Back to all
TL;DR
- Many retailers and brands run AI projects, but only a few see real business impact. Rithum’s CTO Ali Irturk and Head of AI Sebastian Spiegler argue that show how results come when you pair AI with clean data, solid infrastructure, and everyday workflows that use it.
- Story 1 – AI and returns: A fashion brand nearly pulls back from denim and swimwear after seeing 60–70% return rates. Once Rithum unifies the data and applies AI at SKU level, the team finds about 200 problem products, fixes titles and sizing, and keeps the categories.
- Story 2 – AI and product discovery: A beauty brand’s hero moisturizer never appears when shoppers ask an AI assistant for “a moisturizer under $30 for sensitive skin.” By cleaning up price, availability, and “sensitive skin” attributes in the catalog, the brand makes the product visible and recommendable.
- Story 3 – AI and marketplace expansion: A European rollout stalls because teams spend days mapping products into each marketplace’s categories and attributes. With Rithum’s Magic Mapper, AI handles most of the mapping across 100+ channels and 30+ languages while humans review, so expansion moves at the pace of strategy instead of manual work.
If every AI project in your company vanished tomorrow, how much would actually break? For most retailers and brands, the answer is: less than you think. In a recent Rithum and eTail report, nearly all senior leaders said they had adopted AI somewhere in the business, yet only 7% described their implementations as “very effective,” and only 4% said they have a comprehensive AI plan.
Many are spending. But few are seeing results.
Ali Irturk, Chief Technology Officer at Rithum, who has spent 20 years as a technology leader, sees the same pattern in the field. “Around 95% of AI projects fail to deliver any real return on investment,” he says. “Only a small fraction make it out of pilot mode and into the parts of the business that matter.”
Sebastian Spiegler, Head of AI at Rithum and a long-time machine learning practitioner, has a similar view from the technical side. He sees models improving every year, but not outcomes. “Models matter, but outcomes matter more,” Sebastian says.
AI readiness is much more than simply mass applying AI solutions. It’s the combination of product and performance data that actually lines up, infrastructure that can handle complex taxonomies and languages, and workflows where AI is part of how work gets done, not an add-on.
The three stories that follow show how that kind of readiness helped a fashion brand avoid killing profitable categories, made a hero moisturizer visible to AI shopping assistants, and turned a European rollout from bottleneck to growth engine.
Story 1: When return rates almost killed two categories
AI fixes a returns problem that made entire categories look broken
A global fashion brand saw return rates of 60–70% in swimwear and denim on a major marketplace. In weekly reports, ecommerce, finance, and marketplace teams all came to the same conclusion: Those categories seemed to be chewing through margin.
Inside the business, pressure was building. Swimwear looked beyond repair. Denim did not seem far behind. Pulling back hard, or even exiting those categories, started to sound like the only responsible choice.
“You can understand why,” Sebastian says. “If all you ever see is a single percentage for a category, it is very hard to argue that the category is healthy.”
The fashion brand’s instinct to act was right. But the data they were acting on was not. The marketplace reported returns at category level only, stitching together inaccurate data. No one could see which products were causing trouble or why. The ecommerce team was lean and focused on day-to-day operations, not deep diagnostics.
Looking past category averages to SKU-level reality
Rithum started with the data. Orders, returns, reason codes, product attributes, and content were pulled into a single, SKU-level view. Only once there was a shared picture of reality did Sebastian’s team run their returns model.
“Once we looked at it properly, it was clear that the problem was not ‘denim’ or ‘swimwear’ as a whole,” he says. “It was a relatively small group of products driving a very large share of the returns.”
Roughly 200 SKUs were responsible for most of the damage
When the team dug into those products, the issues turned out to be concrete and fixable. Many denim titles never mentioned fit, so a shopper could not tell from the name alone whether a pair of jeans was slim, straight, or relaxed. In swimwear, material and sizing details were vague, so what arrived on the doorstep did not match what the shopper thought they were buying.
From “kill the category” to fixing 200 products
Once those offenders were identified, the remedies were straightforward. Titles were rewritten so fit was clear. Size charts were corrected. Materials changed on the worst performers. Over subsequent cycles, returns in those lines dropped and margins improved. The brand kept both categories.
