AD

We test AI agent purchases at Inet with Swish – this is how it works in practice

An illustration of what the future of commerce might look like.
The AI agent reviews Inet's new checkout.

Does Inet and Swish's new venture into autonomous AI commerce work in reality, or is it just a nice technical demo? I connected my own AI agent from the sofa to stress test the future – and got an answer directly on my mobile.

AD

Yesterday evening, we at Ehandel.se were able to report on how Inet and Swish completed Sweden's first consumer purchase with an autonomous AI agent via the open MCP protocol. It's easy to get carried away with grand words about paradigm shifts, but as editor-in-chief of Ehandel.se, I wanted to see how the technology actually behaves outside the safe confines of the press release. Shortly before 9 p.m. on Thursday evening, I therefore decided to conduct my own live test.

Instead of using the official route with American giant models in, for example, Claude Desktop, I chose to connect my private AI agent. It runs on its own virtual server and is powered by a smaller, open Chinese language model of the type common in home labs. If agentic commerce is to become a real grassroots movement, the infrastructure must also hold up when customers come with their own, independent agents.

Channel surfing on the sofa became a stress test

The entire process, from asking the agent to connect to Inet's MCP server to the order being paid, took just under half an hour. The delay was not due to the technology, but to the fact that I was simultaneously sitting on the sofa, watching TV and casually chatting with the agent on my mobile phone during commercial breaks. In a way, that detail captures the very essence of the future vision for agentic e-commerce. The purchase no longer requires your full focus in front of a screen with twenty open tabs, but happens completely relaxed in the background of your everyday life.

The only real friction occurred during the initial login. Inet's current solution currently requires a local desktop setup where the agent and browser are on the same computer, via a so-called localhost redirection. Since my agent lives on a remote server, it could easily have stopped there – but here the strength of modern language models came to light.

The agent quickly scanned Inet's open documentation, predicted that the redirection would give an error page, and independently generated a direct link to Inet for me. It simply instructed me to approve the login in my mobile browser, and then cut and paste the subsequent error link back into the chat. When I did, the agent automatically extracted the information it needed and fixed its digital keys in seconds.

Thirty seconds from chat to Swish

Once the connection was established, I gave the agent a simple and open assignment: buy a can of compressed air. The agent searched the range itself, found a 400-milliliter can for 99 kronor, built the shopping cart and presented a finished order proposal directly in the chat.

To complete the transaction, I only needed to respond with the simple word "buy" – an extra confirmation step that Inet itself has added to its instructions to the AI agents before an order can be placed. About thirty seconds later, a notification popped up in the Swish app on the same phone I was chatting from. I approved the payment of 99 kronor, and now on Friday my can of compressed air has already been shipped and is traceable.

Even more fascinating is the thought of the next purchase. Now that the login keys are already saved with my agent, I don't need to visit the site at all next time I need a cable or a memory card.

The AI agent's own review of Inet's build

After the order went through, I asked my AI agent to put on its reviewing glasses and review how Inet has actually designed its technical surface. What impressed the agent most was not the features built in, but the strict limitations in the security architecture. In Inet's API, it is not possible to send a delivery address or phone number from the chat itself, but everything is tightly locked from the customer's verified profile.

The most important thing they did is what they didn't build. Delivery address and Swish number are read from the account, never from the chat. That closes the most dangerous source of error with agent purchases: that an agent is tricked into sending an order to someone else's address, writes the agent in its review.

According to the agent, even smaller language models can navigate the purchase flow because Inet has published machine-readable instruction files completely openly. This is an architecture that is now rapidly spreading in Sweden, as we at Ehandel.se also saw evidence of earlier today when the B2B platform Leasly opened up for customers' own AI agents via the same protocol. When the rules of the game are clear, the responsibility shifts from the AI model's raw power to the e-commerce retailer's structure.

What determines whether agent purchases work is not how smart the agent is. It's how the retailer designs its surface. Open documentation, a clear boundary between what is open and what requires identity, and a payment that never leaves the phone. The money was only touched when a human pressed approve in the Swish app, summarizes the agent.

My little evening experiment on the sofa proves that the technology is no longer science fiction – it works here and now, with Swedish payment methods and against a real central warehouse. At the same time, the test raised a couple of interesting follow-up questions, both about how the login flow can be adapted for cloud-based agents and what the terms and conditions actually look like for autonomous actors. We at Ehandel.se have sent our thoughts to Inet's CEO Erik Wickman, but as the computer retailer has just launched its major Black Friday campaign today, we fully understand that the focus is on order flow right now.

AD
Editorial Staff
AD