AI agents in category management: "That's where the margin is waiting, right now"
15 min read
Disclaimer: This article was originally published in German on etailment on November 1, 2026. The interview was conducted by David Wöllenstein. The following is an English translation of the original piece. Please read the German version here: etailment.
Key takeaways
- More than 11 percent of retailers already use AI agents for category management decisions.
- Agents can spot relevant products three to six months earlier, which can be worth millions in revenue.
- AI agents delivered a gross margin improvement of 2.7 percentage points.
What this means for retailers
By deciding earlier, retailers can improve their margins even in today's consumer slump.
Arber Sejdiji, CEO of Zenline AI, on AI agents that prepare assortment, pricing and promotion decisions in the retail head office, on measurability and autonomy, and on why he sees shopping agents as less urgent.
When people talk about AI agents in retail, they usually mean shopping agents: software that searches, compares and possibly even buys products on behalf of shoppers. Arber Sejdiji draws a line between these and agents that prepare assortment, pricing and promotion decisions in the retailer's head office, and he considers the second group the more urgent one. Sejdiji is not just an observer here: his Zurich-based start-up Zenline AI sells such agents, according to the company to retailers including Otto, Futterhaus and Conrad.
So far, research has focused mainly on shopping agents, and the picture is mixed. According to a release from the EHI Retail Institute dated September 10, 2026, more than eleven percent of the retailers surveyed already use concrete agentic commerce applications, and almost half are planning to, as etailment reported. The study was commissioned by the IT services provider Softserve. According to its authors, the online survey of 218 retail professionals is not representative.
Shoppers remain more cautious. In a survey by ECC Köln from September 2025, 61 percent said they could imagine using AI shopping agents. Only 9 percent, however, would let an agent complete a purchase, payment included.
Mr. Sejdiji, you distinguish between shopping agents, which search, compare and possibly buy products for shoppers, and agents that make assortment, pricing and promotion decisions in the retail head office. What is the difference, and why do you consider the second category more urgent?
Arber Sejdiji: A shopping agent searches and compares products for the shopper. AI agents in the retail head office prepare the decisions that come before that: which products make it into the range, at what price and with which promotion.
If a retailer spots relevant products three to six months earlier than its competitors, it captures the revenue instead of them. For a single product, that can quickly run into the millions. If it pulls weak launches faster, it ties up less working capital. And better pricing decisions deliver the margin that retailers urgently need in the current consumer slump. In my view, that is where the urgency lies for retailers: improving margin within the assortment.
I think shopping agents will take off more slowly than some media coverage suggests. Here, I usually advise clients to get the basics in place: a robots.txt that gives AI assistants access, well-structured product data and well-maintained relationships between products.
And to connect the two: a shopping agent compares a retailer's assortment and prices in seconds, and it can only recommend what the retailer has already listed and priced competitively.
With shopping agents, you currently advise your clients to stick to the basics. Many retailers and service providers see it differently and are already investing in agentic commerce. What is your assessment based on, and what signal would make you advise your clients to invest more?
Arber Sejdiji: With shopping agents, we are seeing exactly what current studies show: according to ECC Köln, only 9 percent of consumers would hand a purchase, payment included, over to an AI. Buying behavior in the mass market changes over years, more likely decades. E-commerce has already shown us that. Going all in too early means high costs and a low return on investment. At the same time, many retailers today regret not taking e-commerce seriously sooner. So our mantra for clients is: get the basics right. But that is all it takes for 2027. I would invest more once AI assistants bring in a noticeable share of traffic or revenue.
You compare the pace of shopping agents with that of e-commerce. In Germany, online accounted for 13.4 percent of retail sales in 2024 according to the German Retail Federation (HDE), and most recently for 17.1 percent in the US according to the US Census Bureau. Which market or categories does your comparison refer to? And what does the slower pace in Germany mean for your timeline?
Arber Sejdiji: The comparison is meant as an order of magnitude. According to the HDE, online accounted for 13.4 percent of German retail sales in 2024. So after roughly three decades of e-commerce, less than one euro in seven is spent online. I expect a similar trajectory for shopping agents, starting in categories where shoppers compare a lot and products can be described well through data, such as electronics. For Germany, that means the timeline is likely longer than in the US. That is good news for retailers, because it gives them time to lay the groundwork properly. The decisions inside their own business, on the other hand, they can improve right away. That's where the margin is waiting, right now.
What exactly does an agent decide for your clients today, and what is still decided by a person? Please give an example: a listing, a price change, a delisting. Who signs off?
Arber Sejdiji: Today, our agents prepare decisions, and the category teams make them. Every recommendation comes with a rationale, sources and the expected impact. Take a price change: every week, the agent compares a product's sell-through with similar products in the category and with the prices of comparable items at competitors. If a product underperforms, it proposes an adjustment early, within the retailer's margin targets and price ladders. The category manager reviews it, signs off and rolls the change out in their own system. Before a delisting, the agent shows where demand is likely to go: to a substitute, whether private label or brand, or to a competitor. For a product launch, it backs up the gap in the assortment with demand signals from the market.
