Most examples of agentic commerce start with a consumer use case. Someone asks an AI assistant to find a pair of running shoes, compare a few options and buy the best match. Product discovery becomes conversational, the assistant can increasingly interact with real commerce systems, and the storefront is no longer the only place where a buying journey can happen.
B2B purchasing brings a different set of requirements.
A purchasing manager rarely buys something based on a simple product comparison. They may be ordering on behalf of a specific company location, from a negotiated catalog, at a customer-specific price and under payment terms agreed months earlier. Certain products may require approval, some alternatives may be excluded, and larger orders often involve additional people before anything can be submitted.
Once AI becomes part of this process, the main challenge moves into an area that enterprise commerce teams already know well: identity, permissions, pricing, business rules and system ownership.
An agent may be able to find a product and prepare an order fairly easily. Understanding whether it is allowed to do so for a specific company, under the correct commercial conditions and within the buyer’s authority is where things become more interesting.
B2B commerce has always required more context
Consumer identity is usually straightforward. A person has an account, perhaps a saved address and payment method, and the merchant can personalize the experience around that individual.
B2B identity contains more layers.
Shopify’s B2B model already reflects this. A customer can act on behalf of a company and a specific company location, with that context determining things such as available products, pricing, payment terms, addresses, tax settings and checkout behavior.
That context becomes even more important when an AI agent enters the process.
Imagine someone tells their internal assistant:
Reorder the packaging material we usually buy for our Cologne warehouse and make sure we have enough for the next six weeks.
The request sounds simple because the person asking already understands the surrounding business context. A system executing the request has to reconstruct it.
Which legal entity is making the purchase? Which location should receive the goods? Which supplier catalog applies? Is there an agreed SKU that should always be used? Can another product be substituted if it is unavailable? Which price applies to this customer? Does the buyer have permission to spend €2,000, €20,000 or €200,000? Does the order need approval?
Companies already deal with these questions in B2B ecommerce projects today. Agentic commerce makes them more visible because more of that context has to become understandable to software.
Product discovery is probably the easy part
A lot of current work around agentic commerce focuses on helping machines discover and understand products.
Shopify is investing heavily in this direction. Its agentic commerce infrastructure now includes Shopify Catalog for structured product discovery, the Universal Commerce Protocol for interactions between agents and merchants, and MCP-based capabilities around catalog, cart and checkout experiences.
For merchants, this makes structured and reliable product data increasingly important.
In B2B commerce, however, product data only becomes useful once it is placed in the context of the customer.
A product can exist in the overall assortment without being available to a particular company. Its list price may be irrelevant because the customer has negotiated different terms. One business may purchase the product individually while another buys it by the pallet. Availability can depend on the market, warehouse or company location.
An agent therefore needs to understand more than the product itself. It needs access to the same commercial context that the authenticated buyer would receive inside the B2B store.
This creates a practical architecture question: where does that context come from, and which system is responsible for it?
In many enterprise environments, product attributes come from a PIM, contract prices originate in an ERP, customer information sits in a CRM and Shopify handles the commerce transaction. Those relationships may already work well enough for a human storefront experience, even if parts of them remain implicit.
Agents will put more pressure on companies to make them explicit.
Pricing becomes more contextual
Pricing shows this particularly well.
A consumer might ask an assistant to find the cheapest suitable laptop. Comparing current retail prices is relatively straightforward.
A B2B buyer could ask:
What will this laptop actually cost us if we order 80 units for our German entity under the current agreement?
There may be several correct answers depending on the customer, company location, quantity, market, currency and contract.
Shopify B2B already supports parts of this through catalogs, company locations, quantity rules and B2B-specific payment terms. The complexity usually starts when Shopify is only one of several systems involved in determining the final commercial conditions.
A company might maintain contract pricing in SAP, product information in a PIM, account data in Salesforce and availability in another operational system. A human buyer usually does not need to know where all of this information comes from. They log in and expect the store to show the correct result.
An AI agent needs the same result, together with enough structured context to know that it is valid for the company it currently represents.
