Justin Leibow: Hello and welcome to Product Talk. I'm your host today, Justin Leibow, Chief Platform Officer within the Financial Services organization at EY. With me today I have Mike Le, founder and CEO of Conative AI, an AI-first company that went from startup to award-winning over the past three years, including the U.S. Agency Award for Best Use of AI in a Client Campaign in 2025. In this episode we'll discuss two things: first, how AI is transforming how companies look at marketing, and specifically how Conative AI's approach bridges operations, sales, and marketing to create better go-to-market decisions and overall value. Second, how AI is an iterative process, and how failure isn't always a bad thing — there are lessons learned along the way. Welcome, Mike.
Mike Le: Thank you, Justin. Very happy to be here.
Justin Leibow: I love starting every episode by learning about the person I'm talking to. Your journey — and the risk you took to get where you are — is phenomenal. I'd love to hear your background and how you got here.
Mike Le: I'm Mike Le. I grew up in Vietnam, came to New York in 2005 for my master's degree at NYU, and randomly met my co-founder Jane on the F train. We talked, became friends, and realized we'd both read the same book on immigrant entrepreneurship — Self-Made in America. So we said, let's start a company together, and started CB/I Digital in 2007. Eighteen years later, we've gone through the early days of SEO, early Google Ads, early Meta Ads, all the way to where we are now in the age of AI. CB/I Digital is a performance marketing agency that scaled to over 100 people across New York and Vietnam, working with ecommerce brands and enterprise clients. About three years ago, working with so many brands, I saw a common problem: the disconnect between marketing and inventory planning creates a lot of operational inefficiency. So we built Conative AI, which uses deep learning and agentic technology to enhance forecast accuracy, so brands can reduce overstock and understock and improve cash flow. That's where we are today.
Justin Leibow: Before we get into AI — the fact that you met your co-founder on the subway is an accomplishment in itself; I had my earbuds in the whole time I lived in New York. But let's get into it — leveraging AI early, when it wasn't as trusted as it is today, was a critical step in your journey. How did you make that shift?
Mike Le: I made my first AI hire in 2019 — three years before ChatGPT made AI a hot topic everyone paid attention to. I'd been running my agency for some time and was thinking about the future of the company. I was weighing three areas: internet of things, blockchain, and AI. I decided AI fascinated me the most. I built a team and started researching the space, went through a few failed product ideas, until I landed on the concept that became Conative AI — something I felt I could actually scale. By 2022, we'd already been building on AI for a while, so by the time everyone else started paying attention, we already had a technology head start that helped push the product forward.
Justin Leibow: Were there a lot of challenges early on, given AI was still an immature, untrusted technology at the time?
Mike Le: The biggest challenge at the beginning was that AI was changing very fast. Back then, we weren't talking about ChatGPT and general-purpose LLMs — every problem had its own dedicated, specialized model or model family. For Conative AI, the idea was to use AI and deep learning to enhance demand forecasting, and the key was choosing the right models to build, test, and refine to provide accurate forecasts with lower risk. So the challenge was testing a series of models and architectures to find an approach that could keep evolving consistently over time — because even a winning model today could get upended by something new the next month.
Justin Leibow: That's probably even more challenging today, given the pace things move now — a month feels like light years. How have you continued adapting, testing quickly, and avoiding getting locked into one model?
