drafting campaigns, with changes held for human approval.

The release is well considered, and much of it is useful. It also reflects the consumer-commerce assumptions behind most agentic commerce development today. The agents work with catalogs, carts, campaigns and card-based checkout. They do not address many of the commercial rules that shape a typical B2B transaction: contract pricing, account entitlements, approval authority, purchase orders and credit terms.

That gap prompted us to look beyond the announcements. We reviewed four agent protocols, six commerce platforms and several of the procurement suites operating on the buyer side. The question was practical: what can a manufacturer or distributor deploy today, and what remains dependent on capabilities the market has not built?

The answer is not that agentic commerce has no value in B2B. Some applications are already delivering measurable results. But the strongest opportunities are not necessarily the ones receiving the most attention.

A practical readiness model

The market becomes easier to assess when the available capabilities are separated into four tiers.

Tier	What it covers	Status
1	Seller-side and internal agents: order intake, service assist, merchant operations copilots, catalog enrichment	Ready now
2	Agent-assisted quoting inside first-party channels with an authenticated buyer	Real, constrained
3	Agentic storefronts on UCP and ACP, built primarily around consumer transaction assumptions	Limited B2B fit
4	Contract pricing for third-party agents, entitlements, account hierarchy, approval chains, net terms, punchout, EDI, agent identity and authority to bind	Not yet supported

Most public discussion focuses on agentic storefronts in Tier 3. The more immediate B2B opportunities are concentrated in Tier 1.

Tier 1: ready for practical deployment

The B2B agents with the clearest production evidence generally operate within one company’s systems and controls.

Order intake is a good example. commercetools shipped a B2B Intake Agent in June, co-designed with manufacturer Mirion Technologies. It converts emails, PDFs, spreadsheets and other offline files into structured quotes and carts, matches SKUs and quantities to catalog data, then routes each request to the appropriate business unit, customer account and pricing structure. The agent handles work that would otherwise require a customer service representative to interpret and enter the request manually.

Conexiom provides a larger and more established precedent. It processes more than 1.5 billion line items a year across roughly 40 compatible ERP systems for customers including Sonepar, Graybar and Johnstone Supply NW. The company emphasizes transparent, rule-governed automation—”AI you can see,” in its CEO’s description—rather than broad autonomy. That positioning is consistent with what many B2B organizations appear willing to put into production: automation that can be reviewed, governed and traced.

Merchant operations tools are another established category. Shopify’s Sidekick, powered by Claude Sonnet 4.5, helps merchants work with analytics and store operations. BigCommerce announced its Companion assistant at Commerce Live in April, and Adobe’s CX Enterprise Coworker reached general availability in June. Mirakl says its Catalog Transformer can onboard supplier catalogs in under 24 hours, with most results ready in about two.

These applications may attract less attention than autonomous purchasing, but they address identifiable work and can be measured against labor, cycle time and error rates.

Tier 2: useful within a controlled channel

Salesforce made Agentforce Commerce generally available in late June. Its B2B Buyer Agent supports reordering over WhatsApp and SMS against current contract pricing, including round-trip quoting between cart and quote.

The architectural boundary matters. The agent operates in a first-party messaging channel with a known, authenticated buyer rather than through an open agent protocol. B2B logic remains inside Salesforce, where the system can identify the customer and apply the correct commercial rules. The capability is real, but it depends on staying within that controlled environment.

Tier 3: live, but built for consumer commerce

Agentic storefronts are now operating in the market. Google’s UCP-powered checkout runs in AI Mode and the Gemini app. Universal Cart began rolling out this summer across Search and Gemini with Nike, Sephora, Target, Ulta Beauty, Walmart and Wayfair. Shopify enabled UCP by default for eligible merchants, Adobe committed to both major protocols in February, and BigCommerce has an MCP server in beta.

For consumer brands, these surfaces may become a meaningful acquisition and transaction channel. B2B sellers face a different set of constraints, which Shopify states plainly in its documentation:

If you sell the same products to both B2B and D2C customers, then your products are included in your agentic storefronts using the D2C price that you set. B2B pricing doesn’t display on agentic storefronts.

Shopify also excludes B2B-only products from agentic storefronts automatically. BigCommerce lists a B2B storefront for its MCP server as coming soon. Adobe has shipped B2B drop-in components for negotiable quotes, requisition lists and purchase-order workflows, but has not published a connection between those components and its Commerce MCP surface. At present, that connection should be treated as undocumented rather than assumed.

Hybrid B2B and D2C operators have an additional issue to review. Shopify warns that its exclusion logic may not detect custom B2B configurations built with third-party apps or theme modifications. In those cases, B2B products could appear on an agentic storefront unintentionally. Hybrid stores should include this configuration in their current channel audit.

Tier 4: the unresolved B2B transaction layer

The limitations become clearer at the protocol level. ACP, from OpenAI and Stripe, remains in beta following its April release. UCP, from Google and Shopify, was updated in late August. AP2 moved to the FIDO Alliance in April, while MCP is hosted by the Linux Foundation’s Agentic AI Foundation.

Across these specifications, we found no representation of customer-specific contract pricing, corporate account hierarchy, entitlements, approval chains, purchase orders, net terms, or mappings to punchout and EDI. These are not edge cases in B2B commerce; they determine who can buy, what they can buy, at what price and under whose authority.

The gap is structural. Current commerce protocols generally assume a published catalog with a public price, an individual buyer and card settlement. A typical B2B transaction works differently: price may depend on the account and contract, the buyer acts within an organization under delegated authority, and payment may occur later by invoice. The commerce specifications do not yet provide a way to carry that context through an external agent transaction.

