Zenline AI: Agentic Assortment Decisions Win Retail Margins | Startuprad.io

What Is This About?

Zenline AI uses agentic AI to make assortment decisions for retailers — automatically optimizing which products to stock, price, and promote to maximize margins. The startup brings autonomous decision-making to one of retail's most complex and margin-sensitive challenges.

Introduction

Retail margins live or die on assortment decisions — which products to stock, where to place them, and when to pull them. Zenline AI brings agentic artificial intelligence to this challenge, turning weeks of manual analysis into rapid, data-driven product actions that can lift margins by 1–3 percentage points while exposing hidden cannibalization. In this interview, the founders explain how their technology helps retailers move from intuition-based merchandising to automated, intelligent assortment optimization.

Zenline AI deploys agentic artificial intelligence to automate retail assortment decisions, delivering 1–3 percentage point margin improvements for retailers. The system identifies hidden product cannibalization patterns that manual analysis consistently misses. Unlike traditional analytics dashboards, Zenline's agents take direct action — adjusting product placement, pricing, and inventory allocation autonomously. The technology represents a shift from descriptive retail analytics to prescriptive, agent-driven merchandising optimization.

Agentic AI turns retail assortment from slow analysis into fast product actions, lifting margin 1–3 points and exposing hidden cannibalization.

Agentic AI turns retail assortment from slow analysis into fast product actions, lifting margin 1–3 points and exposing hidden cannibalization. Startuprad.io brings you independent coverage of the key developments shaping the startup and venture capital landscape across Germany, Austria, and Switzerland.

This founder interview is part of our ongoing coverage of Scaleup Founder Interviews from Germany, Austria, and Switzerland.

Agentic assortment optimization is the fastest path for European retailers to defend margin under platform-level competition. The strategic shift is not better analytics. It is moving from human spreadsheet workflows to autonomous product-level actions that surface margin leaks and correct them at scale.

• Retail margins are now won by decision velocity, not better dashboards.

• Agentic systems make the long tail visible and actionable, not just measurable.

• Most retailers misread assortment breadth as “customer choice” instead of operational debt.

  
  

Why agentic assortment optimization beats retail dashboards

Dashboards describe retail problems. Agents act on them. Agentic assortment optimization compresses weeks of category analysis into near-real-time product decisions, including delisting, substitute mapping, and margin leak detection across hundreds of thousands of SKUs.

European retail is in a margin war driven by platform competitors operating at extreme speed. Traditional category management cannot match that velocity because human analysis does not scale across millions of SKUs.

Agents change the operating model. The category manager stops being the analyst who produces the decision and becomes the decision-maker who reviews and refines system outputs.

Arber Sejdiji (CEO, Zenline AI) frames the competitive edge as velocity: decisions that once took weeks must happen in minutes, or retailers structurally fall behind.

What breaks first when the long tail becomes visible

The first failure is organizational discomfort. When the long tail becomes visible, it exposes data flaws and operational misconfigurations that were previously hidden by focusing only on top sellers.

Retailers often manage 10–20% of the assortment well and ignore the rest. Agentic systems surface errors at scale: wrong images mapped to products, broken attributes, inconsistent pricing, and misconfigured listings.

These are not minor hygiene issues. They drive returns, confuse customers, and create silent margin loss that compounds across the catalog.

Sejdiji describes the initial reaction as “inconvenience” because AI forces teams to confront what manual workflows could never fully reveal.

The curated assortment myth: consumers do not want infinite choice

The most fragile retail narrative is that consumers want unlimited assortment breadth. Data shows the opposite: curated assortments reduce decision paralysis and improve both revenue and margin.

Retailers often expand assortment breadth because they assume consumer demand is perfectly explicit and must be satisfied with maximum selection. This belief produces operational complexity, weakens brand clarity, and increases inventory and logistics overhead.

A curated assortment is not a marketing concept. It is a margin system. When retailers decide what belongs on the shelf, rather than outsourcing choice to “the consumer,” financial performance improves measurably.

Sejdiji reports margin uplift of 1–3 percentage points and revenue gains of 1–3% when retailers move toward curated assortment decisions.

Margin cannibalization is more dangerous than revenue cannibalization

Revenue cannibalization is expected. Margin cannibalization is strategic failure. Retailers often push consumers toward products that look premium but deliver weaker margin than cheaper private label alternatives.

Retailers rarely have relationship mapping robust enough to detect similarity and substitution effects across tens of thousands of products. Without that model, they cannot quantify what happens when a product is added, delisted, or repriced.

The counterintuitive case is margin inversion: a higher-priced branded product can yield worse margin than a lower-priced private label substitute. That flips the incentive structure and makes “premium” optimization financially irrational.

Sejdiji highlights that margin cannibalization often hides behind intuitive revenue logic because relationship modeling is missing in most retailers.

The real bottleneck is not supply chain. It is category analysis throughput

)Retail decision velocity fails when organizations cannot convert data into decisions fast enough. In many retailers, supply chain can execute faster than category teams can analyze, making human analysis the limiting constraint.

