What changes when a retailer can make a separate decision for every customer instead of placing thousands of people into the same campaign segment?
In this episode of Startup Builders & Backers, I speak with Ilyas Kurklu, CEO and co-founder of Replenit, about the company’s attempt to move retail CRM beyond fixed audiences, campaign calendars and generic promotions. Replenit calls its product Maestro, an AI CRM Manager designed to decide what should happen next for each customer and pass that decision to the retailer’s existing channel tools.
The distinction between prediction and decision runs through our conversation. A propensity score can estimate what someone might do, but a retailer still has to decide whether to act, which product to recommend, when to communicate and whether a discount is commercially sensible. Ilyas argues that this reasoning should be readable by a human CRM manager. For every decision, the team should be able to inspect why this customer was selected, why the timing was chosen and why a particular product was recommended.
That raises the question every enterprise AI project eventually faces: how do you prove the system created value rather than claiming credit for a purchase that would have happened anyway? Ilyas recommends measuring incremental margin per customer against a randomized holdout. He is skeptical of attributed revenue, influenced revenue and open rates when they do not reveal whether the business made additional money. We also discuss Replenit’s focus on cost per decision and the processing expense that comes with adding further data to an AI model.
Integration is another part of the story. Retailers may already use a customer data platform, a customer experience platform, analytics products and multiple communication channels. Replacing that stack can bring cost, delay, security reviews and regulatory work. Ilyas explains why Replenit has been designed as a reasoning and orchestration layer that works with the systems a retailer already owns.
The conversation also considers customer interest and commercial pressure. A retailer may have contractual brand rules, margin targets and limits on which products can be recommended together. Maestro records brand, product and customer memory, while the retailer sets the rules that govern its output. The human CRM manager then becomes responsible for reviewing outcomes, setting boundaries and questioning the system’s decisions.
For founders and backers, Ilyas offers a refreshingly practical reminder. Building an AI company still means doing the ordinary work of running a business, proving that the buyer can make or save money, answering why an incumbent cannot easily copy the idea and reducing the risk that stops an enterprise customer from saying yes.
Would you trust an AI CRM manager to make individual customer decisions, and what evidence or controls would you require before giving it that responsibility? Share your thoughts with me at Tech Talks Network.
[00:00:00] It's not your fault that your agentic AI systems are acting outside of compliance. There are just too many datasets to guardrail them all. But with Denodo, you can now organize your hundreds of data sources into one layer and govern your agents with a single approach. Try it now with Denodo by visiting denodo.com to learn more.
[00:00:26] What happens when a retailer stops treating thousands of customers as a segment and starts deciding what each person genuinely needs next? Well, today I'm going to be joined by the CEO and co-founder of a company called Replenit. We're going to examine how AI could change customer relationship management from scheduled campaigns
[00:00:53] and into continuous individual decisioning. And I also want to learn more about Maestro, which is Replenit's AI CRM manager, and find out what separates prediction from a commercial decision. And he will also explain why every recommendation needs plain language reasoning, measurable outcomes and boundaries that a CRM team can inspect.
[00:01:19] And for founders listening, there's another useful thread running through this conversation. And that is a clever model still has to solve a very real business problem. Work with systems that customers already own and earn its place through incremental margin. Thankfully, nobody needs another dashboard gathering digital dust. And on that note, let me officially introduce you to my guest now. So thank you for joining me on the show today.
[00:01:49] Can you tell everyone listening a little about who you are and what you do? Okay. Thank you very much for having me, Neil. I'm very happy to be here today. My name is Elias, CEO and co-founder of Replenit. We build AI workforce for retailers. And the first one is Maestro. And then why we built this product? Because like me and the founding team has a really background, has a long background, but they work with the retailers.
[00:02:17] What I can say over here that we built Unicorn from scratch together with my co-founders. And then this was the product we dream of selling it to our customers. And then after some certain time having an experience in the martech field, and we realized that together with AI, now it is possible to do it in a feasible pricing and in a scale. So we built Replenit like around two years ago in Warsaw with six co-founders.
[00:02:47] Today, we are 21 people. And then we have more than like 30, around 40 customers all around the world in six continents, let me call it that way. We call ourselves a startup. Sun never sets. But things are looking great for us. And we are excited as well. Thank you very much. So you just need a presence in Antarctica and you've got all continents covered, right?
