Agentic Commerce:a mandate for your agent, a till for the machines

AI agents search, compare and buy on behalf of people or companies, and they settle the purchase themselves. Nobody taps “add to cart”. For a small business that means two roles: your agent buys for you, and other people’s agents buy from you.

A pattern that creeps in easily: your agent is already spending money, just without a mandate. The harmless case is not even agentic commerce. A subscription runs into its limit, and per the contract the platform tops it up by another ten euros. No AI decides that, a clause does. It gets interesting one step further up, when the agent itself triggers something that costs money.

I know this from my own. They hang off services through MCP servers, services where I have loaded credit up front: an image generator, a search index, compute from several providers. When I have my agent produce an image, it commissions that in my name and burns credit I bought earlier with real money. It works well, right up until a line turns up on the statement that nobody can place: a cryptic vendor name, no receipt in the folder, a mandate nobody ever wrote down. When I first went through a statement like that, I did not spend a second thinking about currencies. I set a limit and wrote down who is allowed to do what.

For a sense of where the German Mittelstand stands on this: 41 percent of companies with 20 or more employees use AI, up from 17 the year before. Businesses under 20 people do not appear in the survey at all, so your share is probably lower. AI agents are among the three fastest-growing fields. And then the same study has the finding I like best: 55 percent consider blockchain economically important, but only 4 percent use it (Source: Bitkom Research, 2026).

They consider it important and do not touch it, which is entirely fine, because none of this needs a coin of your own. What it needs is a mandate. It is only that the till for machines is being assembled in exactly the corner those 96 percent are not looking at.

In April 2026 the International Monetary Fund published a paper on agentic payments. The core mechanism in it is not a coin. It is cryptographically verifiable mandates: scope, limits, the identity of the actor, permitted conditions. So that the final authorisation rests on the explicit consent of the human rather than on whatever the model inferred (Source: IMF, 2026).

Those are the four fields on the form at the top of this article, and what the IMF is describing there is a power of attorney. It stays a tool, and it delivers exactly what this whole debate comes down to: traceability and trust.

Trust here means: tomorrow morning you can say who bought what, with whose permission, under which limit. You can stop it. And if something goes wrong, a person sits at the table, not a model.

That is what the EU AI Act asks for at its core, even when your purchasing agent is often not a high-risk system: human oversight, a log, a way to intervene. The four fields are that, reduced to a sheet of paper. I described that as a trust advantage in “Why European Data Protection Is an AI Advantage”. Here it is daily work: a ledger and a revoke button.

The zoo of protocols emerging right now sorts into three layers. At the top, the mandate: who authorised whom to do what, how much, until when? That is where, for example, a standard sits that Google released in September 2025 with more than 60 partners, from Mastercard to PayPal, and handed over to an independent body in April 2026 (Source: Agenticplug, 2026). In the middle, the checkout: cart, address and order between agent and merchant. That is the layer OpenAI and Stripe are building. At the bottom, the till: card or digital money. For consumers, almost all of it runs over cards today. For machines paying each other in cents, almost all of it runs over stablecoins. Of the five most prominent deployments in early 2026, three settled on cards and two on stablecoins, almost exactly along the line between human and machine (Source: Forbes, 2026).

The card networks know this. Visa sits in both camps, in the foundation behind the machine protocol and in the mandate standard. In March 2026 Mastercard agreed to buy a stablecoin platform for up to 1.8 billion dollars (Source: Forbes, 2026). The networks want to stand at every till, whichever one it turns out to be (Source: Forbes, 2026).

The till is real, and so are the cent amounts. They explain why digital money shows up in this debate. They are not the thesis. Deciding who may do what in your name is something you have been doing for years. No protocol writes a mandate. Someone who knows the business writes it.

The role I described in “Mini-CEO of your agents” simply gains a budget column.

In September 2025 OpenAI launched purchasing directly inside the chat. More than a million merchants were announced. By February 2026 around 30 were live, and after five months the feature was gone. A Gartner analyst: OpenAI had underestimated how hard enabling transactions was going to be (Source: Digital Applied, 2026). No tax handling, no real-time inventory. Customers researched in the chat, then bought in the shop they already knew.

What came out of it has a name in the industry: discover in the AI, buy on your site. You do not have to build a new channel. You have to make your existing one readable for machines. Agentic commerce in B2B starts with the catalogue, not with the chatbot.

Three things.

First: readable. Prices, availability, delivery time and terms as data, not as a PDF and not as text inside an image. Gartner says purchasing agents prioritise vendors with clear, machine-readable terms (Source: Gartner via Two, 2026). That applies to the trade business with spare parts just as much as to the consultant with hourly rates.

Second: a door for agents. Not every machine is a customer. Many are scrapers. Visa cites a 4,700 percent rise in AI-driven traffic to retail sites as the reason it built a protocol with Cloudflare that lets merchants tell legitimate agents from bots (Source: Forbes, 2026). Agent identity is a doorman question, not a ban question. And this is already surprisingly workable with parts that exist today: you expose your service as an interface, an MCP server for instance, and issue one access token per customer. Access and billing both run through that token. Whoever holds one is a customer, everything else is traffic. Agents are already knocking at your doors. I showed that in “When agents hack by accident”. The next step is putting a till at the right door.

Third: a till for small amounts. If you sell something agents need in small bites, a data query, a check, a document, you need billing where the fee is not larger than the price. For that case there is no off-the-shelf answer in Europe yet. If you want to bill in small amounts, you still have to assemble the till yourself for now.

An example, entirely made up. A plumbing operation with seven people has kept its spare parts catalogue as a PDF on its website for years. The purchasing agent of a property manager cannot find it because it cannot read it. The rival two towns over has the same parts as data with price and delivery time. The agent orders there. Not because it is cheaper. Because it is readable.

Machines buy compute from machines, a thousand times a day, for cent amounts, in a currency you can ignore. Those are not people buying sweaters. Those are agents paying for compute, data and API requests (Source: Forbes, 2026). At the beginning of every one of those chains hangs a piece of paper that a human signed. Only you can write it. Only you know what “too much” means in your business.

The agent: search, compare, pay, log, at three in the morning. The human: scope, limit, threshold, term. And the conversation with the customer that no agent should have.

Tonight, take a sheet of paper. Write “Mandate” at the top. Four fields. Fill in what your agent is already buying in your name. When you get stuck, you know where to start.

What I am watching over the next twelve months: whether the mandate standards arrive in ordinary accounting and shop systems, whether a euro till for machines emerges, whether the first disputes appear in which an agent bought the wrong thing and nobody knows who is liable. The EU AI Act will not ask about the currency then. It will ask whether a human was in the loop and whether that is on record.

All names of individuals and companies used in this article are fictitious. Any resemblance to real persons or businesses is purely coincidental and unintentional. The examples are provided solely for illustrative purposes.