This is one reason I expect agentic commerce to increase the importance of clear system ownership. If teams cannot answer where the authoritative price, catalog assignment or customer entitlement comes from, exposing another purchasing interface will make the ambiguity harder to manage.
An AI agent also needs a defined level of authority
Access to the right data is only one part of the process.
A procurement assistant could reasonably search products, compare contracted prices, check availability and prepare an order without much risk. Letting the same system submit a €150,000 purchase order requires a much clearer definition of responsibility.
Most companies will probably need several levels of delegated authority.
An agent might be allowed to reorder known products up to €5,000 automatically. Larger purchases could be prepared and sent to the buyer for confirmation. Orders above another threshold might require approval from a procurement manager. Product substitutions could trigger approval regardless of the order value.
These workflows already exist in many B2B businesses. AI changes how often software will initiate or prepare the transaction, which makes the boundaries more important.
Shopify itself already supports B2B setups where orders can be submitted directly or created as drafts for review. Similar patterns can be extended much further once the ordering process starts outside the storefront.
The key architectural question becomes how the system represents delegated authority in a reliable way.
Approval workflows could become more useful
AI is often discussed in terms of removing people from processes. In B2B purchasing, a more realistic goal is to remove repetitive work while keeping human decisions where they still add value.
Think about a recurring purchase that currently requires someone to notice that stock is running low, check previous orders, search the supplier portal, compare products, calculate quantities, create a cart and then ask a manager for approval.
An agent could handle much of that preparation.
It could monitor usage, identify what needs replenishment, find the contracted products, calculate a sensible quantity, check availability and prepare the transaction. The responsible person could then receive something like:
Based on the last six months of usage, we should reorder 240 units. The contracted price is €18.40 per unit, all products are available and the total order value is €4,416. This matches the previous three orders. Approve?
That already saves meaningful time without requiring a company to hand over full purchasing authority.
For many B2B use cases, this kind of workflow may be more valuable than complete autonomy. The routine parts happen automatically, while financial responsibility remains with the people who already own it.
Important business rules should stay enforceable
As companies start experimenting with internal agents, there will be a temptation to put more and more business logic into prompts or agent instructions.
A rule such as:
Never place orders above €10,000 without asking the procurement manager.
can certainly be useful as guidance. It should not be the only place where that rule exists.
If a buyer has a €10,000 purchasing limit, the underlying system should enforce it regardless of whether the request originates from a website, an internal application, an API or an AI assistant. The same principle applies to product eligibility, catalog access, customer-specific pricing, approval thresholds and other commercial rules.
This matters because enterprise commerce environments change constantly. Teams introduce new channels, replace ERPs, add markets and connect additional applications. Business rules become much easier to maintain when they live in well-defined systems rather than inside the behavior of one particular interface.
Agents can make these rules easier to use, but they should interact with capabilities that already exist and are enforceable elsewhere.
Otherwise the AI layer risks becoming another place where business logic accumulates over time, which would create exactly the kind of maintenance problem many commerce teams are trying to reduce.
The storefront becomes one interface among several
For years, ecommerce architecture has been built around the assumption that customers eventually arrive at the storefront.
They may discover the brand through Google, social media, a marketplace or an email campaign, but at some point they usually visit the merchant’s website and continue the journey there.
Agentic commerce makes that assumption less reliable.
Shopify is already building infrastructure that allows more commerce interactions to happen through AI interfaces. Catalog, UCP, Agentic Storefronts and MCP all point in this direction.
For B2B companies, the range of interfaces could become particularly broad.
A buyer might interact with an AI assistant inside the supplier’s website. Another company could use an internal procurement agent connected to several suppliers. Orders could begin in Microsoft Teams, an ERP interface, a procurement platform or an industry-specific application.
The Shopify storefront will still matter. People will continue to need product pages, account management, order history, service workflows and many other interfaces that work better visually.
At the same time, the storefront may gradually become one client of the commerce platform among several.
That changes how companies should think about architecture. Product data, pricing rules, identity, entitlements and transaction capabilities need to be usable beyond the frontend that currently exposes them.