Mike Le: Let me back up to why Conative AI exists in the first place, because it links to the technology choices we made. As a marketing agency owner for many years, the real problem I saw was brands running into trouble with marketing campaigns because bestsellers were out of stock, or they had too much stock and tried to use marketing to move overstock nobody wanted — which tanks performance. Digging deeper, I saw a lot of brands relying on spreadsheets, gut feeling, and rule-based systems for inventory planning and buying — a very cumbersome process. So the question became: how do we enhance that with AI? Interestingly, this is a problem that, to this day, general-purpose LLMs like ChatGPT don't solve well — it requires specialized forecasting models.A few things matter here: you need a lot of good data from multiple sources, cleaned and fed into the right AI models, and then you measure the accuracy of the output to make sure the error rate is at a level brands can actually use to improve cash flow. There are roughly four tiers of forecasting solutions: rule-based systems, which is what most people use; basic mathematical/statistical models; early machine learning; and then deep learning, which stacks multiple layers of machine learning together for better results. The advantage of deep learning over earlier approaches is that it can take in far more input variables and data streams — rule-based and math-based approaches can only account for a few factors, while deep learning can factor in many more, giving the model a much larger context to produce a better-quality forecast.Testing multiple models in parallel was important early on to find the one with the flexibility to adjust alongside the data stream and produce more reliable results. But eventually you reach a point where you're no longer swapping out entire models — you're fine-tuning layers of an existing model based on your specific data. So today we're no longer constantly changing our core model architecture; we run a few lines of testing in parallel, but it's a much more stable foundation than it was in the early days.
Justin Leibow: As you built that foundation, I imagine data itself — bringing together information across different parts of a company — was a major challenge. How did you see that play out with clients?
Mike Le: That's honestly the biggest challenge I've encountered — the data problem is much bigger than the AI problem. When we built our MVP for Conative AI, it took my team about three months to get something working — collecting data, running forecasts. But once we onboarded a few brands, the MVP started collapsing, because making the data work flawlessly in real time is genuinely hard. I realized data is a massive problem in this industry — it's part of why ERPs get so complex. For Conative, we pull data from Shopify, Google Ads, Meta Ads, Google Analytics, and ERP systems for each brand, and what we found is that for many brands, the data is messy — wrong, missing, or conflicting between sources. The real challenge is you can't feed bad data into good AI models and expect good output. I ended up spending a full year re-engineering our entire data engine — building an enterprise-grade system that could pull data from all these sources in near real time and dynamically map it into a standardized data model, so we could onboard data from any brand, in any industry, from any source, consistently. That year of work on data quality is what I'd credit most for our forecasting accuracy today.
Justin Leibow: That makes total sense — a lot of companies jump into "AI first" thinking AI will solve everything, without focusing on data cleanliness and organization. If you start an AI journey without understanding your data, you're going to fail. You built the app, thought you were ready, and then ran into a full year of data work.
Mike Le: It also depends on the type of data. If you're training an LLM, you're mostly working with documents and images — technology today handles that reasonably well. For Conative, we're dealing with live sales, marketing, and inventory data streams that move fast and at high volume — that requires a much more complex, robust data architecture to handle correctly.
Justin Leibow: As you looked at these data streams and how companies are siloed — operations, inventory, marketing rarely talk to each other even though they clearly influence one another — did you run into that challenge with clients?
Mike Le: Yes — even at the demo or onboarding stage, we always share a specific story with brands about why this matters. We worked with a brand where about 80% of revenue came from just 15 top SKUs. What's crazy is that at any given moment, half of those SKUs were out of stock — meaning they were losing roughly half their potential revenue at any given time, and nobody realized it, because the inventory buying team wasn't looking at marketing, and marketing had no visibility into inventory. When we streamlined marketing and inventory together for that brand, sales lifted 40%. That story helps brands see the problem and connect the dots. There's real effort involved in getting inventory and marketing teams on the same page — Conative's dashboard becomes the tool that breaks down that silo and lets both sides engage with the same data, making better, more informed decisions together.
Justin Leibow: Knowledge is power, as they say — and giving teams a shared dashboard means they can actually act on that knowledge, adjusting forecasting, inventory, and marketing based on a shared source of truth. On the people side — has AI felt like a threat to some of the people whose jobs are literally running inventory and the books?