[00:03:13] If there was a retail existence in Antarctica, definitely, definitely we will try that. And obviously, I know it's an incredibly exciting week for you guys, a very busy week. You're talking at the event and things. There will be a lot of announcements coming out of that. And this episode will be live after the event. So I will include links to a lot of the news that comes out there. So I urge listeners to check that out because there is some exciting stuff in there. But a question I've got to ask you, because retailers already have prediction tools.
[00:03:41] So what is the difference between predicting what a customer might do and deciding the next action a business should take? Okay, definitely. Actually, amazing question. It's a challenging one. If I can answer you, like the prediction is a probability, but a decision is a commitment that we give about the consequences. If you have Maestro, what would change in your life? And then the retailers are drowning in the first.
[00:04:08] They know that they need to communicate with customers. And then they always say, okay, I like to purchase. And then what they do, I like to purchase more than 70%, filter out the people, put 40,000 people in a group, and then they send them, here is your 10 pounds, come and buy it. And we realize that it doesn't work. In general, like if you go to the retailers and then ask about their performance, there
[00:04:35] are some cases they have, oh, we have more than 2% conversation. They brag about it, but like the, I read this like that. Oh, so 98% is not converting. So we need to ask the question, why? And at Replenit, you describe a reasoning layer that accounts for context and intent. So what evidence tells a retailer that this reasoning is reliable and what guardrails prevent maybe overconfident decisions like we see with AI? No, definitely.
[00:05:05] I think if you are hiring a workforce from an AI solution, retailers should never trust any reasoning that they are receiving. They should trust the measurement that they are providing, and they should be able to read every decision like they would read a colleague's recommendation. And then to do so, we do three kinds of evaluation.
[00:05:31] Very first one, like the against the control group, but like it is not for the workflow, not for the model accuracy on the slide too. And second, every decision our system writes in a plain language. Why this customer? Why now? And why this product? So a CRM manager, actual CRM manager can come and audit it, whether selection criteria,
[00:05:57] who, what, when, and why has been sustained properly. And the third, like the reasoning layer that we shipped this March, fine-tuned, and then they developed over the Google Cloud. We can do it on a volume. So like when you produce a reasoning and then the decision for a retailer, of course, what is the cost of this decision? And then what is the return of every decision?
[00:06:25] So like the cost per decision, actually, our new metric that we are tracking, and then the retailers are taken into account. Okay, Replenit has given a million decision for us. So we get this return, and then our decisioning done nonstop, 724. So retailers shouldn't think of whether we execute the communication.
[00:06:49] No, they need to control whether those decisions were right, and the consumers get what they needed by engaging with this communication layer of the retailer. And Maestro is designed to work with existing data and orchestration systems. So how do you add an intelligence layer without creating another silo or requiring a costly replacement program? Tell me a little bit more about how you do that.
[00:07:17] So like throughout our experience, like what we have seen, retailers, when they have a new product, things are quite looking good in the presentation. You know, it's quite easy to show. If you get this platform, it's going to be a rainbow and sunshine. But when it comes to reality, you've got to integrate this, change, that, check, this, they're part of it, legality, security, GDPR, and so on and so forth.
[00:07:46] So like they're fed up with this. Okay, I'm going to get this. What is the switch cost? I'll imagine that I switch it. What is the timeline for ramping up to get, to catch my past performance? Then I will improve it. And we realize that like if you're going to create a product for a retailer, it's not going to be, it's not going to be worth it. They're going to say, okay, guys, like the idea is good. It doesn't make any sense. I always use this analogy. I'm going to do that again.
[00:08:15] So like if I go to Volkswagen CEO and I tell him that, look, man, we have a device. We're going to plug it into the 2000 model Volkswagen. It's going to be Tesla. Tesla autonomous driving. It is not possible for the cars. That's why like the electric car companies are on the rise. But it's possible for MarTech. It's possible for retailers. The current tech stack has the connectors, but they are not connected to each other. So what we built, we built Maestro on top of it.
[00:08:45] And then the Maestro orchestrates that tech stack. That's why we put the name of Maestro. So like the retailers actually only check the outcome of the tech stack plus Maestro's results. And before you join me on the podcast today, I was reading about your work with Bico that uses connected product signals to shape post-purchase journey. So tell me more about that.
[00:09:10] And what IoT data can reveal that a conventional campaign or customer profile. I am not 100% sure that I can share like about IoT data I can do or not. Because like the Bico is extremely important partner for us. And I don't want to get into that detail. Yeah. I would like to roll back the question. Why the data is important and how important why you are utilizing with AI? Like you have a lot of data.
