Agentic commerce explained: are businesses ready for AI agents?
Agentic commerce: buying and arranging things online through an AI agent that can search, compare and take action on a user's behalf.
Drafting an email, writing a LinkedIn post, coming up with a recipe for whatever’s left in the fridge. For a lot of people, that has become perfectly ordinary in a very short time. The next step is AI that actually does things for you. Buying a train ticket, switching to a cheaper energy supplier, filing an insurance claim. That’s agentic commerce. But are businesses actually ready for it?
The short answer: not even close. Agentic commerce means buying and arranging things online through an AI agent, software that searches, compares, fills in forms and pays on your behalf. The Agentic Experience Index by In The Pocket, a September 2026 study of 245 large companies and public service providers in the Netherlands and Belgium, shows that most organisations aren’t set up for it yet. An AI agent can find just over three quarters of them and gets a complete answer to roughly half of its questions. But finishing a task on its own, such as placing an order, submitting an application or switching provider, works at fewer than one in ten.
I talked about this recently on BNR Nieuwsradio, the Dutch news and business radio station. Below I go through the key questions, some of them the ones the editors put to me.
What exactly is agentic commerce?
Agentic commerce means an AI agent doing business with a company on your behalf. The difference from a chatbot like ChatGPT is that a chatbot gives you advice, while an agent does the work. Think of it this way: the language model is the brain, and the agent is that brain with digital hands and feet. It goes to a website, looks things up, fills things in and closes the deal.
Say you tell your agent: my printer’s run out, order new cartridges that arrive before Friday and cost no more than €30. For an online shop, simply showing up in the answer isn’t enough. The agent also needs to see which cartridges fit your printer, whether they’re in stock and when they’ll be delivered. And it needs to be able to set up the order, so all you have to do is hit ‘confirm’.
The past month showed that this is coming faster than many companies expect. On 8 September, Meta launched Muse in the United States, a personal agent that can also do your shopping. On 29 September, OpenAI followed with Dots: agents that keep working around the clock, each from its own computer in the cloud. Dots isn’t yet available to private users in Europe, by the way.
What jobs can an AI agent do for you?
In principle, anything you’d now spend an evening hunched over your laptop for. In The Pocket had an agent carry out, sector by sector, the tasks people search for most. For online shops: checking a price, seeing whether something is in stock and adding a product to the basket. For energy suppliers: calculating a price and switching. For banks: working out a car loan or mortgage and opening an account. For insurers: filing a claim. And for public transport: looking up a departure time, planning a journey and buying a ticket.
Those tasks fall into two types. Looking things up (what does this cost, when does the train leave) and getting things done (order it, switch provider, submit it). Looking things up goes reasonably well. Getting things done still goes very badly.
How well prepared are Dutch and Belgian companies for agentic commerce?
It varies hugely by sector. In The Pocket tested at three levels.
Can an agent find you? That works for just over three quarters of companies.
Can it read your information? Around half of the lookup questions got a complete answer.
Can it finish the job? Of the 380-plus action tasks, only 19 were completed from start to finish, about five per cent.
HR and payroll companies lead the pack. There, an agent could fully complete at least one job at 58 per cent of companies, such as requesting a quote or registering as self-employed. At banks, telecoms providers, government, health insurers and public transport, it didn’t succeed in a single test. At energy suppliers, it worked in one per cent of tests. And at online shops, an agent could set up an order at eight per cent of companies. Mind you, that meant putting something in the basket, not checking out.
Why do AI agents get stuck on websites?
There are two causes: technology and strategy.
Technology: many sites only load their prices and buttons after the page opens in your browser. To us humans, that looks fine. But an agent that doesn’t view the page through a browser and only reads the raw files gets an empty shell. This is the biggest technical culprit, and also the cheapest to fix.
Strategy: a deliberate choice. A login, a firewall that turns robots away, or an identity check. For a bank or insurer, that makes perfect sense: the company wants to know who its customer is.