The team went from “kill the category” thinking to “fix these products.” AI did not replace anyone’s judgement; it showed exactly where that judgement needed to go.
For Ali, this is what AI on a solid foundation looks like. “If your own systems are fragmented or disagree with each other, the model is just as blind as you are,” he says. “Getting to a high-quality, centralized view of the data is what lets AI behave like a scalpel instead of a hammer.”
Story 2: The hero moisturizer the AI assistant could not see
When AI-powered shopping leaves your hero product out of the conversation
Search habits are shifting toward natural language. Shoppers still type keywords into search boxes, but more often they are talking to AI-powered interfaces and asking full questions.
“Show me a skin moisturizer under 30 dollars for sensitive skin.”
A global beauty brand, curious about how it appeared in that environment, tried this kind of query in an AI-driven shopping experience. The company had a flagship moisturizer that matched the request perfectly on price and skin type.
The assistant suggested a competitor’s product.
“When an AI assistant becomes the starting point for a lot of shopping journeys and your hero product is invisible, that is more than a quirky result,” Sebastian says. “It is a warning sign.”
The catalog blind spots behind an invisible product
The team looked at the feed the assistant was using. From the system’s perspective, the catalog was cloudy. Prices were unclear and inconsistent. Stock status was hit and miss. Attributes that should have been central to this query, such as “for sensitive skin,” were sometimes buried in long descriptions, or missing altogether.
To a human reading the product page, the offer was clear. To an AI agent reading fields, it was not.
“If your catalog cannot answer basic questions about your products, you cannot expect an AI agent to answer them for you,” Sebastian says.
Cleaning up product data so AI can finally find you
The next step did not involve changing the assistant. It involved changing the catalog the assistant read.
Rithum connected the brand’s feed and focused on the handful of fields that matter most for this sort of discovery. Each relevant product was updated to expose a reliable price, an accurate view of availability, and clearly structured attributes for skin type and primary concerns. In effect, the agentic AI platforms could finally see what the brand already knew about its own moisturizer.
Once the data reached that level of clarity, the same natural language query produced a different outcome. The assistant recommended the brand’s moisturizer and placed it at the top of the list because it matched the budget and clearly matched the “sensitive skin” requirement .
The formulation did not change, and neither did the campaign budgets. What changed was the agentic AI’s confidence that this moisturizer was a good answer to the shopper’s question.
Ali sees a wider shift in what it means to be present in the market. Retailers used to focus on search results pages and shelves. Now, they also must think about how AI agents see their products and whether the PDP and other data builds enough trust to recommend them.
In The 2026 AI revolution in commerce report, more than half of respondents named inventory and demand management as their top AI priority and many reported early wins there. Far fewer said they felt ready for AI-led discovery. The invisible moisturizer is what that gap looks like in front of a shopper.
Story 3: Europe, taxonomies, and turning days into minutes
Marketplace expansion slowed by taxonomies, rules, and languages
A fashion brand mapped out an expansion across Europe. On slides, the plan looked straightforward. The category was in demand. There were plenty of relevant marketplaces. The catalog should have travelled well.
“Europe is not one tidy market,” Sebastian says. “It is dozens of marketplaces, each with its own taxonomy, category rules, and attribute requirements, spread across several languages.”
Every time the brand added a new channel, teams had to repeat a heavy process. They needed to understand that marketplace’s category tree, decide which products fit where, match attributes to local rules, and check that the language on the listing matched how people in that country actually shop. Preparing a few hundred SKUs for a single marketplace could take several days.
“What started as a growth plan turned into a queue of manual work,” Sebastian says. “The limiting factor was not appetite for expansion, it was capacity.”
Using AI-powered mapping to tame European marketplace complexity
Rithum built Magic Mapper, which is powered by RithumIQ, for this exact situation.
At its core are models trained on successful listings across many channels. Given a product, the system suggests how it should be categorized and which attributes it needs to meet the rules of a particular marketplace. A French source listing can be mapped into a German or U.K. taxonomy with proposed categories and fields already in place. Merchandisers then work through a human-in-the-loop workflow, reviewing and correcting suggestions instead of building every mapping from scratch. Those corrections feed back so the system improves as it learns from real decisions.