[00:09:40] Most of the retailers have a lot of data too. So they have different silos. So like it is important, but you're going to give it to the LLMs and then they get an answer. And what will be the size of the data? So like if you, can you do this? Like you can get all of the data that you have, push it to the cloud and then get a response. You can take it. It's going to be super expensive for you.
[00:10:04] So like when you would like to do a testing and then the targeting, you're going to get the data. Okay. And then you need to put it in an order and structure. And then you slowly start feeding it to the LLM and then fine tune it and then train it. Then after this, start predicting on top of it. So what we do on our end, our like the AI researchers and then the head of AI is working on it. Okay.
[00:10:33] We have built the first predictions. How we can add further data so we can improve our prediction quality and engagement quality. So that's why like the, I cannot tell the detail of the IoT details, but it is extremely important when you add a new data. Did it impact the end result, the consequences of the decision? Is it positive, negative?
[00:10:59] Does it worth adding and then the making a prediction because every new data point that you add to the LLM, it's going to create a processing cost on your size. So if you are a business, you need to think that way too. So what is the cost of doing so? So if I utilize it, can I make a profit on top of that? So that's why every data is important to understand whether for that prediction specifically, are you going to use or not?
[00:11:28] This doesn't mean that, for example, we are working on the communication layer. Maybe IoT data right now is not important, but you will make other decisions for that company too. So that data could be relevant for them. So it is not useless. For that specific given prediction, it's not working. It doesn't give any clue. But like for other use cases, it can be useful. That's why like the companies like Beko, Estelada, L'Occitane chooses us,
[00:11:57] not for the, you know, services and then the decisions that we provide them, but also they see that like the potential in the future, how we can enhance the understanding of those customers so we can provide a better experience of communication so we can increase customer lifetime value. So this is our direction. That's why like I cannot tell you that this data impacts positively, that data impacts negatively, but definitely you got to test, try, find out the outcome,
[00:12:26] and then take it to the next level. And of course, when we're talking around any new tech project or AI project in particular, there's a strong focus on ROI now. And one of the things that stood out to me about Replenit is you offer a strong return on investment commitment. So I've got to ask, what kind of metric should a retailer use to prove that their system created incremented value rather than just taking credit for a sale that would have happened anyway?
[00:12:54] Because the old IT mantra, you can only improve what you measure has never been more appropriate. But tell me a little bit more about the metrics that they should be measuring. Definitely. Definitely. We have a strong return on investment commitment. Definitely. Because incremental margin per customer against a randomized holdout, like the spray and pray, like they should be valued and measured. Not attributed revenue, not revenue influenced,
[00:13:24] or not a vanity metrics that we call it. Oh, open rates has been increased. Okay. What is the impact on revenue? Because I'm running a business on my end. So like what we are doing on our side, first guys, you are not going to have the switch cost. Like what I mean with switch cost, we are adding our system, and then this system makes it autonomous. Maestro runs the whole tech step of yours.
[00:13:53] And on top of that, after that, when it is ready, once your data is with us, you are ready for prediction. The other day, next day, 24 hours. So you're not going to wait for it. So like the retailers don't think, oh, this is a black bus. Let the algorithm learn for two months. And then, yeah, it's gradually increasing. No, because like the life cycle is clear for us. What happened with that customer? And then we are pretty much concrete. When will be the replenishment purchases?
[00:14:23] What are the potential cross-sells? What is the substitutes? Who are the silent customer? Maybe Neil has a profile of high-end customers, so he doesn't need any discounts, but like he has been impacted by fear of missing out. So new arrivals would be recommended by Ilya Cezar. You know, I'm not much of a, like the fun of like the new arrivals, but like the mostly motivated by a discount. So we can protect the margins and we can do it in a scale on an individual basis.
[00:14:49] So like that's why when you look at the return of investment, think of what are the additional capabilities Replanet is adding? Because currently no retailer without Replanet can do individual engagements. And we really personalize for every customer of those. So that's why they send millions of emails. And then they have lots of unsubscriptions
[00:15:18] after sending each email, but batch and blast. And then this creates, this decreases the value of the channel. How many emails you receive, Neil? You tell me like there's a promotion and how many of them you open and then click and make a purchase. This is so desperate. This is so desperate. I don't think that such enterprise companies are doing such a desperate moves to get a couple of boxes from the customers. And you know what? Like when you send the customer some personalization,
[00:15:48] put 40,000 people in a group and then the show, this is last three items and then create that urgency. 53% of them are not happy what they purchased. So because you're guiding your customer to the wrong product. And imagine that you're going in an offline store and saying the sales clerk, what you need and what you have purchased, he knows and he's selling you something that you do not need. I don't think that you will go there. I don't think that you will go talk to their sales clerk.
