In short: Agents acting for buyers make the B2B customer indirect. The order still reaches you, while intent, comparison and context stay with the agent. Gartner expects 90% of B2B buying to be agent intermediated by 2028. Businesses where sales and support absorb commercial complexity are the least ready, and the response starts with process reengineering, a mandate model and clear owners.

Who is your customer when an agent places the order?

This series has so far looked inward, at agents producing change faster than an engineering organisation can absorb it. The same shift is now arriving from the other side of the boundary. Customers are starting to send agents of their own, to search a supplier's catalogue, compare its terms and place orders on their behalf.

Gartner has put a number on the direction of travel. Among its strategic predictions for 2026 and beyond, it expects that by 2028, 90% of B2B buying will be AI agent intermediated, pushing over $15 trillion of B2B spend through AI agent exchanges. Investors studying the market have suggested the timeline may prove aggressive, and the exact figure probably matters less than the word Gartner chose: intermediated.

An intermediated customer is an indirect one. Something now stands between a business and the person who decides, built by someone else, working for someone else, and keeping what it learns. This is the Absorption Gap reaching the front door: demand arriving in a form, at a speed and through a channel that most businesses were never designed to absorb.

What do you keep when an agent does the buying?

You keep the order, and possibly more of them, since an agent may compare and reorder more often than a busy buyer would. What you risk losing is the signal around the order. The comparison with competitors, the hesitation over a price, the alternative considered and dropped, the moment an account manager would have spotted a churn risk: all of it now happens inside the agent, on a platform you do not control.

The loss is easy to underestimate because nothing breaks. Revenue can hold for a while as the relationship thins. The customer's preferences, constraints and history move into the agent's memory, which may hold the same knowledge about your competitors, and a supplier the agent knows only as a catalogue, a price and a delivery promise becomes the easiest one to replace.

The mechanism has a direct consequence for architecture. Whatever a business exposes to agents becomes most of what they know about it, so the quality, completeness and governance of that surface decide how it is compared. Inside the engineering organisation, the series called this context capacity. Facing outwards, the same discipline shapes the information a buyer's agent uses to decide where to buy.

Why does B2B feel this harder than consumer commerce?

Consumer commerce has drawn most of the attention, and it has already shown how hard the problem is. OpenAI launched Instant Checkout in ChatGPT in September 2025 and stepped back from it in March 2026, saying the first version lacked the flexibility it wanted and letting merchants keep their own checkout. Reporting at the time pointed to difficulties onboarding merchants, keeping product data accurate, supporting carts with several items and connecting loyalty memberships.

B2B carries that complexity and adds its own. A business customer is rarely one person with one card. It is an organisation with entities, sites and cost centres, negotiated prices and contracts, billing rules that differ by entity, budgets, and approvers who must sign off above agreed thresholds, and a buyer's agent will have to carry all of it.

B2B is also not starting from nothing. Large accounts have bought machine to machine for decades through EDI and punchout catalogues wired into their procurement systems. Those integrations were rigid and costly, so they were reserved for the largest customers while the long tail went through sales, phone, email and the web shop. Agents could bring the machine to machine pattern to every account, which is both the opportunity and the exposure.

The architecture then depends on who sends the agent, and three personas seem to matter most.

The direct buyer sends its procurement agent to source, order and track. That agent needs the catalogue, the customer's contract prices, availability and the right to order within the customer's approval rules. How much of the relationship the supplier keeps depends on whether the agent orders through the supplier's own surface or through a marketplace in between.

The reseller buys to sell on, and its agent sources across several suppliers to fill its own customers' orders. The end customer is two steps away, so the signal the supplier keeps is thinner still, and the agent will judge it largely on data quality, availability and delivery reliability.

The partner embeds a supplier's capabilities through APIs inside its own platform, so that its users order without ever seeing the supplier. It is the most intermediated persona of the three, and it calls for contracts, service levels and usage controls designed for software as the customer.

Each persona implies a different surface, a different trust model and a different share of the relationship. Treating them as a single "agent channel" would probably expose too much to one and too little to another.

Was your sales team the API all along?