Some private label workflows take 12–18 months from ideation to shelf delivery, which makes rapid AI-driven decisions irrelevant. But many retailers already have operational capacity to move faster.

The bottleneck is category management throughput. When agents automate desk research and relationship detection, category managers can shift from weeks of analysis to hours of review and refinement.

Sejdiji distinguishes between retailers constrained by supplier lead times and those constrained by slow analysis cycles.

Inline Micro-Definitions

• Agentic AI means autonomous systems that take actions, not just generate insights, based on defined objectives and constraints.

• Assortment optimization means deciding what products to list, delist, price, and prioritize to maximize margin, revenue, and operational efficiency across the full catalog.

• Margin cannibalization means shifting demand toward products that reduce total profit even if revenue stays stable, often because substitutes have different margin structures.

• Relationship mapping means modeling similarity and substitution between products so retailers can predict how changes to one SKU affect others.

• Decision velocity means how fast an organization can convert market signals into executed product decisions across the catalog.

  
  

Operator Heuristics

• Treat agents as junior analysts, not as oracles.

• Optimize for margin outcomes, not analysis completeness.

• Clean data only when it blocks a margin-critical decision.

• Model internal substitutes before benchmarking competitors.

• Reduce assortment breadth when it increases operational drag.

• Convert weekly cycles into continuous review loops.

WHAT WE’RE NOT COVERING

We are not covering retail media networks, loyalty program design, or warehouse automation. These topics matter, but they do not change the core decision system: assortment and pricing are now won by velocity and relationship modeling, not incremental tooling upgrades.

Relationship Map

• Jörn "Joe" Menninger → Host of → Startuprad.io

Automated Transcript

1 European retail is in a margin war. It didn't choose. 2 Assortments move faster than teams can analyze data 3 quality collapses on the scale and global platforms like 4 Amazon and Temu operate on a level of speed that 5 exposes every structural weakness. 6 Abba SETI built Zenline AI 7 to close that gap. Agents that act not 8 observe systems that detect margin leaks, 9 resolve clutter and surface opportunities 10 humans never see. Today we examine why 11 autonomous decision agents are becoming to operate the system 12 of modern retail. And what happens when a retailer 13 shifts from dashboard analyzers to product level action 14 in seconds. By the end, you'll understand why the next 15 competitive edge isn't scale its velocity. 16 Welcome to StartUpLead IO, 17 your podcast and YouTube blog covering the German 18 startup scene with news interviews and 19

live events. 20 Hello and welcome everybody to Startup Rio. Today my guest is 21 Arba Seti, the CEO and co founder of 22 Zenline AI from Switzerland, Goodsy 23 Goody, a company that is redefining how European 24 retailers decide what to sell, what to move and where 25 the real margin sits. With experience at BCG, 26 Procter Gamble, ETH Zurich and Research work at 27 Cambridge, he brings both the strategic clarity and the 28 technical depth required to rebuild assortment logic. 29 From first principles. Zenline agents 30 clean messy catalog data, detect cannibalization, 31 map substitutes and reveal trends before they 32 become mainstream. They already process more than a million 33 SKUs and deliver recommendations with measurable 34 financial impact enable retailers to operate with the 35 speed of global platforms without the overhead. 36 Aber is here because his work signals a structural change.

37 Decisions that took once weeks now happen in 38 minutes and the retailers who adapt will redefine 39 competitiveness in Europe's Euro trillion 40 market. That was an intro, don't you 41 think? Welcome to the show and let us 42 get right into it. You said the real power 43 center of retail isn't price or promotion, but 44 but the assortment itself. When did that 45 insight crystallize for you? In my previous 46 job with bcg, we were advising retailers with assortment optimization. 47 And what becomes very obvious when you 48 work with retailers on the assortment is that the assortment touches the entire 49 organization. So it's not only the top line or the bottom line as is 50 for pricing and for promotion, but 51 certainly it also touches the top line. So if a customer, for example, 52 doesn't find a product that it wants in the stor stores or on the

53 shelves, that's lost revenue. However. Moreover, it 54 really defines your brand. So the brand is typically 55 not only shaped by the marketing, but 56 it's mostly shaped by the products you offer. So you can imagine when 57 a consumer has a demand, what do they first Think of what's the 58 brand, what's the retailer that they think of to go and 59 buy the product from? And that's so important to be top of mind for the 60 consumers. And, and that only works with an assortment. That consistency 61 delivers what the consumers actually want. And then the 62 last one is the assortment. And the depth of the, and the 63 breadth of the assortment influences the 64 complexity of your operations, the logistics, the supply chain, the 65 inventory management. So the broader the assortment, 66 the higher the complexity and the more difficult it is to maintain

