Agentic Commerce in B2B: Following One Order from Agent to Delivery
Agentic commerce lets AI agents carry out shopping tasks on a buyer’s behalf, from finding suitable products to preparing a purchase. For a B2B supplier, the practical question is whether those tasks can proceed under the right account, with the right prices and purchasing controls.
TL;DR: Agentic Commerce in B2B, in Brief
- Public product information helps an agent identify a suitable supplier.
- Private account context determines the available assortment and commercial terms.
- A valid order depends on the seller’s data and rules, including contract pricing and approval.
- In this walkthrough, the agent assembles the cart; the buyer confirms and places the order in the seller’s checkout.
What Is Agentic Commerce?
Agentic commerce is the use of AI agents to perform shopping tasks on someone’s behalf, with authority defined by the user and participating systems. In B2B, those tasks must account for the buying organization’s products, terms and permissions, even when a person retains the final purchase decision.
The difference becomes clearer when a buyer needs to get something done. An assistant that explains a filter’s specifications answers a question. An agent that retrieves approved filters, checks account pricing and builds a cart advances a purchasing task. Degrees of delegation vary; human confirmation can remain part of the process.
Machine activity is growing, although it should not be confused with purchasing. The joint playbook reports Prerender’s analysis of 20 enterprise retailers, where AI crawler traffic grew approximately 3.7-fold over a 12-month period. This is retail crawling data, not a measure of B2B orders placed through agents.
One B2B Order, Built by an AI Agent
A procurement agent needs the seller to answer progressively more specific questions: whether it carries the products, whether this account can buy them, and whether the proposed order meets its requirements. The following fictional example follows those questions through a buyer-confirmed purchase and the delivery updates that follow.
A building services contractor operates three job sites and buys from a regional distributor. On Monday, its purchasing manager instructs the agent: “Restock site 3: fittings and filters, delivery by Thursday.” The manager supplies the site’s approved replenishment list, including quantities and equipment references. The resulting cart totals \$2,340. All scenario details and amounts are illustrative.
Each stage depends on information or controls supplied by the seller.
Step 1. Find the supplier
The agent first needs to establish that the distributor carries the required kinds of fittings and filters and serves the job site’s location.
That starts with the public catalog: product identity, categories, core specifications and general availability signals. Negotiated prices need not be exposed. Deciding what AI assistants can see in a B2B catalog gives the seller a useful public presence while protecting customer terms.
Clear business identity, shipping policies and machine-readable capabilities also support agent discoverability. Universal Commerce Protocol, or UCP, provides a standard way to discover a store’s supported commerce capabilities.
Inaccessible information can prevent the agent from establishing suitability. The Invisible Catalog explains the underlying visibility and access decisions.
Step 2. Act for the right account
The seller must establish which buyer and organization the agent is authorized to represent before returning private commercial information.
The contractor’s account determines its terms and assortment. The agent therefore needs an authenticated buying context, rather than a company name entered as ordinary text.
This is the boundary between the catalog’s public and private layers. People and machines can see the public information; access rights determine who receives the fuller commercial representation. The UCP identity specification likewise distinguishes an authenticated agent from a request authorized on behalf of a user.
As Oleg Zhuk, CTO of Virto Commerce, explains in the playbook:
“In B2B, a product only exists in the context of a buyer. So the question is never just ‘can AI see my catalog.’ It’s ‘which version of my catalog should AI see, and where does that stop.’”
Without that context, a plausible-looking cart could contain the wrong prices or products the contractor cannot purchase.
Step 3. Match the products
The agent must convert the replenishment list into specific catalog items without changing what the site needs.
For the fittings, that means matching the specified diameter, material and connection type. For the filters, it means checking the equipment reference and required dimensions. A similar description is insufficient when a physical difference makes the part unusable.
Units of measure matter just as much. A box of 50 fittings is a different order quantity from an individual fitting. The response must make the selling unit and pack contents clear enough to compare with the site’s requested quantity.
The agent also needs the assortment defined by the account’s contract-based catalogs. An alternative filter should be considered only within the contractor’s permitted substitution rules. If the list leaves a specification unresolved, the agent should ask for clarification before adding a replacement.
Step 4. Check price, stock and delivery
The agent needs a purchasable offer for the selected items, including the account price and a delivery option that meets the site’s deadline.
The seller must return the applicable contract price, any qualifying volume discount and stock information relevant to fulfillment. “In stock” somewhere in the network does not answer whether the warehouse serving this site can dispatch the filters in time.
A compatible agent can use Universal Commerce Protocol to build a cart at the signed-in customer’s contract prices. Where supported, Model Context Protocol provides the connection through which an AI application calls external tools and retrieves data. The seller’s systems still supply the commercial answer.
Before its platform change, Lavazza by Bluespresso maintained individual customer price lists that were not consolidated online. Digital ordering needs access to those agreed prices.
If price, stock or delivery cannot be confirmed, the agent should flag the unresolved requirement rather than present an estimate as a commitment.
Step 5. Respect the limits
The contractor’s purchasing limit must be enforced by the seller’s ordering process, regardless of how the cart was assembled.
In this example, the cart exceeds the manager’s purchasing authority and requires an approver. The contractor’s rule is configured in the distributor’s portal. The agent must be able to discover that approval is required, and the workflow must prevent the purchase from proceeding without it.
Approval workflows can route purchases according to spending limits, buyer roles and organizational responsibilities. The requirement here is that the same control remains effective when an agent prepares the request. A budget written into the original prompt cannot substitute for that control.
If a quantity or substitute changes during review, the process must reassess whatever approval conditions that change affects.
Step 6. Hand off and place the order
The completed cart must reach the buyer intact so the purchase can be reviewed and placed through the seller’s checkout.
