Agentic Commerce opens up new opportunities for merchants and manufacturers – from more efficient purchasing processes to more advanced personalization. At the same time, it raises new questions around trust, control, liability, and competition.
In the first part of our interview with Paul Krauss, we covered the basics of Agentic Commerce, the technological developments driving the trend, and the opportunities it creates for commerce. In this second part, the focus shifts to the challenges: What risks emerge when AI agents independently select and buy products? How should businesses prepare? And what regulatory guardrails does agentic commerce need?
1) Where are the greatest risks when autonomous systems select, compare, or even buy products – for example in terms of transparency, control, and liability?
The legal situation in the EU has not yet been conclusively clarified. At its core, this is a liability question, similar to what we know from robotics: if an autonomous system causes damage, who is responsible? The platform, the merchant, the model provider, or the user?
There are mechanisms to solve this. Shared payment tokens with limits are a starting point, because they prevent an agent from purchasing without control. Audit trails, defined escalation paths, insurance solutions – all of this will come, just as it did with online payments twenty years ago.
The real problem is therefore less the law and more trust. Trust is slower than technology. An agent that buys the wrong product is not a UX bug; it is a reputational problem – for the platform, for the merchant, and for the entire category. The first precedent cases will come soon, and they will slow adoption for one or two years. That is no reason to panic, but it is a reason to prepare.
There are also three structural risks that are underestimated. First, manipulation: prompt injection on product pages, fake structured data, Agent SEO as the next dark discipline. Second, concentration: 90 percent of coding agents choose Stripe as the payment default. If such default effects spill over into the consumer market, we will have digital quasi-monopolies that are not yet covered by antitrust law. Third, the lesson from IoT: ten years ago, the refrigerator was supposed to order milk. It failed not because of AI, but because of interoperability and liability. These same two issues will determine the future of Agentic Commerce.
2) How much decision-making freedom will consumers actually be willing to hand over to AI agents in the future?
This is highly category-specific, and every merchant and every consumer will find this out for themselves. The current ECC figures are revealing in this respect: 11 percent say “definitely,” while 50 percent say “for certain purchases.” That is an honest answer. Consumables yes, sneakers maybe, wedding suit no. Travel partly, insurance gladly, therapy never.
There is a significant adoption gap between the US and Germany. In the US, 30 to 45 percent of people already use generative AI for research. In Germany, the figure is significantly lower. This is not primarily a technical question, but a cultural one. It will converge, but more slowly than some roadmaps suggest.
On the merchant side, this is a prisoner’s dilemma, and Alex Graf put it sharply in his brand eins interview. His position: do not go to prison voluntarily. Anyone who connects to Google’s UCP or similar protocols gives up their data and their decision-making power, and in return gains nothing they would not have had anyway. Anyone who does not connect risks no longer being found by the agent. Both options are uncomfortable, and both are real decisions.
Graf is relatively alone in the industry with his skepticism, but his argument is not trivial: Agentic Commerce first needs to solve a problem that the customer actually has. With printer cartridges, yes. With sneakers, probably not.
This will not be decided by the industry, but by the consumer – and the consumer will decide category by category.
3) What requirements must companies meet to be ready for Agentic Commerce – technologically, organizationally, and in terms of data quality?
Data quality is eighty percent of the work. If a machine does not understand products in detail, they do not exist as a preference. Granular, up-to-date, complete product attributes are the entry ticket. Real-time inventory is mandatory, not optional. An agents.txt or llms.txt file and a grounding page that explains how the catalog is structured is a good start. It costs little and has an immediate effect. Feed optimization is an entirely different discipline, and anyone still writing texts manually today will have a problem.
Architecturally, this means API-first, headless, and ACP- and MCP-ready. Anyone building a proprietary checkout today is investing in a model that will no longer matter in the agentic economy. Agents do not interact with graphical interfaces. They need machine-readable data and structured APIs. Stripe Checkout, Shopify Checkout – the infrastructure is already there. Most companies no longer need to build; they need to integrate.