You talk about a three- to six-month head start and revenue in the millions per product, and elsewhere about 2 to 4 percent more margin. How do you measure that: against what baseline, and how do you isolate the effect of the agents from seasonality, price levels or marketing? Please use an anonymized example.
Arber Sejdiji: Our agents make recommendations on pricing, assortment and consumer trends, so we measure different effects depending on the use case. Let me take an example from pricing. We measure with an A/B test: one group of comparable products runs on the agents' recommendations, the other runs as before. That way, seasonality, price levels and marketing affect both groups equally. One example is a mid-sized European furniture retailer with around 9,000 SKUs. New products that were selling poorly used to show up in the weekly reports only after three to four months, and then had to be cleared at steep discounts. The agent continuously compares the sell-through of every new product with similar products in the same category and flags outliers after just a few weeks. In most cases, a small price adjustment was enough. Gross margin rose by around 2.7 percentage points, and there was significantly less slow-moving stock in the warehouse.
Do the 2.7 percentage points describe the difference between the test and control group, or a before-and-after comparison? How long did the test run, and with how many items per group? And which metric do you mean when Zenline elsewhere talks about 2 to 4 percent margin and a 3 to 5 percent EBIT uplift?
Arber Sejdiji: The 2.7 percentage points of gross margin describe the difference between the test and the control group (A/B test). The duration depends on how price-elastic the products are: for beauty products it is more like four to six weeks, for electronics more like eight weeks. You need 15 to 20 percent of the assortment to prove that the measures are effective.
With some clients, we measure the profitability of the assortment and improve gross margin. In that case, we talk about gross margin improvement in percentage points. Other clients want to grow absolute margin, so there we measure the improvement in absolute margin. Depending on the industry and the company's maturity, we typically see a 2 to 4 percent improvement in gross margin.
Where have your agents got it wrong? What was the most expensive or most instructive mistake at a client, and what did you change afterwards?
Arber Sejdiji: Of course, our agents get things wrong sometimes too. So far, that has not been costly, because the agents give recommendations to category managers rather than acting on their own. Mistakes are caught by the retailer's people who work with the agent, and their corrections feed straight back into the agent. To make sure as few mistakes as possible even reach the team, our agents work in several stages: a "senior agent" checks every recommendation against the company's strategy and filters out anything that doesn't fit with high confidence.
Can you give a concrete example of a wrong recommendation? What did the agent propose, how was the mistake caught, and what changed afterwards?
Arber Sejdiji: A wrong recommendation might be, for example, a product listing that the team had already delisted based on internal information the agent did not yet have. That is why, at the start, it is essential to capture all of this information together with someone who knows prompt content and prompting itself inside out. At Zenline, that is the forward deployed engineer, who sets this up together with the retailer.
An agent is only as good as the data it can access. According to EHI, only 39.2 percent of the retailers surveyed rate their data as fairly good to very good. How far can an agent get with the data a typical retailer has today, and how much groundwork is needed before it delivers useful recommendations?
Arber Sejdiji: Very far, and without any groundwork on the retailer's side. We start with public data: the agents read a retailer's assortment from its website and benchmark it against the market. So the first insights, and even recommendations, come without any data transfer at all.
Internal data comes in after that: sell-through, margin, master data and inventory, and later loyalty, traffic and promotion data, for example. The more data, the better the recommendations. Weekly uploads are enough for this. No IT project or ERP integration is needed.
Data quality is the agents' first job: they check attributes against titles, descriptions and product images, fill in missing values from technical data sheets and manufacturer websites, and merge duplicates.
The groundwork on our side is mainly about understanding the retailer's logic: how does it define substitutes, what role does each category play, which margin rules apply, and what does the company's assortment strategy look like? We build this together with the team during the pilot.
As a long-term goal, Zenline describes an "autonomous organization" in which a "swarm of agents" makes almost all commercial decisions on products, prices, promotions and placement. How far is that from today's reality? And who is responsible when such a system delists a bestseller or misprices a category?
Arber Sejdiji: Large retailers such as Amazon, Temu and Shein are already very close to the autonomous organization. German retail still has quite a way to go. Our optimistic estimate is five to seven years. But that is exactly the path we want to open up for other retailers, step by step:
Today, agents prepare decisions and people sign them off. First, tasks run automatically where mistakes are cheap and easy to correct, such as product data. Next come decisions with clear guardrails, for example price adjustments within fixed margin and price limits. For far-reaching decisions such as delisting a product, sign-off stays with people for as long as the company wants it to. Our job is to make every recommendation traceable: with a rationale, sources and the expected impact. No category team today would deploy a system whose decisions nobody can check.
And who is liable for a wrong decision once agents act without sign-off: the retailer, Zenline, or both, shared by contract?
Arber Sejdiji: The agents make recommendations. Decision-making authority stays with the team, as it always has.