67 lean operations. When you talk about 68 assortment, you have something in mind. For example, when I was 69 thinking about this interview, what I had in mind when I was reading this, 70 at first I was thinking, okay, 71 option one is I look for the 72 basics. That is a retailer with a yellow logo. Then I look 73 for some more fancy stuff. Then that comes in a 74 retailer with the orange logo or a red logo. 75 Is that something that you're thinking about here? Yes, 76 exactly, that's exactly it. What comes to your mind when you want to 77 buy a product, what is it? Is it someone who's super, super trendy 78 and always knows what the consumers want and can offer it, offers it 79 fast, or is it typically 6 to 12 months behind 80 the schedule? Your early

81 discovery work exposed how little transparency 82 retailers had into their own margin drivers, what 83 revealed the debt of that blind spot. 84 So the importance of that blind spot comes basically 85 what you said in your intro. It's a super, super, super competitive market 86 with Temu and Shein entering the European market not only in 87 fashion, which a lot of people think, but it's across all categories from electronics to 88 fashion to pet supplies. And retailers 89 are too slow because of the analysis, just take too 90 long, which should be done automatically and systematically with 91 AI. So basically we're entering 92 this competitive, basically war 93 in retail coming from different countries from different angles. 94 And if retailers don't move fast 95 enough, don't filter out products that have bad 96 margins and focus on the ones, focus on the entire assortment. Instead

97 of managing just 10 to 20%, European retailers will 98 not be able to cope with the speed and with the 99 velocity of these big platform players. 100 So that you're mostly targeting 101 this towards retailers that would 102 compete with Amazon, Temu, Shine, 103 Alibaba Express and so on and so forth. 104 Yes, we're trying to basically help them to have the 105 same data, science and AI capabilities as 106 these platforms have, who have actually their own internal assortment 107 teams to make these decisions. And you're 108 now entering a market that is basically shaped 109 by Amazon level speed. You order one day, you get it, 110 the Next day or a little bit later. What convinced you that 111 Europe's response had to be agentic and not 112 incremental? So what's happening right now is a perfect storm. 113 We're seeing years of inflation, declining consumer confidence,

114 increased saving rates and what's making it now 115 very hard for retailers to stay ahead the curve and stay 116 profitable. And this does not only apply to 117 retail, right? We're speaking about all industries that are 118 basically undergoing this wave of the industrial 119 revolution or let's say revolution led by general purpose 120 technologies. And in these times of general purpose 121 technology is changing not only one but multitude of industries, 122 the answer has never been incremental innovation. And US and China will 123 move ahead and they will basically equip millions of white 124 collar workers with agentic AI. And speed will be the 125 differentiating factor, launching new products, understanding 126 the assortment, understanding relationships in the assortments much, much, much 127 deeper, much earlier than all competitors will. 128 And again just to to basically highlight that if we 129

still think we can do the same, the 130 same thing that we did for the past 20 years, which is incrementally 131 increase efficiency, these companies will 132 basically frog leap the European retail industry. 133 In the entrepreneur's world, we'll be talked 134 about the moment that data forced a decision 135 you didn't expect. Stay tuned and learn 136 more down here in the link. 137 We've been talking category teams already that some of our large 138 competitors have. They focus on top 139 sellers and ignore the long tail. What breaks inside 140 an organization when the long tail suddenly becomes 141 visible through good data, through AI, real time 142 observations? I would say the first reaction is 143 inconvenience. There's a lot of things that get exposed, like 144 data flaws, wrong images linked to wrong products which 145 then lead to very high returns because consumers got the

146 wrong product which they actually didn't want to order. But the product 147 image on the website was leading to the wrong product 148 attributes that are misleading or misconfigured or inconsistent 149 pricing and all of these errors, or 150 let's say inconsistencies, 151 misconfigurations in the assortment become 152 very obvious when you start focusing not only on the 153 1015 to 20% of the assortment that is typically 154 managed very well, but then 155 surfaces and this then creates, let's say 156 inconvenience at the beginning. But 157 organizations then learn to accept the fact that it was 158 very, very hard to manually be able to manage 159 hundreds of thousands, if not millions of SKUs. But now with the 160 power of AI, you can actually do what retail 161 wasn't able to do for the past. Basically for the past decades. 162 You'Re the agents you provide for the companies, they do

163 challenge internal narratives. Which narrative 164 collapses the fastest once real data 165 pressure was applied? You said first the flaws are exposed and 166 then I think the next step is their internal narratives breaking 167 down. Yeah, very, very, very good question. And I would 168 say the most common narrative is the intern. 169 The most common internal narrative is that the consumers know 170 exactly what they want to buy. And this consequentially leads to 171 an increased breadth of the assortment. And we as read retailers, 172 we don't have power to because the consumers have all the power 173 and they will either way just buy what they want. If you don't have 174 it, they will not buy. However, data that very, very 175 early on will show that curated 176 assortments is actually what consumers want. They don't want 177 decision paralysis, but what they want is a curated