The very latest generation of AI agents changes a lot. Agents such as OpenAI’s Dots and Meta’s Muse get their own virtual computer with a browser in the cloud. They simply click through a website the way you and I do, so they can already clear some of the technical hurdles. But faced with a login or DigiD, they too hit a closed door.
Does it matter if your company isn’t ready for AI agents?
In the short term you won’t lose much revenue over it, but things could move fast. A third of Dutch shoppers say they want to use AI agents to find the best price, and in Germany around six per cent of online shoppers would now let AI complete a purchase entirely. So we’re happy to outsource the comparing, but not yet the paying. Rightly so, because agents still make mistakes. OpenAI itself says you should check any important work Dots does.
Being found and read properly, however, already matters enormously. More and more people put their questions to ChatGPT rather than Google, and over the coming years agents will become an increasingly normal kind of personal assistant. Take two online shops selling the same product at the same price. Shop A has beautiful photos, but its information is hard for a machine to read. Shop B has clear product data. When your agent makes the comparison, shop A doesn’t even make the list.
How do you optimise a website for AI agents? Is this the new SEO?
Partly, but it goes a step further. SEO is about being found, which matters for agents too. Agentic commerce is also about transactions. That leaves two key questions: ‘Can AI find me?’ and ‘Can AI do business with me?’
What agents will prefer is hard to predict, because an agent adapts to the preferences of its ‘owner’ and of the tech company behind it. Does your agent choose on price? On sustainability? On reviews? On the companies it can most easily talk to technically? Or on commercial deals you know nothing about? For twenty years, companies have thought about how to win people’s attention. Now they also have to think about how to get selected by machines. And we as a society have to think about how transparent that selection should be.
Should companies actually keep AI agents out?
That’s mainly a commercial decision. Amazon blocked Meta’s Muse on 20 September, twelve days after launch. Anyone trying to shop on Amazon via Muse gets a message saying an unauthorised AI agent breaches the terms of use. Amazon says Muse doesn’t identify itself as an agent and that Meta never asked for permission. Meta says users have to approve every purchase first and that the agent can’t see passwords or payment details. But Amazon also has its own shopping assistant. So this is also about who gets to own the customer.
In The Pocket’s test agent gets stuck because it announces itself as a bot. Muse is kept out of Amazon precisely because it doesn’t. Either way, access is not guaranteed.
Shopify, on the other hand, has fully embraced Muse. There, Meta is an official sales channel and you can simply check out through Muse. That opens doors for the many small online shops running on Shopify.
For a giant like Amazon, locking the door is low risk, because people will come anyway. A mid-sized online shop or energy supplier doing the same is choosing to be invisible. At the same time, let’s be honest: this technology isn’t mature yet. All sorts of things will go wrong, such as agents buying things the customer never wanted. Those get sent back, and returns are expensive. So the most sensible route, in my view, is a door with a doorman: you let agents in, but through a controlled entrance and with clear agreements about what they’re allowed to do.
What can companies do right now?
Three things:
One: test it yourself. Have a few different agents carry out your company’s most important customer tasks. Not just looking things up, but actually ordering, booking or switching. Within an afternoon, you’ll see where they get stuck.
Two: fix the accidental blockers. Make sure prices, stock levels and terms are readable by a machine too. A human sees a page with twenty insurance policies and colourful charts. An agent just wants to read: premium this much, excess this much, these terms, these are the steps. That often comes down to how your website is built.
Three: build a dedicated entrance for agents and decide what they’re allowed to do. Standards for this are now emerging. OpenAI developed the Agentic Commerce Protocol with Stripe, and in January Google and Shopify unveiled the Universal Commerce Protocol. With these, an agent can see exactly what something costs and how to order it. This does raise new design questions. What can an agent do on your customer’s behalf? How do you know the agent really has that customer’s authorisation? And what happens when it makes a mistake? An agent might be allowed to book a table, but not to change a bank account number on a whim. For now, the most likely model is: the agent prepares, the customer hits ‘confirm’.
Want to hear the conversation in Dutch? Listen on BNR Nieuwsradio.