Identity standards present a related constraint. A survey published this spring found no production-ready mechanism to trace authorization across multiple agent hops, narrow permissions as work moves between organizations, or correlate an audit trail when one company’s agent delegates to another’s. Those gaps matter if a buyer agent calls a procurement agent that then communicates with a seller agent. We also found no published standard for a machine-verifiable credential confirming that an agent has authority to bind a company to a purchase.

Regulatory obligations are developing while the technical standards remain incomplete. Article 50 of the EU AI Act took effect on August 2. The Commission’s guidance brings agentic systems within its disclosure obligations, with penalties reaching €15 million or 3% of global turnover. Organizations operating quoting or service agents in the EU should assess whether those interactions fall within the current requirements.

What the market forecasts actually say

Gartner projects that 90% of B2B buying will be AI-agent intermediated by 2028, representing more than $15 trillion in spend. The figure now appears frequently in agentic commerce presentations.

The interpretation depends heavily on the word “intermediated,” which Gartner has not publicly defined in detail. It could include a buyer using an LLM to research suppliers as well as an agent placing an order. The projection is useful evidence that AI will influence buying research and routing, but it should not be read as a forecast that $15 trillion in transactions will be executed autonomously.

Forrester offers a narrower forecast: 20% of B2B sellers will be required to engage in agent-led quote negotiations during 2026. That prediction points to a more immediate planning question—whether sellers can respond when an agent participates in the buying process—even if the final transaction still follows established B2B controls.

Gartner’s broader agentic AI research adds some caution. It expects more than 40% of agentic AI projects to be canceled by the end of 2027 and estimates that only about 130 of the thousands of vendors claiming agentic capabilities meet its definition. The category is advancing, but the label remains much broader than the production evidence.

The buy side is moving first

Current evidence suggests that B2B agent activity will reach many sellers through procurement systems before it reaches them through consumer-style chat or search interfaces.

Coupa reports more than 450 customers running its agents in production and more than 400 custom agents built in Agent Studio. One payment-batch agent ran 14 batches comprising 2,395 payments and $20.1 million in its first five weeks. Coupa also exposes more than 30 procurement, invoicing and contract tools through MCP for external AI systems including Microsoft Copilot, with agent-to-agent connectivity expected this month. SAP shipped Joule agents across Ariba and Fieldglass in June.

These examples operate inside the buyer’s governance environment. They can use an established approval matrix, spend limits, audit log and supplier master, while keeping authority to commit the organization within existing controls. That environment makes deployment materially easier than an external agent transaction spanning several companies.

Existing B2B channel patterns reinforce the point. Grainger’s disclosures show substantial order volume originating through its website, eProcurement and EDI connections, and KeepStock inventory programs rather than open-web search. Conexiom’s 1.5 billion annual line items reflect another large stream of orders arriving through documents and system-to-system traffic. An agentic roadmap built only around storefront discovery would overlook much of this activity.

Consumer results also argue for measured expectations. OpenAI phased out in-chat Instant Checkout in March, roughly six months after launch, in favor of merchant-handled app checkout. Walmart said its in-chat purchases had converted at about one-third the rate of click-out transactions. The result does not predict B2B performance, but it shows that a more agentic transaction is not automatically a better-performing one, even in consumer commerce.

For B2B sellers, the practical implication is to prepare the channels where buyers already transact while continuing to monitor the newer storefront protocols.

Priorities for the next 12 months

Four workstreams stand out from the current evidence.

1. Make commerce data machine-readable and current

Agents cannot reliably use contract pricing, entitlements or inventory that resolve only through a nightly ERP batch. Improving access to that information is foundational integration work, regardless of which agent interface eventually calls it. Deloitte found that 87% of B2B suppliers are upgrading or planning an ERP upgrade, creating a practical opportunity to address these data paths as part of work already underway.

2. Treat punchout and procurement networks as an agent on-ramp

A buyer’s agent may be more likely to reach a seller through Coupa, Ariba or JAGGAER than through a consumer AI surface. Sellers should assess their coverage of those networks, the quality of the catalog and pricing data exposed through them, and any customers still relying on manual workarounds.

3. Deploy Tier 1 where the operating case is clear

Order intake, service assistance, catalog enrichment and quoting support can be evaluated against labor, cycle time, accuracy and service levels. Those measures provide a more defensible business case today than an assumed conversion lift from agentic storefronts.

4. Define the controls that current standards do not provide

Cross-company standards for identity, delegation, audit trails and authority to bind remain incomplete. Any implementation that moves beyond internal assistance should state how those controls will work, who owns them and where human approval remains required.

What is ready—and what is not

Agentic commerce is real, but the term covers capabilities at very different stages of maturity. Internal and seller-side agents are already producing measurable results. First-party buyer agents can support B2B transactions when identity and commercial logic remain inside a controlled platform. Consumer-oriented agentic storefronts are live, but they do not yet carry the context required for many B2B purchases.

The narrowest version of B2B agentic commerce—an external third-party agent transacting autonomously against contract pricing on credit terms—is not yet established. We found no named production deployment with a quantified result, and the current specifications do not represent several commercial concepts required to support one.

That does not call for waiting. Manufacturers and distributors can deploy the operational use cases that are ready, improve the pricing and entitlement services future agents will depend on, and strengthen their connections to procurement networks. Those investments have value now and preserve options as the standards mature.

Workstream 1 starts with a plain question: how legible is your commerce data to a machine today? The McFadyen AI Commerce Readiness Audit answers that part for free, with an automated check of how your storefront and product data present to AI systems. It covers the data foundation every tier above depends on, not the protocol gaps described in Tier 4.

B2B Agentic Commerce in 2026: What Actually Works, in Four Tiers