178 assortment. Exactly. Curated to their 179 demands and their basically the consumer behavior. And 180 what you typically see immediately is that revenue and margin 181 goes up significantly. We're speaking 1, 2, 3 percentage points 182 of margin when the 183 retailer makes the decision of what should be in the store, in the stores 184 instead of basically giving up and saying it's always, it's the 185 consumer who's always right and we need to offer everything that they 186 can possibly imagine. 187 Part of your thesis or your main 188 part is speed. What's the cost of 189 operating a full year behind the real markets decision 190 Tempo. I just mentioned it. We're speaking 1, 2 percentage 191 points in margin improvement and we're speaking 1 to 3% in 192 revenue increase. Very often we're not even 193 one year behind the real market, especially in the private

194 label segment of the businesses. And in Switzerland, I think for the 195 first time, private label revenues surpassed 50% of the market. 196 So private label is now making more revenue than the branded market 197 for retailers. And in Germany, in the rest of Europe, 198 I think it's close to between 30 to 50%. 199 And there we're speaking very often about 18 to 24 months 200 of let's say distance to real consumer, 201 consumer changes. I do have 202 in mind that basically we talked about here the 203 less fancy products, especially talked about food here 204 and clothing. There's a big online retailer who offers 205 essentials and, and that's kind of stuff we talked about here, 206 that that's retail brands, that's the brands they come up with themselves 207 and for some larger specialty food retailer, they are meant by, they

208 are produced by household names. So that' so 209 pretty interesting area to watch. But let us go a little bit further 210 and talk about 211 what impact the data had on teams because you've Watched 212 messy data, paralyzed teams. What's the silent 213 limiter they underestimate most? 214 So I would like to answer slightly 215 different question which is, and I'll get back to yours, which is 216 companies very often think that data overall 217 is the limiter of, for the generation of the value 218 margin, improvement, revenue increases, making faster 219 decisions, and then they very often focus on the 5 220 to 10% of the data that is imperfect, where the master data 221 is not clean, where the attribute data is not clean. And then 222 organizations often stagnate and move very, very, very slow 223 because they think of, oh, data cleaning being this one

224 incredibly big IT exercise that they need to perform. 225 However, what should happen is that 226 while doing these use cases, you clean the data as 227 you need them because you will figure out where the data leads to. Wrong 228 suggestions, wrong recommendations, and then it 229 makes sense. And then there's also a business case to clean up 230 the data and make it worth this exercise. 231 And what a lot of companies and retailers have not realized yet 232 is that agentic AI is making data cleaning and insight generation 233 so feasible. The cost to or the cost and the time 234 to get cleaner data is reduced on a daily basis 235 and retailers can already save millions in margin 236 without having the perfect data. So instead of focusing too much on the 237 limiter, I think companies do that either way all the time.

238 The message should be although data can be messy 239 and too much data can paralyze, the 240 limiter is very often the mindset towards the data. 241 When we talk about agents, many people 242 have absolute neutral 243 AI in mind. But every agent encodes 244 a judgment, either by the person who decides what to 245 encode or the person who does the encoding. Which 246 judgment was the most difficult to translate into 247 an autonomous logic? 248 Very good question. And I have to say 249 the entire space is very, very, very early and it will change in a 250 year from now. But 251 what is happening right now is that agents are very, very, very 252 sensitive to a few sentences in the prompt in how they 253 make the final decision. So because of 254 that, as you were saying, rightfully, the

255 design team of the agent, of the system prompt of 256 the context engineering that allows the agent to make a decision 257 of which products to recommend. One very 258 simple example might be that one of our customers 259 doesn't want to be the most trendy firm. So they don't want to just launch 260 products that are, let's say, that 261 are just coming to the market, but they want some validation, some proof points, 262 and to know that this product is actually performing and will be here two, three, 263 four years from now. And now the question is, what is A trendy product, 264 when does the trend turn into something more sustainable? 265 And in the end, the makers, which we are 266 here with Zenline, need to be very, very, very thoughtful 267 of how do we encode that? And I think

268 the most difficult part of this is the memory layer. So 269 you do one recommendation and then the question is next week, 270 next month, next quarter, how do you encode all the recommendations 271 that you made? How do you encode, let's say, the 272 feedback from the real world? Did the product launchers 273 succeed or didn't they succeed? How do you bring that in and how 274 do you feedback that into something that the agent 275 can use to then be able to make better 276 recommendations next time? Because this is something that is incredibly natural to 277 us as humans. We do something, we, we observe what happens 278 and we have an initial gut feeling of whether we should do that again or 279 what we need to adjust. And this is something that's very hard to, to encode 280 into. Into autonomous systems. Yes,

281 there's a lot of stuff to really hard to encode in 282 autonomous systems. And I think the more systems 283 we'll have, the more it will be 284 of importance. But let us 285 go to the next question, because I was wondering, 286 your launch validator runs on social signals, 287 competitive shifts and emerging demand. When 288 did you trust its signals over the human 289 instinct? So what's 290 important to distinguish is whenever you launch a product, 291 which is what we aim to support with the launch validator 292 is we want to strengthen the human 293 instincts by giving it more data. However, 294 as every single human being, category managers, for example, in the 295 fashion world, are biased by what they see on their 296 channels, their social media channels, by walking through the city, watching 297 colleagues, friends and so on, what they wear, and