[00:16:18] So that is the problem of retailers. And then communication is the most important thing. Years ago, I don't want to name the brand, but it was a really good big brand like enterprise brand. And then the executive told me, Elias, you know what we are proud of? We are proud of the engagements that we don't send to our customers because we don't want to send it just because we have a discount. We don't want to send it to everyone. We want to find out who wants and who is motivated because of that
[00:16:46] and who needs that product too. Because if you sell something excessive to me, I'm not happy with that. I will see that that product I don't need is staying over there and it's going to annoy me day to day whenever I see that too. So all experience by the time after the lockdown boom is not good anymore. We want to change that. And then there so far we are doing good. And as you can see from return of investments and then the customers, brands, less than two years old company precedes stage and where we are able to
[00:17:16] onboard really global enterprise brands. And one of the things I wanted to bring up as well today is we will have a lot of startup builders and backers listening who might be inspired by the journey that you've been on. So what have fundraising, enterprise sales and category creation, all these things you've been involved in throughout your career, what have they taught you that maybe another founder building an AI company could be listening that they could maybe start applying this quarter and take away? I will start with the,
[00:17:46] it's a business at the end of the day, Neil. And then every business has some standard procedures like accounting, finance, office management and so on and so forth. So what I see with the new founders, they want to do only the fun parts. They're executing, like delivering the products, selling it, dealing with the customer, doing the marketing of it, but they don't want to do the repetitive tasks. For example, they don't pay attention to the human resources. So very first one, like you are opening up
[00:18:16] a business, it doesn't have, it has a traditional requirements. Very first one, you need to put it into your account. So you need to be agreeing on that you are going to do the unnecessary, you see unnecessary tasks, but it's a mandatory thing. Then the second part of it, it's a business. It needs to make money and the product should enable the buyer to earn more money. So that's why I like our obsession with Google Cloud and then the decreasing the cost part
[00:18:45] were extremely important because we knew that like some platforms or some of them will try to move to the decisioning and then they will explain, oh, we did the partnership with cloud and then we built the cloud force, okay, and then they give that access to cloud and then the millions of data poured in the cloud and then they will enter the one prompt. I'll see your bill. So like, that is not something business desire for. They want something,
[00:19:15] they want to invest money on something, generates revenue, profit in return. So you got to focus on that as a leader of your company, a founder of your company. Okay, I'm going to produce a product. How much it's going to generate? Why we are giving a commitment? Because we are confident that it's going to create the value. It's not a marketing gimmick. It's not like the vanity metrics. It is the end result. So founders, startup founders should go
[00:19:45] and justify it. And the third one, if they want to be invested in, you need to show that there is a potential. There is a need. Why others shouldn't do that? For example, we received a lot of questions. We have MarTech background and then they always ask why Salesforce shouldn't do it. And they need to find an answer for that. You know, I cannot explain it over here for you right now. But like, they need to find, even they don't have a competitor. Who might be their competitor and what
[00:20:15] they can do? You're really pretty much, you know, kind of okay with doing the, you know, like some repetitive tasks. And then if you really focus on the profitability, what do you provide to your customers? And then the third one, you need to show that there is a potential market and then they light at the end of the tunnel, like first, two, three paying customers. Reasonable investors who actually really believe that like the smart money is supporting, you know,
[00:20:44] like the business-minded founders will end up with funding. So ours were like that and then we end up with great founders like Moose Capital, Matish Visi, Waspoint, Eleven Labs' CEO, Matish Danievsker is our angel investor too. There are a lot of names, I don't want to name it, but when you just advise these three pillars in a proper way, they are really supporting you. decision engines are also something that can recommend the action most likely to improve
[00:21:13] the metric, but commercial incentives often conflict with the customer's actual interest. So how can retailers set boundaries around those recommendations just to ensure that they get that balance right? I don't want to name brand, but like the, since we have this, like the background and then the experience, we knew that like the, some retailers are selling brand A and brand A is Monday said, you cannot recommend
[00:21:43] my buyers from a product B or any other product or a competitor specifically. So like the, what we are doing at this point, like with Maestro, when Maestro gets into control, it's the driver anymore. So a CRM manager becomes the co-pilot. So like when we are onboarding a customer, they have also interface that, okay, brand A cannot be recommended to brand B
[00:22:12] and then the, you can, you need to take into account the coloring. You need to upload what is the, like the, what was the good messaging, high open rate messages with the customers, with the emails. So we get all of that and we also create a brand memory, product memory and a customer memory and then this brand memory is being managed by the CRM manager because he or she knows that like