In most B2B businesses, a commerce function absorbs the complexity of the customer's organisation: its hierarchy of entities and sites, its billing rules, its approvers, the special conditions agreed over years. Self service aims to let customers act alone within those boundaries. Getting there is harder than it looks, because it takes a very high bar of architecture, product, design and engineering for every process between the core product and commerce to run without a seam.

Where that bar is not met, people fill the seams. Sales representatives key in orders that arrived by email, customer service resolves the price the system could not, technical support corrects the master data a customer tripped over, and the account manager knows which approver to chase. In B2B organisations I have advised, this human layer is often the real integration between product and commerce, even where the technology foundations themselves are sound. It works at human volume because people quietly compensate for what the processes leave out.

Buyers were already moving away from that layer before agents arrived. In a Gartner survey of 646 B2B buyers conducted in August and September 2025, 67% said they prefer to buy without a sales representative, and 45% reported using AI during a recent purchase. An agent takes that preference to its limit: it cannot call the account manager, and it will not wait for customer service to fix a price.

Figure 1. In many B2B businesses, people are the integration layer between commerce and the core product. A buyer's agent cannot use that layer, so the processes behind it have to be reengineered before they are exposed.

Figure 1. In many B2B businesses, people are the integration layer between commerce and the core product. A buyer's agent cannot use that layer, so the processes behind it have to be reengineered before they are exposed.

Exposing today's processes to agents as they stand would mostly automate the queue in front of those people. The reengineering comes first: order, pricing and approval processes that run end to end without human compensation, with people kept for the exceptions and the relationship. It is a business transformation with an architecture inside it, and it rarely fits inside a technology roadmap alone.

What is a buyer's agent allowed to do?

Every transaction an agent makes for someone raises the same questions. Who is acting, and on whose behalf? Within which limits, agreed by whom? Can the business prove it afterwards, and can the customer revoke it? Together the answers form a mandate, and the mandate seems likely to become the central object of agent commerce.

Payment networks have started to formalise it. Google's Agent Payments Protocol, announced in September 2025 with more than 60 organisations, records the user's instructions as cryptographically signed mandates: an intent mandate capturing what the user wants, and a cart mandate fixing the exact items and price. Google describes the chain from intent to cart to payment as a non repudiable audit trail for authorisation and authenticity.

A B2B mandate needs more. It has to carry the buying organisation's own rules: which entity is buying, which contract applies, which budget it draws on, and which purchases still need a human approver. Designing Intent Capacity showed that a spec bundles intent, context and policies with different owners, and a buyer's agent arrives carrying exactly that kind of spec, written by someone else.

Figure 2. Your own agents are governed inside your control plane. Your customers' agents can only be verified at the boundary, against the mandate they carry and the policies you enforce.

Figure 2. Your own agents are governed inside your control plane. Your customers' agents can only be verified at the boundary, against the mandate they carry and the policies you enforce.

The asymmetry matters for architecture. A company's own agents can be governed from the inside, through the control plane described in Governance as Architecture, where control capacity measures how much agent action an organisation can authorise, constrain, trace and reverse within its risk appetite. A customer's agent sits outside that plane, and the supplier can only verify it and constrain it at the boundary, through the mandate it presents and the policies enforced against it.

What should the agent facing surface expose?

If the mandate defines what an agent may do, the agent facing surface defines what it can do. In practice that means the capabilities a buyer needs, exposed as APIs and increasingly through MCP servers that agents can discover and call: catalogue and availability, contract pricing, order creation within approval rules, order status, returns and invoices. Policies matter as much as data. Delivery promises, return conditions and payment terms have to be machine readable, because an agent cannot pick them up in a conversation with sales.

Estates built around APIs and events start from a better position. At La Redoute, moving to APIs and events helped reduce order to shipment from days to about two hours. The aim then was speed for human customers, and the same property, capabilities exposed as services instead of being locked inside screens, is what an agent needs now.

Bot management changes meaning at this surface. For years the goal was to keep automated traffic out. When Visa announced its Trusted Agent Protocol with Cloudflare in October 2025, it noted that existing bot detection can mistakenly block legitimate purchases made by agents, and cited a rise of more than 4,700% in AI driven traffic to US retail sites over a year, based on Adobe data. Its protocol and Mastercard's Agent Pay both build on Web Bot Auth, which lets an agent sign its requests so that a merchant can tell a registered agent from an anonymous bot.