Organizationally, the biggest hurdle is the planning horizon. Three-year roadmaps do not work in a market that shifts every quarter. What works are small bets: an agent checkout for one brand, a pilot with one product category, rapid learning. An innovation budget instead of a large enterprise project. And a team that understands data quality as an ongoing task, not as a project with an end date.
The biggest problem for German companies is not a lack of awareness. We all know what is coming. It is an implementation problem driven by “let’s wait and see what competitor X does.”
4) Will Agentic Commerce change competition in retail because AI agents prioritize different criteria than human shoppers?
Yes, fundamentally. An agent ignores artificial scarcity. An agent is immune to staged pricing, even though agents can respond to ads and can be manipulated. An agent does not evaluate product images, but product data. The agent compares datasheets and decides based on utility.
Advertising will also become more fragile. Paid placement in agent output destroys trust faster than paid placement in classic search, because the user assumes loyalty from the agent. As soon as it becomes visible that recommendations have been bought, trust in the entire channel collapses. It is no coincidence that consumers today trust LLMs more than influencers, according to the ECC survey. Platforms know this, and they will have to be more careful than Google ever was with Search Ads. At the same time, OpenAI, for example, has a massive problem with its burn rate.
This also changes the market structure. Compute, model, marketplace, and logistics are becoming integrated. Amazon controls AWS, Marketplace, and logistics in one hand. That is a vertical stack that European players will find difficult to compete with, even if StackIT is positioning itself. The answer is not to rebuild the same thing, but to pursue one of two strategies: alliances or radical specialization. An EU marketplace hub made up of Allegro, Zalando, and bol.com would be conceivable. So would niches where AI cannot provide generic answers because of a lack of data or because of complexity.
One thing is certain: margins will decline in the commodity segment, because perfect comparability erodes all margins. Margins will hold in brand and service. Exactly where many German merchants have invested the least in recent years.
5) What regulatory or ethical guardrails does Agentic Commerce need in order to create trust and prevent misuse?
In my opinion, the EU AI Act only touches Agentic Commerce at the margins. It regulates risk classes of models, not the purchasing autonomy of an agent acting on behalf of a consumer. This gap will need to be closed, and four topics will define the debate. The trustworthiness of AI has so far been more of a topic for high-risk systems, but for consumers it will become a tangible conflict of objectives at the latest once advertising models are involved.
First, liability. If an agent makes a wrong purchase, it must be clear who is responsible: the platform, the merchant, the model provider, or the user. This can be solved, similarly to autonomous vehicles, but it requires clear allocation of responsibility. Otherwise, uncertainty will paralyze the market.
Second, transparency around recommendation logic. Which criteria did the agent apply? Were placements paid for? Is there pay-to-rank in agent outputs? Consumers have a right to know whether a recommendation is neutral or financed. Without this transparency, trust erodes, and the channel loses the very quality that makes it useful.
Third, identity and mandate. Who authenticates the agent as the legitimate representative of a specific user? Shared payment tokens with limits are a technical starting point, but legally we need clear rules around authorization, revocation, and liability limits. Consumer protection questions are also still open: the right of withdrawal for agent purchases, and the burden of proof in disputes.
Fourth, competition. If an agent default reaches 90 percent market share, it becomes relevant under antitrust law. The EU has a tool in the Digital Markets Act that can be applied, but it needs to be translated to agent ecosystems. Otherwise, we will build platform monopolies that are harder to break up than app stores or search engines.
Trust does not arise through regulation alone, but it will not arise without it either. The industry would be well advised to be proactive here before the first scandal leads the debate on its behalf.
Agentic Commerce will not be judged by what is technically possible alone. What matters is whether businesses can establish reliable data, clear control mechanisms, and a trustworthy framework for decision-making. At the same time, platforms, merchants, and lawmakers will need to address unresolved questions around liability, transparency, and competition.
For businesses, that means Agentic Commerce should neither be dismissed as a short-lived hype nor approached through large-scale initiatives too early. What is needed instead are targeted pilot projects, a strong data foundation, and a willingness to continuously test and evaluate new technologies. After all, how quickly AI agents are adopted in commerce will depend not only on their capabilities – but also on how much responsibility customers are willing to hand over to them.