Mike Le: It's a real challenge, and it took time to find the right way to communicate with users so they feel safe rather than threatened. Inventory planners and buyers are under real pressure — a mistake can cost the brand real money, so they're naturally protective of their process. At first, people don't trust AI — they'll say, "I have 20 years of experience, how could AI know this?" So the first thing we usually do is let them run their own method in parallel with AI for two to four weeks, comparing outputs and digging into any discrepancies. What we've found is that after about two weeks, most people realize AI is already in a very similar range to their own forecast — without the 20–30 hours of manual effort it took them to get there. When there is a discrepancy, digging deeper usually reveals a good reason behind the AI's number. After that, most people voluntarily let go of their old process, because it's no longer necessary. And because we build a risk-control layer into the mechanism, even if a brand blindly trusts the AI's output, we've already accounted for that risk in advance so it doesn't expose them to major downside. Combining that transparency with letting people see the comparison for themselves is how we build trust. I also always remind people: you're very capable at your job, but it's genuinely hard work — there simply wasn't a tool in the market that could help you do it more effectively until now.The shift to AI agents in 2025 changed this conversation even more. Before, even in Conative's early days, an inventory manager had to dig through a dashboard full of numbers, graphs, and visuals to find insights. With agents, it's like a chatbot that knows your data — the data can talk to you. You just ask, and the agent does the job. That lets you focus purely on the decision, because the one thing AI can't do is make the decision or take responsibility for you — if there's a mistake, it's still the person who owns it. You always want to remain the leader of the AI, not the other way around. That approach has let inventory teams using Conative simply enjoy the upside — the tool speeds up their world without threatening their role.
Justin Leibow: That fear of AI taking someone's job is real, but it seems like it's really a way to enhance the job — freeing people to focus on critical thinking instead of the manual grind to get there.
Mike Le: Right. In our journey, there are really two areas of AI we've developed: AI forecasting, which is about building trust in the forecast itself, and AI agents, which is about how you interact with AI to boost your productivity and get your role's work done with ease. One is about trust, the other is about efficiency — but on the agent side, I think the train has left the station. It's no longer a question of whether AI will take your job — you simply need to get on board and become a master of it. If your competitors are using it and you're still doing inventory planning and buying on spreadsheets while everyone else has moved to self-driving cars, you're going to fall behind. I think the pace of AI progress has made people much more aware of that than they were even six months or a year ago.
Justin Leibow: Last question — if you look into your crystal ball, where do you see this in five years? What's changed?
Mike Le: Honestly, I don't know exactly what happens in five years — this space changes every month, every three months. But I'd say there's still a lot of room for AI to grow. In forecasting specifically, large language models by nature aren't built for this kind of problem — ChatGPT is still a famously bad chess player; it doesn't even reliably remember the rules, which is why solving complex, structured problems still requires specialized models. But I think that's merging — AI overall will get more intelligent over time, to a point where we have a much broader understanding of what it can do than we do today. Right now we mostly use AI for productivity, efficiency, and creative ideation. In the future, I think how humans engage with technology fundamentally changes — five years is a long horizon in this space, and I don't know if we'll reach AGI by then, since that would change everything. But I do see AI becoming more centralized, where a single model can solve many different types of problems and function as something like a digital employee on the team — executing tasks intelligently. That's a trend already starting, and I want Conative to be at the forefront of building those digital employees for inventory and demand planning.
Justin Leibow: That trend toward digital employees is fascinating — and given how hard six months is to predict right now, five years is a real stretch. It's an exciting time, though for some people it's a scary one — there's just so much to consume, and if you're not experimenting and trying things, you'll get left behind.Mike Le: If we really embrace AI, we're on the winning side of that, because as humans we can now do far more with AI's support than we ever could alone. It's an exciting time — we just need to jump in.
Justin Leibow: Unfortunately we're out of time. Mike, thank you for joining me on Product Talk — I learned a lot about Conative AI, about your background, and the real risk you took to get here, and how that's paid off. Congratulations, by the way, on Conative being named Best Product in Q4 by Products That Count — well deserved, I should have mentioned that at the top.
Mike Le: Thank you — I'm really honored, and I enjoyed our conversation.
Justin Leibow: For our listeners — hopefully you enjoyed this. Download it, watch it wherever you get your podcasts. Thanks again, Mike.
Mike Le: Thank you, Justin.
[Closing from SC Moatti, Founder and Chair of Products That Count, thanking listeners and pointing to productsthatcount.org — standard show sign-off, not Conative-specific.]