Who owns the agent channel?

In most organisations, nobody yet. Security sees agent traffic as a threat to block, marketing counts it as a referral source, sales owns the accounts the agents act for, and product may not see it at all. Redesigning processes for agents falls between these roles, and so does deciding which agents to trust and what each may do.

A split along these lines might work, adjusted to each organisation. Product would own the agent channel as a channel, with its own personas, journeys and measures. Technology would own the surface and the trust model, security the verification policy, and sales with customer operations the redesign of the processes their people currently hold together. The commercial terms an agent can accept, from price to credit, would stay with whoever owns them today, now written so that a machine can read them.

The CTO's part is to bring business and product into the work early. Engineering can build the surface. Whether the surface is worth exposing depends on the reengineering behind it, and that belongs to the whole business.

Should you wait for the standards to settle?

There is a reasonable case for waiting. The protocol landscape is crowded: OpenAI and Stripe's Agentic Commerce Protocol, Google and Shopify's Universal Commerce Protocol, Google's Agent Payments Protocol, Visa's Trusted Agent Protocol and Mastercard's Agent Pay, with Web Bot Auth underneath and MCP and A2A for agent connectivity. One industry roundtable on agentic commerce counts at least six protocols competing to govern agent identity, authorisation and payment, and the OpenAI retreat shows that even a well funded first move can be withdrawn within months.

For the protocol choice itself, waiting is defensible. For everything underneath it, waiting looks expensive. Process reengineering, a capability surface, machine readable policies and a mandate model are useful whichever protocol wins, and they take far longer to build than a protocol adapter. The same reasoning runs through Not Yet: the investment worth making early is the one that pays off across several futures.

Taking part may also be worth more than watching. Most of these protocols were written by platforms, payment networks and large retailers, around consumer checkout. The B2B questions, such as organisational mandates, approval chains, contract pricing and liability when an agent exceeds its mandate, remain largely open. Suppliers that bring their constraints to standards bodies and industry groups now may find the result fits them, while those that wait may inherit a model designed around someone else's checkout.

The question for your agent channel

The customer is becoming indirect. A business that understood its customers through its people will increasingly have to understand them through what it exposes and what it can verify, and that work starts in the processes those people hold together today.

If your best customer's agent placed an order tomorrow, would any of your systems know who it was acting for, and would anyone in your organisation own the answer?

Read further

FAQ

What is an intermediated B2B customer?

An intermediated B2B customer is a business that buys through an AI agent acting on its behalf, so the supplier deals with the agent rather than with the people who decide. The order still reaches the supplier, while the intent, comparisons and history behind it stay with the agent and its platform. Gartner expects 90% of B2B buying to be agent intermediated by 2028.

How is a customer's agent different from my own AI agents?

A company's own agents can be governed from the inside, through a control plane that authorises, constrains, traces and reverses what they do. A customer's agent belongs to someone else and sits outside that plane. The supplier can only verify it at the boundary, by checking its identity and the mandate it carries against the supplier's own policies.

What is a mandate in agentic commerce?

A mandate is the verifiable statement of who an agent acts for, what it is allowed to do and within which limits. Google's Agent Payments Protocol records it as signed intent and cart mandates that form an audit trail. In B2B, a mandate also needs the buying organisation's rules: the entity, the contract, the budget and the purchases that still require a human approver.

Why are B2B businesses less ready for buyer agents than consumer retailers?

Many B2B businesses rely on sales, customer service and technical support to absorb the complexity of their customers' organisations: entity hierarchies, billing rules, approvers and negotiated prices. That human layer works at human volume but cannot be used by an agent. The processes behind it usually need reengineering before they can be exposed to agents safely.

Should we wait for agentic commerce standards to settle before investing?

Waiting on the choice of protocol is defensible, since at least six compete for agent identity, authorisation and payment. Waiting on the foundations looks costly: reengineered processes, a capability surface, machine readable policies and a mandate model pay off whichever protocol wins. B2B suppliers may also benefit from taking part in standards work, because most current protocols were designed around consumer checkout.

Sources