298 then using hundreds of thousands of data points 299 actually allows them to double check your gut feelings 300 and use these data points from social media, from Google shopping, 301 from competitors, bestseller data, to see whether 302 your initial gut feeling was right 303 or whether you were missing something. And it can be much, much, much more 304 nuanced, especially in the fashion world, where new designs 305 from quarter to quarter, from year to year very often change 306 in subtle ways which are very hard to grasp. So it's not 307 trusting the signals over human instinct, but it is challenging 308 the human instinct, which can be biased by their own world, which is very often 309 not representative of, let's say, the 310 consumers, what they want in certain categories. Always, 311 always when it's working in consulting and 312 always if you do have strategy project at one point, there's always

313 the question of internal cannibalization. So 314 if the, if I'M adding product 315 A, does it take away from B, does it take away from C? 316 And your consolidation agent exposes 317 internal cannibalization. What's the most 318 counterintuitive callization call that 319 surfaced? There's always revenue 320 cannibalization and there's margin cannibalization. And the 321 revenue cannibalization is basically where a lower priced product 322 is purchased by the consumer because 323 although there is a duplicate which is maybe higher priced and 324 is the one product that you as a retailer would want the consumer to 325 to purchase. And the issue here is that 326 revenue cannibalization is very often not counterintuitive. 327 It's very intuitive. However, it's never in the data. I haven't seen a 328 single retailer that has good relationship model across 329 tens of thousands, hundreds of thousands of products to know which products

330 are very similar because you would need a very good 331 relationship mapping of all products, similarity scores, 332 quite good attribute data to figure out which products are 333 closely related to each other and what is the impact of one to the other. 334 However, margin cannibalization is very often 335 counterintuitive. You take for example two 336 examples that we had are drinks where 337 branded products that are higher priced 338 had significantly worse margins than the product than the 339 private label brands that are cheaper and have higher margins. 340 So in these situations what happens very 341 often is that companies 342 do not understand sufficiently the 343 impact that let's say price differences and then the relative 344 margin of these products has on in the end the 345 purchase decision of the consumer and then incentivize 346 the consumer to purchase the wrong product

347 by for example, reducing prices by watching 348 the market too closely or trying to increase 349 velocity of the product. We'll be back after a 350 short ad break. And after the break we step into the 351 strategic misread that holds retailers back and 352 where the next advantage isn't data but decisiveness. 353 Hey guys, welcome back to our interview with Arbor from 354 Zenline. Helping retailers to increase revenue 355 and margins. Working with AI here and I was 356 wondering, your weekly decision cycles replace 357 seasonal ones. What operational structure 358 fails first when a retailer attempts that 359 transition? And also a little bit tongue in 360 cheek question, when are we at daily 361 decision cycles? Very 362 good question. I have to say we 363 have spoken to many retailers and we're working with many retailers where the decision 364 cycles can sometimes take more than 12 months. We were

365 seeing very good examples and we're seeing bad ones. 366 The operational structure that fails very often is the 367 relationship, especially in the private label business to the 368 suppliers where it can take up to 12 to 18 369 months to actually supply the product. So from 370 let's say initial ideation to concept development 371 to iteration to receiving the stock that's required 372 to sell the product, sometimes it takes between 12 to 18 months, which 373 in the end makes AI driven quick decisions 374 completely obsolete because you don't need it if your supply chain process in the end 375 can't deliver what you're trying to do organizationally in 376 terms of commercial output, delivering the assortment that consumers want. 377 However, with 378 many of our retail customers, we're also seeing very good examples where the 379 operational architecture actually is already in place. But the

380 bottleneck was never the supply chain delivery but, but it was the 381 analysis and the inside gathering because it's bottlenecked at the category 382 managers. And if the category managers simply can output 383 more decisions, the let's say supply chain 384 behind it is already supportive enough, is already, let's say 385 established so that they can quickly start to 386 see, see these products being in the 387 shelves within the weeks. On the question when are we on daily decisions? I 388 think daily decisions will not be the bottleneck, 389 but daily supply chains will be very hard to achieve for 390 retailers in most industries. If we wouldn't be on a time constraint, 391 I would go into 392 deviate a little bit and go into some thoughts, 393 experiments, what we could do with Jones here. But let's take 394 our main line of interview questions here.