[00:22:41] what is the company goal? For example, increasing margin, then putting the setup, sending the guardrails and moving forward. And after that, one of the most important thing, okay, it is autonomously working and autonomously learning from the end result. So when we make the predictions and in the decision, not only the predictions and the decision, we also look at the decision result, then start learning from that and fine tune
[00:23:11] our next predictions for the given customer only, not 40,000 people. So enterprise customers often want proof before sharing enough data to produce it. So I'm curious, the very early days are how do you overcome that early trust problem and what can other founders learn from that process? Because you have a great product, you just need to get that trust so you can get it over the line. Okay. So that's a very good question too. That was also one of the challenges that we learned in the past. And then the redesign
[00:23:41] our product. If enterprise customer, they have a hesitation, okay, instead of Neil Hooch, they are telling me that one, two, three, make a prediction. This is the past and then I go back, one, two, three, does this and then the potential, the decision of a purchase will be this. They are matching on there and how to communicate with that given customer or we have like we are utilizing for the secure, we are following the necessary GDPR
[00:24:10] and AI regulations for example for every given decision that we do for individual level who, why, and why. We keep record of it so we can audit it. That's why that's why like when you do so and then they see the need and then they start testing and then initially right now we don't even offer testing most to the most of the products because like I already have a lot of case studies which is working in different you know like the platforms and different words
[00:24:39] so like if there is a concern they can leave the contracts anytime because we don't do dot-com balloon dot-com boom type of contracts you need to sign five years and then you cannot exit you have to pay whether it works or not it's not like that if it doesn't work if it doesn't deliver the commitment that we give to the customer we don't remorse them so simple as that if it doesn't work they can leave the contract anytime and no one is leaving because
[00:25:08] they want a working product they want an end result and they want to control this so that's why I like to what was the challenges of the past challenges of ours which said that okay we are going to design a product not to solve our problems solve the really customers problems and since me and my founding team mostly focused in the past of the customers we know end to end what they are looking for well thank you so much
[00:25:37] for sharing your story with me today for everybody listening I know it's a big week for you guys with the AI in Retail Summit you'll be working with Google on that I'll include links to all the announcements that come out of that event and anyone wanted to just find out more about Replanet and everything that you're doing there or connect with you and your team where would you like them to go? Oh definitely very first one I can definitely guide them to our website we have really good blog posts that our AI engineers and AI researchers
[00:26:07] are sharing on a regular basis so definitely subscribe our newsletter on our website we have a LinkedIn page we announce I have an open public LinkedIn page anyone can reach out to me by messaging directly when I see the really good messages for example we are 21 people but three of them are through direct messaging to me so like when we find the talent we respond we are not like that and then the last one
[00:26:36] I have a Medium page I'm sharing my thoughts on Medium and then it can be it can just like write my name and surname and then find it on Medium but I will be sharing with you so you can place it over there too and in the case anyone would like to reach out please feel free to send me any kind of messages I'll definitely read them all and then I find what is in it for me and what is in it for you clearly I'll respond
[00:27:06] them all awesome absolutely love it a real pleasure talking with you today we hear a lot around technology and AI and especially dealing with customers etc but what I love about what you're doing here is we're not talking about replacing anything with AI we're talking about using AI leveraging it to manage every customer relationship individually continuously intelligently and commercially it seems like everyone's a winner here so a big thank you for sharing thank you very much thank you very much
[00:27:36] for your time it was a pleasure having this session with you I think one of the biggest takeaways from our conversation today is that retail AI should be judged by the decisions that it improves not the volume of messages that it can send and my guest made a strong case for randomised holdouts incremental margin and an audit trail that explains who received a recommendation why it was made and why the timing made sense and there is
[00:28:06] also a useful lesson here for founders listening enterprise buyers seldom want another replacement program they're motivated by evidence that a new capability can fit in with their current stack respect their rules and produce a result that they can measure so a big thank you to him for sharing the thinking behind replenet and maestro and given founders listening some real gold takeaways there and you can learn more on
[00:28:35] the replenet's website newsletter linkedin page and my guest profile i'll include links to everything in the show notes over at techtalksnetwork.com but over to you what would make you trust an ai system to manage individual customer decisions again techtalksnetwork.com that's all things me you'll find everything there along with 4000 interviews all the tech events that i will be attending this year if you want to meet me in person
[00:29:04] but that's it for now so thanks for listening as always bye for now you