What becomes of the category manager when agents steer assortment and pricing? Does the role shrink, change, or become redundant?
Arber Sejdiji: How has the job of a Wall Street banker changed since machines started making the decisions? More decisions are made, at much shorter intervals. The people behind them are now more responsible for strategy, for understanding the market and for improving the model. We are seeing something similar in retail. Decisions such as price adjustments, listings or promotions are becoming more frequent. Category managers who used to spend two weeks adjusting planograms will have a far more strategic role in the future: understanding their customers even better and serving their needs even more precisely with their own assortment.
A provider like Zenline works with several retailers, some of which compete in the same categories. Competition authorities, including Germany's Federal Cartel Office (Bundeskartellamt), have been looking into whether pricing algorithms can lead to prices converging for years. How do you make sure your clients' pricing decisions remain independent of one another? And are data or models used across clients?
Arber Sejdiji: No, they are not. Of course, we get this question a lot, especially on pricing. Data and models are not shared across clients. Every client works in a completely separate environment with us. Client data, agents and any input stay with the respective client, and there are no shared learnings or interactions between the agents of different clients. Internally, too, we have processes that define which employees may see which data from which client. We are ISO 27001 certified, and our databases are hosted in Europe.
It is also important to note that we do not train our own models. We teach existing models relevant retail workflows without training them on retailer data. Retailers share their most sensitive numbers with us. If they couldn't rely on no competitor benefiting from that, we probably wouldn't have a single client.
You say data and models remain separate. Do competing clients use the same base models and the same pricing workflows? And how do you prevent the same logic from pushing prices in the same direction?
Arber Sejdiji: Data and models stay separate, and you can think of it the same way as with people. Large consulting firms enforce a strict ban on exchanging information between consultants working for competing clients. The same must of course apply to AI agents, otherwise nobody would hand over their data. The terms for this are "Chinese wall" or multi-tenant separation. Accordingly, every retailer has its own customer base, its own customer behavior, its own assortment and therefore, of course, its own assortment and pricing strategy. As a result, every pricing workflow is completely different.
In early September, Anthropic released a template for commerce agents. Alongside a shopping agent, it includes a merchant agent for inventory, pricing and marketing. Are merchant-side agents becoming a standard feature of the big AI providers, and what is left for specialists?
Arber Sejdiji: One thing is bringing client-specific knowledge into the models. That takes a very, very deep understanding of the business: bringing different data sources together and understanding what they mean. It is an enormous amount of work to make an agent enterprise-ready, so that it doesn't confidently give an unfounded answer.
The second is that there are very specific software applications, such as planograms, pricing management or promotion planning, in which people still spend hours every week. ChatGPT or Claude, in other words the big model providers, are not going to go into this 30-year-old software stack to improve it.
How do you assess the Anthropic template specifically, with a merchant agent that recommends prices and promotions? And how do you view established planning software vendors that are building agents into their own products?
Arber Sejdiji: The Anthropic agent is perfect if you have all your data highly structured and readily available to the agent. Especially if you are very advanced in IT, have good scrapers of your own, competitor data, market data, social media data and so on, it is of course an excellent tool. In reality, our clients don't have this data readily available, and making it properly accessible is very complex. So it is well suited to small online retailers. Larger retailers need a lot of groundwork and, ideally, a lot of market context rather than just working inside their own assortment.
You point to the consumer slump, in which retailers urgently need margin. But in weak years, retailers also cut IT budgets. What should a mid-sized retailer with a few thousand SKUs and no in-house data science team do over the next twelve months, regardless of which provider it works with? And roughly how much does it cost to get started?
Arber Sejdiji: Don't try to build your own workflows with Claude. The complexity that comes with it is consistently underestimated. Ask to see an eight-week plan, and measure it against your own three criteria for a successful pilot, such as delistings or margin improvement through pricing. Then decide whether to roll it out.
What can a retailer prepare on its own, independently of a provider? And what is the rough cost range for a pilot?
Arber Sejdiji: The retailer doesn't need to prepare anything. We can start with public data from its online shop.
Which of your theses do you think is most likely to be outdated a year from now?
Arber Sejdiji: Most likely the thesis that every recommendation has to be signed off by a person. That is right today, because trust has to grow first. But the models are improving so fast that the recommendations will soon be reliable enough to implement many of them without sign-off. I could imagine that, a year from now, the first price adjustments within fixed guardrails will run automatically. Category managers will then set the guardrails and focus mainly on the big questions, such as assortment strategy, private labels and suppliers.
About the interviewee: Arber Sejdiji is CEO and founder of Zenline AI. The Zurich-based start-up develops AI agents that prepare assortment, pricing and promotion decisions for retailers. According to the company, its clients include Otto, Futterhaus and Conrad.
About the interviewer: David Wöllenstein is an editor at etailment and "Der Handel". He writes about e-commerce, retail technology and digital business models, most recently with a focus on agentic commerce and the use of AI in retail. He previously worked at the advertising magazine Campaign Germany.