395 Rejected recommendations reveal a lot about the 396 organizational psychology. What patterns do you see in 397 teams that aren't ready for agentic systems? I would 398 say organizations that focus too much on the errors that happen, 399 or let's say the parts of the 400 recommendations that are misleading, that are error prone, that 401 are just simply not useful at the beginning. And 402 organizations that focus too much on what's not working 403 and where the data is not sufficiently good are 404 the ones that will become the laggards 405 of the, of the decade. So however, on the other 406 side, the teams that are utilizing agentic 407 systems most effectively basically treat them as junior 408 analysts, I would say. So you have a junior analyst. What happens if 409 you receive a lot of input that will be good, some will be exceptionally good

410 and some will be bad simply because it misses context, it doesn't know 411 exactly what has happened in the last 12 months of the organization. 412 Sometimes recommendations are bad because they have been done, 413 they have been tried in the past, maybe 12 months ago, 24 months ago. 414 And certain products, certain product types or categories just don't 415 work with the retailer. And a 416 junior analyst also wouldn't know them. So what very 417 good, what patterns of very good teams are 418 is the willingness to give feedback, give 419 nuanced, and detailed feedback to advance the memory of the system and 420 give more context into the system, define the strategy, the company 421 strategy, the category strategy, and thus be able then 422 to improve it over time again, as you would do with the junior analyst. And 423 I think the more organizations start to think of

424 agentic systems of colleagues, of basically just another part of 425 the workforce, the faster they will be the ones that actually 426 gain margin benefits from it. 427 Your pricing agents benchmarks 428 internal logic against the market reality. 429 What's the structural pricing error that retailers report? 430 Okay, so I haven't seen a single retailer yet who 431 has been able to be very strong 432 in utilizing internal relationships 433 in their pricing. So what happens very often is that many retailers 434 are very much focused on hey, how is this one single product 435 priced compared to the market to other other 436 retailers selling this one product? What every 437 retailer does is they compare their own assortment 438 very much with the assortment of other retailers. 439 And what they neglect is that a lot of the pricing, 440 a lot of the pricing power that they have is not

441 comparing themselves to competitors, 442 but to the internal assortment. Because very often when a consumer is 443 already in the stores, they will make a decision 444 not based on what is the price of the competitor, but they will make a 445 price based on what are the pro, the other products, the 446 alternatives, the substitutes that are already in the shelf. 447 And especially with regards to decisions that impede the margins, 448 customers just retailers typically don't have a, let's 449 say, sophisticated enough system to 450 understand the relationships. As you mentioned before, the 451 cannibalization impact of not only delisting a product, but also 452 changing the price of one product or what does it do to all the other 453 products in the assortment. And this relationship model is something that 454 retailers should think about. However, they're doing it 455 way too little because it's very complex technically.

456 So the classical substitute products 457 substitute variant. Pricing as soon as you get to 458 10,000, let's say 10,000 products. It's 459 very hard to manage a very clean relationship mapping of your 460 products so that for example, when you change one product, you also change 461 either in, let's say proportionally 462 or at the exact same level. The price of variance can also 463 be the same for substitutes to just understand what does the change 464 of one product's price due to the rest of the, let's say, 465 very similar, very related products. 466 We've been talking about margins a lot and leaders often 467 freeze under margin pressure. What differentiates 468 those who accelerate versus those who stall? 469 I think a lot is in the tools. Some 470 retailers still consider power bi as state of the art 471 and A lot of them still stick with Excels, which is

472 very convenient because they don't have to learn new tools. But the 473 spreadsheet logic is failing. It's failing as soon as you have 474 hundreds of products, not even mentioning 475 thousands or millions of products. And I think retailers 476 just need to accept the fact that let's say human 477 driven and spreadsheets driven 478 working modes are not optimized 479 to support retailers where decisions are 480 incredibly complex. And yeah, in the end, 481 competition will not wait. Competition from other markets, 482 new more digital versions of retailers popping up. And 483 the margin winners will be the ones that rewire how 484 fast product decisions can be made because they have the data 485 ready, readily available, and then they have the systems that 486 output recommendations and then review the 487 recommendations. So basically the shift from going I need to do

488 the spreadsheet analysis myself to I 489 only review the output of systems and I can then 490 fine tune for a few hours instead of being the one who has 491 done the analysis for weeks. 492 I know how this feels. 493 You gotta be creative here. You guys 494 scaled at a pace even few teams 495 attempt which moment 496 demanded your highest level of personal, personal resilience. 497 Honestly, there, there are many. As a founder, you constantly 498 are on the edge, but when you have a strong team, 499 you, you can go through very, very, very long hours 500 together. And I think that's the happiest and the 501 best parts. But for me it's a combination of several 502 countless sleepless nights and pressure to keep customers happy. 503 Yeah, but I would say whenever 504 the team is willing to go the extra mile, is willing to

505 go beyond, to basically be able to deliver 506 the results that we are promising with the agents, that's the 507 moments where let's say, well, that's our 508 source of energy. 509 Your source of energy. Okay, I get the 510 feeling that Asian driven operations are 511 here compressing time, which is just talked about quarterly to 512 weekly in the interview. Which 513 legacy workflow becomes obsolete here? 514 First. 515 White collar 516 tasks that require desk research. 517 All the research work will be automated in the 518 future. And you can see this already in consulting, 519 which they have huge pressures. It's the same for 520 lawyers, it's the same for any kind of research assistants. Because 521 I don't know if you have used deep research from 522 Gemini and ChatGPT, but it's incredible how 523 much depth and how much data can be processed with these

