For as long as any of us have been in business, whoever decided what to buy was a person, and so was whoever signed it off. Every trick in the marketing book, from the shop window to the retargeting advert, assumes a buyer with a mood, a memory and a few seconds of attention to spare. That assumption is starting to wear thin. The payment networks are building the plumbing that lets software spend money on our behalf. When the buyer is an agent carrying a set of instructions, it starts by checking whether your terms fit, and a business it can’t check is never in the running.

In March 2026, Santander and Mastercard completed what they called Europe’s first live end-to-end payment executed by an AI agent, running through Santander’s real payments infrastructure on Mastercard Agent Pay. It was a single transaction in a controlled environment, and both companies were careful to say it wasn’t a commercial rollout. What caught my eye was the description of how it works: the agent initiates and completes payments “within predefined limits and permissions” that the customer has set in advance. That phrase describes a customer unlike any we’ve sold to before.

Most of us have already got used to AI as an adviser. We ask it for three options on a new office chair or a supplier for printer toner, read the list and click through to buy the one we like. A human still makes the final choice, and the website still gets its moment to persuade.

Agentic payments can take that moment away. The customer says something like “reorder the toner when we run low, from anyone reputable, under £60, delivered by Friday”, and the agent goes off and does it. Google’s Agent Payments Protocol, launched in September 2025 with more than 60 partners including Mastercard, American Express and PayPal, formalises exactly this arrangement: the customer’s instruction is recorded and signed as what the protocol calls an intent mandate, and so is the basket the agent eventually settles on, with its items, price and terms locked in. Mastercard’s own scheme works along similar lines, allowing only registered agents to transact and only within what the customer has agreed to. For a seller, the useful part is the record. Under schemes like these, an order can arrive with evidence of what the customer actually authorised, though what reaches you is something to establish with your provider. The customer is still the person who gave the instruction, not the software carrying it.

Selling to the mandate

So the buyer that turns up at your checkout is, in effect, a written brief with a payment card attached. It has a price ceiling, a deadline and a short list of things it’s been told to care about, and it’ll compare you with everyone else on those terms first. Your photography and your countdown timer may still count for something, depending on how the agent has been built, but they won’t get you past a brief you don’t fit.

This is where the job changes for most small businesses. If the agent is checking you against a mandate, then the work of your website and product data is to make that check easy and to pass it honestly. The questions an agent asks are the dull ones customers have always asked, and plenty of businesses still make them surprisingly hard to answer: what exactly is this, what does it cost all-in, is it in stock, when will it arrive and what happens if it’s wrong?

Search felt much like this in its early days, which I watched from inside an agency. Google’s crawler didn’t care about anyone’s brand campaign either. The businesses that did well were often the ones whose pages a machine could make sense of — a good number of them small firms that simply described what they sold more clearly than the big names did.

That’s why I think agentic payments could do small firms a favour, though not the favour of a level playing field. Agents may well be steered towards approved merchants, integrated suppliers or whichever platform they were built on, and Walmart’s ChatGPT checkout converting three times worse than its own website shows that even a giant can find agent-led selling a poor channel. What clean, checkable terms do is remove a disadvantage you may not know you have. “Call for a quote”, prices hidden behind a login, delivery charges that appear at the last step and returns policies written to put people off are all likely to fail an agent’s check, and nobody will ring to say so. You won’t even see the lost sale, since the agent never mentioned it had been there.

Who pays when the agent gets it wrong?

The trust rules are still being written, and I’m not sure anyone yet knows how disputes will settle. If an agent buys the wrong size, somebody has to decide whether it overstepped its instructions or the customer wrote careless ones, and the public effort so far is going into identification rather than blame. Mastercard’s announcement at the end of September 2026 is about telling agent-led transactions apart from human ones and helping banks judge whether they look normal; its partner Skyfire put the aim as making every transacting agent “identifiable, accountable and auditable”. The release had nothing to say about chargebacks or liability. For a small business that leaves three questions to put to your payment provider before you accept an autonomous order: will it tell you when a payment was initiated by an agent, what record of the customer’s authorisation comes with it, and who carries the loss if the customer disputes it?

My guess is that the signed instruction will end up doing much of the evidential work, since a recorded brief and a locked basket make it far easier to see who got it wrong. Seeing who got it wrong and deciding who pays are different jobs, though. The record is evidence; the scheme rules and your provider’s terms decide what that evidence is worth. Either way, sellers who describe their products accurately will have the easier conversation.

What to do about it now

None of this needs a project plan this quarter. Mastercard cites a projection that around one in ten consumers will routinely use agents to make purchases by 2030, so there’s time. Most of what makes you easy for an agent to deal with makes you easy for a person too, and only the last item needs a conversation with your payment provider:

  1. Put the full price where a machine can read it. If delivery, quantity or VAT change the total, state the rule on the page rather than revealing it at the final step. An agent working to a ceiling needs to work out the all-in figure before it asks for authorisation, and a total that only appears at checkout reads as a seller that can’t be held to a number.
  2. Make stock and delivery dates real. A lead time that depends on destination or order size is fine as long as the dependency is stated. An agent with a Friday deadline is likely to pass over “usually ships in 3–5 days” for a page that commits to a date.
  3. Check the product data you already publish. If you run a Google Shopping feed or carry product markup, make sure the names, variants, identifiers, prices and availability in it match the page. Nobody needs to commission an API for this.
  4. Write returns and terms a stranger could follow. If disputes end up being settled against recorded instructions, clear terms protect you as much as they protect the customer.
  5. Decide whether you want agent orders at all, and ask your provider those three questions first. I argued in August 2026 that letting agents onto your site is a product decision, and that holds even more once they can pay. Make the call on purpose, then make your checkout match it.

Shops spent the best part of a century learning how to catch the eye of someone walking past. Sooner or later a toner order will land that nobody looked at your website to place. Whether it comes to you or to a competitor will turn on a price, a date and a returns policy, and those you can tidy up this afternoon.