Turning days of manual mapping into minutes of review
“With Magic Mapper in place, that team went from spending days to spending minutes on the same mapping work,” Sebastian says. “They now support more than a hundred channels and over thirty languages without constantly increasing headcount.”
The tangle of taxonomies and languages did not disappear. It became manageable.
For Ali, this is a clear example of the kind of work where AI belongs. “People are very good at noticing when a mapping is wrong or when a product needs an exception,” Ali says. “They are not good at repeating the same mapping task hundreds of times. AI should handle the volume so your people can focus on the cases that really need their attention.”
Again, the decisive factor was not a particular model architecture. It was the groundwork: consistent product data, reliable connections into channels, and a clear place for AI inside the workflow rather than on the edge of it.
What AI readiness really looks like
Across these three stories, the turning point is not a clever algorithm appearing out of nowhere. It is the moment when the organization has done enough groundwork for any sensible model to be useful.
Inside Rithum, that change started with basics. Rithum pulled catalog, orders, returns, and performance signals into a single, trusted data layer. The team upgraded the infrastructure so systems could handle billions of events, a range of taxonomies, and multiple languages without constant fire drills.
Only after that did AI move from pilots into day-to-day work. Returns analysis, catalog clean-up, marketplace mapping, and other core processes now use AI as part of the standard flow. Governance has grown alongside this, with clear guardrails and named owners in each team who are responsible for how AI is used.
Ali describes AI readiness in four parts:
- Data that lines up across teams instead of fighting between systems.
- Infrastructure that can absorb new models and use cases without constant reinvention.
- Workflows where AI sits inside the tasks that matter, not off to the side.
- Governance and ownership that keep usage grounded and safe.
“It is not a quick, one-off project,” Ali says. “It is an ongoing transformation. But that is what turns AI from something you talk about into something that shows up in your KPIs.”
For a retailer or brand, three questions give a quick sense of that readiness:
- When you see a painful return rate, can you identify the SKUs and attributes behind it or only a category average?
- When you test natural language queries that customers might use, do your hero products appear or does a competitor own that space?
- When you add a marketplace, are you planning for days of mapping work or for minutes?
If those questions are hard to answer, the issue is not a lack of AI. It is a lack of readiness.
The AI engine under the hood
Rithum’s way of operationalizing this philosophy is RithumIQ, the AI engine that runs under its platform.
Rithum’s team describes RithumIQ as the intelligence layer under the platform. It takes the chaos of multichannel ecommerce, billions of signals, and messy catalogs, and turns them into simple signals customers can act on.
The goal is a kind of self-healing ecommerce layer that learns from behavior, adapts where it needs to, and automatically fixes certain kinds of issues.
RithumIQ processes around $50 billion in total sales volume each year and roughly 2.4 billion transactions every day. That scale and variety feed the models behind the three stories in this article—models that can spot patterns in returns at SKU level, recognize which attributes shape discovery in particular categories and markets, and understand how different marketplaces interpret the same product across languages and taxonomies.
From Sebastian’s perspective, the design matters as much as the volume. The system runs on a modern data lake, uses multilingual and multimodal models trained on commerce data, and keeps humans in the loop through specialist engineering teams who monitor and refine behavior.
The decision in front of retail leaders
AI is now part of retail. Almost every large organization uses it somewhere, often in planning, customer service, personalization, or content. Yet only a small minority feel those efforts are working very well, and even fewer can describe a clear strategy for what comes next.
The question is no longer whether to use AI. It is what you are willing to change so AI can actually matter.
One option is to keep collecting tools and pilots: chatbots, recommendation engines, writing aids for marketing teams. The other is slower and less glamorous. It involves cleaning up data, upgrading systems, and reshaping workflows so AI can work on returns, discovery, and expansion rather than only on surface tasks.
Sebastian often talks about a spectrum of AI use. At one end, AI acts as a personal helper. In the middle, generic models sit inside existing workflows. At the far end, models are trained on your own data and wired into the processes that move your P&L.
The models will keep improving whether anyone is ready or not. The real decision for retail and brand leaders is where they want that improvement to show up: in a few clever demos, or in the stubborn parts of the business that customers, operators, and finance teams all care about.
Want to learn how Rithum can help? Talk to our team today.
Talk to our teamQuotes have been lightly edited for clarity.