What a realistic B2B agent flow might look like
A useful way to think about this is to follow a normal purchasing process.
A user asks an agent to reorder a regularly purchased product.
The system first needs to establish who the user is, which company they represent and which company location they are acting for. That determines the relevant catalog, commercial conditions, shipping context and potentially the payment terms.
The agent can then interpret the request, retrieve the necessary product and commerce data and prepare the transaction.
Before anything is submitted, the system checks whether the requested action falls within the user’s authority. If it does, the order can proceed. If additional approval is required, the transaction can be turned into an approval request or draft.
The resulting order should also contain enough context for the business to understand what happened later. Who initiated it? Which agent acted on the request? Which rules were evaluated? Who approved the transaction? What eventually entered the ERP?
When you break the process down this way, the AI model is only one part of the architecture.
Identity, pricing, catalogs, permissions, workflow orchestration, transaction execution and auditability all matter just as much.
That is very close to what we already see in complex Shopify Plus projects. The quality of the storefront is important, but a large part of the real work sits behind it in the systems and processes that determine how commerce actually operates.
Agentic commerce will expose weak B2B architecture quickly
Many B2B commerce platforms work today because experienced people know how to compensate for inconsistencies.
A salesperson knows that a certain customer always receives an extra discount. Someone in operations knows which warehouse should actually fulfill a specific item. A buyer knows that one SKU should never be ordered for the Austrian subsidiary, even if the system technically allows it.
This kind of knowledge often lives in people, spreadsheets, email threads and unwritten processes.
That is manageable as long as humans remain closely involved in every transaction. It becomes harder once companies want software to handle more of the work.
If an important rule only exists in someone’s head, the agent cannot reliably apply it. If the source of truth for a price is unclear, giving an AI access to five systems does not resolve the ambiguity. If ERP and PIM data disagree about availability, the agent still needs a defined rule for which system to trust.
In that sense, agentic commerce creates a useful test for B2B architecture:
Could another system understand enough of your commercial rules to prepare a correct order without an experienced employee interpreting the data first?
For many organizations, the answer today would probably be no.
That is not necessarily a problem. Fully autonomous purchasing is still an emerging area. But the same work that makes a commerce architecture easier for agents to understand also tends to improve integrations, operations and long-term maintainability.
What enterprise teams can prepare today
There is no need to redesign an entire B2B platform around autonomous buyers before there is a real use case.
Some preparation is useful regardless of how quickly agentic purchasing develops.
Product and commercial data should have clear ownership. If Shopify, the ERP, PIM and CRM disagree, every additional channel inherits the same problem.
Permissions should also distinguish clearly between information access and transactional authority. A user or agent may be allowed to retrieve prices and prepare orders without having the ability to submit them.
Approval workflows should be treated as a normal part of digital commerce architecture. Some transactions can be automated end to end, while others should intentionally stop for review.
It is also worth keeping business rules outside the AI layer wherever possible. Agents work best when they interact with clearly defined commerce capabilities rather than carrying dozens of hidden rules inside their own instructions.
These are fairly conventional architecture principles. Agentic commerce simply gives companies another reason to take them seriously.
B2B agentic commerce is a delegation problem
Shopify’s investments in Catalog, UCP, Agentic Storefronts and MCP make the broader direction clear. More commerce journeys will begin in AI interfaces, and some will also finish there.
In B2C, much of the current focus is naturally on product discovery, recommendations and conversion.
B2B introduces another layer because the person requesting something, the company paying for it and the system executing the transaction can all be separate actors.
An agent needs enough context to understand which organization it represents, which catalog applies, what the agreed price is, which payment and delivery terms are valid and how far the user’s purchasing authority extends.
Sometimes it may be able to complete the order. In other situations it will prepare the transaction and ask for approval. In some cases it should stop because the available data or permissions are not sufficient.
Those boundaries will be central to how useful and trustworthy agentic B2B commerce becomes.
For enterprise teams, the work therefore starts in familiar territory: clean data, clear ownership, reliable permissions, explicit business rules and systems that can expose those capabilities consistently across more than one interface.