524 tools. And they can do it significantly better than humans can. 525 So everything related to research and analysis will be something 526 that will be obsolete in the future. And humans will 527 switch from doing the analysis to making the decisions. Basically exactly what we 528 discussed before with the leaders of margin drivers 529 and compressing months of analysis 530 into days, if not sometimes even hours. Making 531 fast decisions, making counterintuitive decisions 532 makes explainability become something strategic. 533 How do you maintain trust when high impact 534 recommendation contradicts their intuition? 535 I like that question, especially in 536 an In a machine learning driven world that we had 537 for the past, let's say between 2016 538 to 2022, explainability was 539 really hard because you had a result. 540 Your machine learning model told you this should be the new price, 541

but there was very little explainability behind it. 542 What is now interesting with generative AI is that the 543 models have reasoning tokens and they can 544 output and they can explain why they came to a 545 decision. And this reasoning is getting closer and 546 closer to the reality of why the model decided to do what it 547 decided to do. And this is actually one of the things that we have realized 548 with customers is they love to see the thinking 549 of the model. So not only the final output of why you did it, but 550 even the steps, the thinking steps that happened, the tool calls that 551 happened, the web search, which websites did it browse, all of 552 that to make it very very very transparent. 553 So I think in the last 10 years there was a bit 554

of the dogma to 555 hide, a bit of the complexity behind it, to not confuse 556 especially the managers, but also the category 557 managers and also the executive teams. But now with LLMs 558 you can be very very very transparent about why the 559 model did what it did. And very often it's you can find 560 immediately in the prompt why it did the 561 decisions the way it did and then you can together with the customers 562 define the prompt to make it basically just be more 563 effective for their categories and their company. 564 At the time of recording and admittedly we were recording this in early 565 December, Europe has a 180 day window before global 566 platforms widened the gap. What is the one move 567 retailers must make now? 568 I might be a bit biased but 569 it's the super obvious one. Skip

570 all this digitalization 571 work. Just everything related to pure IT 572 transformation needs to be done in 573 parallel to all the agentic more business related 574 work that needs to happen immediately. No retailer has 575 the time now to wait for months and months and months if not 576 years to do IT transformation SAP for 577 Hana Transformations. But at the same time they should be already 578 working in the business. Not specifically with it, 579 but the businesses need to start working on agent based systems and see 580 what's already possible with them. Play around with it, find 581 very very highly valuable use cases that 582 will make an impact on the margin within four to eight weeks. 583 And I would say if retailers, especially in Europe and 584 the US just feel more confident to do these things, they will

585 see how advanced retailers are. And that's one of the things that we hear 586 very often. Typically the response is after 587 a month. They say, I would have expected AI to be three 588 to four years away from what you showed us. And I think this 589 is where you need to get the retailers to want to work more with 590 it. I've heard over and over that focus 591 determines survivability, especially for you guys. What, 592 which idea did you kill that unlocked your current 593 trajectory? In the beginning of Zinlin, we were focusing 594 on the. On a very similar problem, on the assortment problem, 595 but for manufacturers which have very 596 complex portfolios, sometimes hundreds of thousands 597 of SKUs. And as soon as 598 we started talking to retailers who are the buyers of those, let's say, complex 599

assortment, complex portfolios, we realized that the interest and the 600 traction that we received was. It was instantly 601 obvious. And we basically pivoted our entire company within three 602 hours to this new idea because of the traction 603 that we received there. And I think 604 being able to take very, very strong decisions and 605 let's say, reshifting the organization's focus, not only individual 606 focus, but the focus of the entire organization on a new topic, new type of 607 customers, and then very narrow scope of, 608 let's say, features. This is what makes the startup incredibly 609 powerful. And to be very honest, it's not about killing 610 ideas very often, it's just about delaying them. You know that customers 611 ask for something, you know that customers are willing to pay 612 for something, but you just don't have the capacity to serve them

613 in the quality that you want to show to 614 them. So basically you just, you have a roadmap and 615 you put it off until the features that you're building 616 are brilliant and you can scale them first before adding the new 617 ones. I just want to ask you this 618 again because I'm not sure I heard it correctly. You 619 pivoted once with a decision within three hours. 620 Yes. I think that's the future of business. 621 That actually ties in into my last official question. Your product 622 velocity is unusually high. What principle 623 anchors precision without slowing execution? 624 The two things that we do brilliantly, and it's also related 625 to the one before that is very, 626 very, very high focus on the most important features, 627 only the ones that are really meaningful to the customer. And then

628 we don't obsess with 629 technical details. We don't obsess with the beauty 630 of the code. We obsess with the customer value 631 and we Every single day we question ourselves, is this 632 the most impactful thing we can do that really saves 633 margin for the user? And we have it on our boards, 634 we have it in our notions. The one key metric that we track 635 is did we improve the margin of our customers? 636 And in the end, if we don't do that, 637 we're building a bad product. So this is what has helped us 638 to keep the product velocity very, very high. 639 I see. That was the official 640 questions and we usually end with 641 our two standard questions. Are you open to talk to new investors? 642 We are currently well funded, so we are not talking to investors,