The buyer checks the line items, supplies the purchase order number and confirms the job-site address. The account’s agreed payment terms, Net 30 in this example, govern payment. Delivery and order placement are completed in the storefront, with the seller remaining the merchant of record.
This keeps the final commitment with a person on the buyer’s side. Preparing a purchase does not grant the agent authority to submit it.
The handoff also needs to withstand interruptions. A retried request should recover the intended state rather than create an additional cart or repeat a completed operation. The seller should be able to tell the buyer whether the order succeeded before anyone tries again.
Step 7. Follow the order through delivery
The order must remain understandable after checkout, when shipment events and changes determine whether the site receives what it needs.
Suppose the fittings ship first and the filters follow separately. An authorized agent retrieving status needs line-level information to explain what remains outstanding and whether the original delivery commitment still holds. It should not report the whole order as delivered because one package arrived.
onX, the Commerce Operations Foundation’s Order Network eXchange standard, provides a common language for order, inventory and fulfillment operations across compatible systems. Here, its role is to connect the order with warehouse and logistics updates.
Rescheduling, cancellation and return requests must reach a system that can accept or reject them under the seller’s rules. Acknowledging a request is different from confirming a change.
Both shipments arrive by the agreed deadline. A later scheduled replenishment can prepare another cart, checking current terms before buyer confirmation.
Where This Order Can Break
An agent-prepared purchase can fail wherever the seller cannot supply the required information, enforce a rule or preserve transaction state. The table below turns the walkthrough into checks for the commerce team. These are potential failure points to investigate, not a ranking of their frequency or severity.
|
Step
|
What goes wrong
|
What the seller’s platform must provide
|
|---|---|---|
|
Find the supplier
|
The relevant products or service area cannot be identified.
|
Accessible public product information, business identity and supported capabilities.
|
|
Right account
|
The agent sees list prices or an unrelated assortment.
|
Authenticated buyer context and access controls for private terms.
|
|
Match products
|
A selling unit is misread or an unapproved substitute is selected.
|
Explicit units, specifications, permitted assortment and substitution constraints.
|
|
Price and stock
|
Customer prices remain offline, or stock cannot be tied to delivery.
|
Readable account prices, relevant inventory and available delivery options.
|
|
Approval
|
The purchase proceeds without the required authorization.
|
Enforced approval rules that remain effective for agent-prepared purchases.
|
|
Handoff
|
The cart is lost, duplicated or submitted again after an uncertain response.
|
Persistent cart state, buyer-bound handoff and safe handling of repeated operations.
|
|
After the order
|
Shipment changes or return requests cannot reach the responsible system.
|
Authorized status access and connected workflows for post-order requests.
|
Fig. Seven failure points in an agent-prepared B2B order.
Can an agent access the catalog information this order needs?
How to Check Whether an Agent Can Buy from You
Test an ordinary repeat purchase against the information and controls available in your platform. Use a real account with known terms, then compare the agent-prepared result with the expected order. Record where the process needs manual intervention, what caused it and which team owns the missing answer.
- Discovery: Have we decided which catalog information is public, which requires authentication, and whether an agent can find the public portion?
- Identity: Can the seller establish the correct buying account and restrict access to its authorized commercial terms?
- Matching: Can the requested items be resolved to permitted SKUs with unambiguous selling units and pack contents?
- Offer: Can the agent retrieve account prices, relevant stock and a delivery option that satisfies the request?
- Approval: Does a purchase requiring authorization remain blocked until the appropriate approval is recorded?
- Handoff: Does the buyer receive the intended cart, and can interrupted or repeated operations be recovered without duplication?
- Fulfillment: Can an authorized request retrieve meaningful shipment status and reach the correct process for a change or return?
Turn each unanswered question into a specific integration, data or workflow task.
Put AI to work in B2B commerce
How Virto Commerce Supports This Order
Virto Commerce provides the seller-side commerce capabilities that a compatible procurement agent needs to interact with a business. Its role here is to expose the catalog and ordering processes while preserving account context and seller control. It does not build buyers’ procurement agents or create and train foundation models.
The UCP adapter connects compatible agents to existing commerce APIs without a separate integration for each agent. Contract prices and assortment remain account-specific; approval requirements belong in the seller’s configured workflows. The agent prepares the cart, while the buyer completes the purchase in the store’s checkout.
In Virto Commerce, catalogs, price lists, and contracts are separate entities, and xAPI returns product information according to access rights. This lets you make products publicly discoverable while keeping customer-specific prices and terms available only to authorized buyers and agents.
As a founding member of the Commerce Operations Foundation, Virto Commerce also provides an onX adapter for connected order operations. For customers using the seller’s own interface, ordering in plain language uses order history and account context to prepare an editable, priced cart.
Give compatible agents a route to your checkout
FAQs
It translates a buyer’s request into actions such as finding products, retrieving account terms and preparing a cart. In the workflow described here, the buyer then reviews and places the order in the supplier’s checkout. Access and permitted actions depend on the participating systems.
In a conventional storefront journey, the buyer performs each search, comparison and cart update. With an agent, some of that work is delegated through a goal expressed in natural language. The supplier still needs to provide reliable commercial information and enforce the applicable purchasing rules.
No. Software can turn a handwritten count sheet into a cart without independently pursuing a purchasing goal on the buyer’s behalf. That is AI-assisted document processing. It can support an agent workflow, but the extraction step alone does not establish one.
Choose a representative repeat order and test the seller-side requirements from discovery through fulfillment. Establish the public/private catalog boundary, then check account access, product matching, commercial terms, approval and checkout handoff. Assign owners to gaps before expanding the range of tasks an agent can perform.