An AI agent that compares suppliers, requests quotes, negotiates a price and places an order without a human clicking: that is no longer a lab demo. Gartner estimates that by 2028, AI agents could intermediate more than 15 trillion dollars in B2B purchases, and that 90% of B2B buying would go through them. For a Belgian company that sells or buys in B2B, the question is no longer whether this is coming, but how fast, and how to stay in the loop.
B2B commerce ran over the phone, by email, then on portals and marketplaces. The next layer is the agent: autonomous software that pursues a buying goal, makes decisions and acts, rather than just answering. We talk about agentic commerce when an agent can search, compare, negotiate and buy on behalf of a company, within limits set in advance.
The underlying market is huge. According to Mordor Intelligence, global B2B e-commerce is worth 36.86 trillion dollars in 2026 and should reach 61.66 trillion by 2031. That volume is what sharpens the interest in automation: every point of friction removed from a buying process quickly adds up to millions. Two maturity levels need to be distinguished, and they coexist today.
The agent prepares, recommends and fills the cart. A human buyer validates every order. This is the dominant level in 2026, and the easiest to govern.
The agent decides and executes on its own within defined limits: budget, approved suppliers, thresholds. The human only supervises the exceptions and sensitive cases.
The advantage of B2B is that many purchases are repetitive, constrained by framework contracts and governed by clear rules. That is exactly the ground where an agent adds value, because the decision is bounded and measurable.
The agent compares suppliers, references and catalog prices, and shortlists according to your buying rules.
continuous shortlistIt requests quotes, compares terms and negotiates price and lead times within limits set in advance.
machine to machineThrough an agentic payment protocol, the agent places the order and pays within a traceable, auditable frame.
bounded executionStock monitoring, triggering of recurring reorders and respect of negotiated framework contracts.
fewer stockoutsFor an agent to buy without human intervention, a trust layer is needed: proving the agent really acts on the buyer's behalf, authorizing a payment, keeping a record. That layer took shape very fast in the autumn of 2025 around competing open standards, which merchants will sooner or later have to handle.
The Agent Payments Protocol, announced on 16 September 2025 with more than 60 partners including Mastercard, PayPal, American Express and Salesforce. An interoperable, auditable agentic payment framework.
The Agentic Commerce Protocol, first release on 29 September 2025, already live in ChatGPT. An open standard linking buyers, agents and merchants to complete a purchase.
In parallel, Visa (Intelligent Commerce) and Mastercard (Agent Pay) launched their own frameworks in 2025 to govern payments triggered by an agent. The practical consequence is simple: a serious B2B merchant will have to make its catalog readable by agents and accept several protocols, just as it had to accept several payment methods twenty years ago.
Forecasts vary by firm, but the direction is the same. McKinsey puts the global agentic commerce opportunity at 3 to 5 trillion dollars by 2030. Bain estimates that agentic AI could account for 15 to 25% of US e-commerce by the same horizon, that is 300 to 500 billion dollars in sales. On the B2B side specifically, Gartner's forecast is the most striking: 90% of B2B purchases handled by agents and more than 15 trillion dollars of intermediated spend by 2028.
These figures are projections, not certainties. What they mostly say is that the players who matter (payment networks, software vendors, marketplaces) are already betting on this scenario and building the infrastructure. Waiting for everything to settle amounts to letting your competitors learn in your place.
Agentic protocols rest on verifiable data and trust frameworks: agent authentication, an explicit buyer mandate, an audit log. That layer is what lets an agent negotiate, contract and pay at high frequency with minimal human intervention. Without it, no serious autonomous buying, and it is what a company should demand from its tools before opening the door to agents.
Agentic buying inherits precise risks. Fraud: Visa warned in late 2025 of rapidly escalating risks tied to agentic commerce. Agent washing: Gartner estimates that of the thousands of vendors claiming agentic capability, only about 130 offer the real thing. Uncertain value: over 40% of agentic AI projects will be canceled by the end of 2027, for lack of clear return or risk control. Liability: who answers for a wrong purchase made by an agent? The rule stays the same: validate on your own data, cap the spend, keep a human on the trade-offs.
Your customers will delegate part of their buying to agents. A poorly structured catalog becomes invisible to them. Making product data machine-readable becomes a commercial issue, not just a technical one.
Automating sourcing, quotes and recurring reorders frees up time and cuts errors. Provided you set clear limits and keep your hand on high-stakes decisions.
A company's purchasing data is sensitive: suppliers, negotiated prices, volumes, margins. Entrusting those flows to an agent means controlling where the data goes and who sees what. A rollout that respects the GDPR and the EU AI Act, on infrastructure you control, remains simpler to govern than an opaque remote service. An agent that commits money must be traceable, capped and supervised, by design.
Structured product data, price, availability and terms cleanly exposed. An agent only buys what it understands.
Budgets, approved suppliers, validation thresholds. Decide what an agent may do on its own and what requires a human.
AP2, ACP, payment integration. Logging, agent authentication and verifiable spending caps.
Supervise sensitive cases, monitor fraud and drift, document for the audit and the regulator.
Molderez Consult helps Belgian companies prepare their catalogs and purchasing for agentic commerce: agent-readable data, scope of autonomy, guardrails and governance, on infrastructure they control.
Discuss my use case