AI · AI in everyday life
When the AI Agent Shops for You – What Actually Arrives in Europe
Agents that place orders themselves are reality in the US and not yet legally settled in Europe. What already happens today, where the liability question is open, and what to watch when AI recommends products to you. As of August 2026.
By Boaz Lichtenstein Prefer us on Google
The story has been the same for two years: soon an agent will handle the shopping. It knows your preferences, compares prices, orders and tracks delivery. For consumers in Germany the position in August 2026 is more sober and more interesting at once. The first part has long since arrived – AI assistants recommend, compare and sort products, and they do it for millions of people daily. The second part, the actual purchase, has not arrived here yet, and the reason is not technical. This piece separates what is already happening from what is still being claimed.
Key takeaways
- Recommending: yes. Buying: not here yet. Purchase inside the chat is limited to the US as of August 2026.
- The brake is the liability question, not the technology: who pays for an order the agent misunderstood?
- Two protocols are sorting the market: the Agentic Commerce Protocol from OpenAI and Stripe, and the Universal Commerce Protocol from Google and Shopify – the latter now with Amazon, Meta, Microsoft, Salesforce and Stripe among the participants.
- A recommendation is not a verdict. Rankings emerge from data sources and partnerships nobody outside audits.
- Your right of withdrawal is unchanged – it attaches to the merchant, not to the route to the contract.
- The data price is higher than for search: occasion, budget, address, payment details in one flow.
- For merchants the recommendation phase is what counts now, not the missing checkout.
What already happens today
The part that gets talked about less is the more effective one. When somebody types “which running shoe for wide feet” into an assistant, a shortlist of three to five products appears – and for many people that shortlist replaces the comparison they would previously have done across several websites. It shifts influence away from results lists and review portals towards the data sources a model builds its recommendation from.
Infrastructure has formed on the supply side. OpenAI and Stripe have put forward the Agentic Commerce Protocol, an open standard for how agents and merchants communicate; Google and Shopify are driving the Universal Commerce Protocol, which in spring 2026 was joined by Amazon, Meta, Microsoft, Salesforce and Stripe among others. That houses this different agree on shared interfaces is the real signal: a buying surface is being built here, not a feature.
How merchants prepare is covered in the agentic commerce guide and more concretely in agent-ready commerce.
Why the purchase is still missing here
The short answer: because nobody can say who is liable. The longer one is worth having, because it shows this is not a translation problem.
If an agent orders something you did not want – wrong size, wrong variant, twice over, or a product it inferred from a misread phrase – several bodies of law apply simultaneously in the EU: consumer law for the contract, payment services law for the authorisation, data protection for the profile, the AI Act for the system. None of them clearly answers whether the model provider, the merchant, the platform or you carry the risk. For providers that means launching in the EU without settled liability is an experiment with other people’s money. Accordingly, every rollout begins in the US.
That gap will close, though not through a single piece of legislation – and until then an uncomfortably honest formula applies here: the convenient half is missing, the influential half has already arrived.
What to watch – now
Treat the recommendation as a shortlist, not a verdict. An assistant names products because they look good in its sources, not because they were tested. For anything above pocket money, a second independent source belongs in the process – along with the question of who paid for the first one.
Remember who you are buying from. When you move from the chat into a store, the merchant is your counterpart: warranty, withdrawal and delivery sit with them. The assistant has nothing to do with it, even though it brokered the purchase. Since June 2026 that merchant has to provide a digital withdrawal function in Germany – a right that becomes practical precisely when a recommendation did not fit.
Be frugal with the profile. A shopping agent wants to know what you need, what you will pay and where it goes. That is a denser picture of you than any search history. What to consider is covered in AI and what you type into it.
Separate convenience from delegation. An agent that prepares and lets you decide saves time without risk. An agent that orders relocates decisions – and the question of when you want that is best answered before the function is available here. How to use assistants usefully otherwise is set out in AI assistants in everyday life.
Bottom line
In Germany the shopping agent is not yet a buyer but already an adviser – and as an adviser it influences more purchase decisions today than any ranking list. The most useful stance is therefore neither enthusiasm nor rejection but ownership: take the recommendation, check the selection, know your counterpart, keep the profile small. When the checkout does launch here, it changes the convenience – not the responsibility.