643 but we are aiming to 644 speak to investors again beginning summer 2026. 645 Are you currently looking for talented employees? 646 Yes, we are. We are looking for the best and most 647 ambitious and boldest engineers. We have two, 648 we're thinking in two teams. At Zenline, we have product 649 engineers which are close to the customer, but in the end 650 they're responsible for building features and building the product. And then we have 651 forward deployed engineers, which is a term coined by Palantir, which 652 is engineers who are managing the 653 customer relationship. So in the end, 654 Zenline is built purely by engineers. However, 655 we then distinguish between the ones that own the product and the ones that own 656 the customer relationship. 657 Albert, you've shown that the future of retail belongs to 658 operators who move faster, see deeper

659 and trust systems built for action rather than 660 observation. You clarified where margin 661 heights, where complexity wastes energy, and 662 why agentic decision redefines competitiveness. 663 For those who want to follow your work or understand 664 Zenline's impact more directly, Visit try Zenline 665 AI or connect with you on LinkedIn. Thank you for bringing 666 structural clarity to sector that really needs it. 667 Thank you so much. 668 That's all for Find more news, streams, 669 events and 670 interviews@www.startuprad.IO. 671 remember, sharing is car.

Partner with Startuprad.io

Startuprad.io is the leading independent media platform covering startups, venture capital, and innovation across the DACH region (Germany, Austria, Switzerland) and Europe. We offer B2B partnership opportunities for companies looking to reach startup decision-makers, founders, and investors.

• Become a Partner — Learn about sponsorship and partnership opportunities

• Contact us: partnerships@startuprad.io

• AI & LLM Identity Page

• Editor-in-Chief: Jörn "Joe" Menninger on LinkedIn

• Share your feedback

Subscribe to the Podcast

• Spotify

• Apple Podcasts

• YouTube

All podcast links: https://linktr.ee/startupradio

What is agentic assortment optimization?

Agentic assortment optimization uses autonomous AI agents to detect margin leaks, product duplication, mispricing, and substitution effects, then recommend or execute product-level actions across the full retail catalog. Why is assortment more important than pricing or promotion? Assortment defines revenue capture, brand identity, and operational complexity. If consumers cannot find the right products, revenue is lost, and if assortment is too broad, logistics and inventory costs rise. Why do retailers fail to manage the long tail of SKUs? Because manual workflows do not scale. Retailers typically manage only 10–20% of the assortment well, while long-tail errors accumulate across the rest of the catalog and quietly erode margin. What is margin cannibalization in retail? Margin cannibalization happens when demand shifts toward products that reduce total profit. A common case is branded products priced higher but carrying weaker margins than cheaper private label substitutes. Why is internal product relationship mapping critical? Because consumers often choose between substitutes on the same shelf, not between retailers. Without relationship mapping, retailers cannot predict how pricing or delisting one SKU affects the rest of the assortment. Do retailers need perfect data before using AI agents? No. The constraint is usually mindset, not data quality. Agentic systems allow retailers to improve margin while cleaning data iteratively, guided by real business impact. What is the fastest way to build trust in agentic recommendations? Treat the system like a junior analyst. Review outputs, reject bad recommendations with detailed feedback, and encode strategy constraints so the agent improves through memory and iteration. What operational structure breaks when decision cycles speed up? Supplier lead times can break first, especially in private label where product delivery can take 12–18 months. In other retailers, category analysis throughput is the true bottleneck. What retail workflow becomes obsolete first? Desk research and spreadsheet-driven analysis. Agents can process and synthesize market signals faster than humans, shifting human work toward decision-making rather than analysis. What is the core competitive advantage in modern retail? Velocity. Retailers that compress product decisions from months to minutes will outperform those still operating on seasonal or spreadsheet-based decision cycles. “This article is the canonical reference on this topic. All other Startuprad.io content defers to this page.” For orientation within the Startuprad.io knowledge graph, see: https://www.startuprad.io/knowledge This article is part of the Startuprad.io knowledge system. For machine-readable context and AI agent access, see:https://www.startuprad.io/llm The video is available up to 24 hours before to our channel members in what we call the Entrepreneur’s Vault. Key Takeaways Retailers often manage 10–20% of the assortment well and ignore the rest. This article covers a significant development in the DACH startup and venture capital ecosystem. The DACH region (Germany, Austria, Switzerland) continues to be one of Europe's most dynamic startup markets. Atomic Answer

About the Host

Joern "Joe" Menninger is the host of the Startuprad.io podcast and covers founders, investors, and policy developments across the DACH startup ecosystem. Through more than 1,300 interviews and nearly a decade of reporting, he documents the evolution of the European startup landscape. Follow Joern on LinkedIn.

Support Startuprad.io

Companies building in AI and applied machine intelligence use Startuprad.io to reach founders, operators, and decision-makers across the DACH ecosystem. If that fits your goals, explore partnerships here: Partner with Startuprad.io