Latest / The DTC Marketing Podcast with Fexingo: Direct-to-Consumer Brands, Shopify, and E-Commerce Growth / How DTC Brands Use AI for Predictive Inventory Management
Transcript
- Lucas: One of the most painful things for any direct to consumer brand is the inventory guessing game. Order too much and you're sitting on dead stock eating up cash. Order too little and you miss sales, frustrate customers, and hand market share to a competitor who actually has the product. Luna: And historically, the way brands solved that was either with a spreadsheet that some poor operations manager updated manually every week, or just gut feeling from the founder. Lucas: Right. But in the last couple of years, a bunch of DTC brands in the 10 to 100 million dollar revenue range have started using AI to predict demand at the SKU level — not just for seasonal trends, but week by week, sometimes even day by day. I want to look at one specific example today: a mid-size activewear brand called VITAL. They're not a household name, but their numbers are really instructive. Luna: I actually know VITAL. They make those seamless leggings that have the little pocket for your phone. They've been growing fast. Lucas: Exactly. So VITAL was using pretty standard forecasting — they'd look at last year's sales by month, add a growth percentage, and place bulk orders to factories overseas. The problem was that their demand was getting more volatile. A post from an influencer could spike sales of a particular color overnight, and they'd be out of stock for six weeks. Meanwhile, they'd have three months of inventory on a style that wasn't moving. Luna: So they brought in a machine learning model. What did they feed it? Lucas: A few different data streams. First, their own historical sales data — that's the obvious one. But they also pulled in weather data for their top 20 markets, because they noticed that a cold snap in a region would kill sales of their lightweight shorts and spike demand for fleece-lined tights. And they started tracking social media trend velocity — how fast certain keywords were growing on Instagram and TikTok. Luna: So the model could say 'hey, this particular shade of blue is trending on TikTok in the Midwest, and it's going to be 10 degrees below average there next week — ramp up production of blue tights for that region.' Lucas: Exactly that. And the model wasn't some exotic deep neural net. It was a gradient-boosted trees algorithm — XGBoost, specifically. That's pretty common in retail forecasting because it handles tabular data well and is interpretable. You can look at which features mattered most. For VITAL, the top three were: sales from the same week last year, the weather forecast, and the volume of mentions of a specific product on social media. Luna: And the results? What changed operationally? Lucas: Within six months, they reduced stockouts by about 40 percent. That's huge — that's revenue they would have lost. And they cut excess inventory by 25 percent. So their working capital improved significantly. They also started doing something they hadn't done before: dynamic rebalancing. They'd have inventory allocated to different fulfillment centers, and the model would suggest moving units from a slow region to a fast one before a demand spike hit. Luna: That sounds like a massive change in how the operations team works. Were there growing pains? Lucas: Absolutely. The biggest hurdle was data quality. Their Shopify backend had a lot of dirty data — duplicate SKUs, missing timestamps, returns that weren't properly logged. They spent about three months just cleaning that before the model could run reliably. And the team had to learn to trust the model over their own intuition. There was one instance where the model predicted a big demand spike for a particular style of sports bra, and the ops manager overrode it because he 'felt' like that style was done. It turned out the model was right — a celebrity had worn it on a TV show, and they missed the wave. Luna: So it's as much a cultural change as a technical one. What about the cost? Is this accessible for smaller DTC brands? Lucas: The software itself can be surprisingly affordable. There are off-the-shelf tools like Inventory Planner or Lokad that start at a few hundred dollars a month, and they integrate with Shopify directly. The bigger cost is the human time — the data cleaning, the training, the change management. For a brand doing 10 million in revenue, I'd budget maybe 30 to 50 thousand dollars for the first year if you include a part-time data person or consultant. But the ROI can be huge. VITAL's gross margin improved by about 4 percentage points just from fewer write-offs and fewer lost sales. Luna: Four points on a 50 million dollar business is two million dollars. That pays for a lot of data cleaning. Lucas: Exactly. And I think the trend is only accelerating. More brands are layering in real-time signals — Google Trends data, ad platform conversion data from Facebook and TikTok — to make these models even more responsive. There's a company called Fero that does this for Shopify brands, and they've seen demand triple year-over-year. Luna: It makes me wonder about the limits though. Can a model like this handle a completely new product with no history? Lucas: That's a great question, and it's the toughest case. For new product launches, you're basically using what's called a 'cold start' approach. You look at similar products, category averages, and any pre-order data. It's less accurate, but still better than a random guess. Some brands use Bayesian models that update quickly as real sales come in. After a week of data, the model usually converges pretty fast. Luna: So the playbook for a DTC brand right now seems to be: start with clean data, pick an off-the-shelf tool that integrates with Shopify, run a pilot on one category, and then expand once you've built trust. Lucas: That's exactly right. And don't try to build your own model unless you have a dedicated data science team. The off-the-shelf tools are already good enough for most brands. The competitive advantage comes from how well you use the predictions, not from the algorithm itself. Luna: You know, speaking of things that keep the show running — we're able to dig into these kinds of specifics because this podcast is listener-supported. We don't have sponsors pushing us to talk about generic tools. A couple of dollars a month from listeners who find value in these conversations genuinely makes a difference. If you've gotten something out of episodes like this, you can support us at buy me a coffee dot com slash fexingo. Lucas: Yeah, it really does go directly to production costs and keeping us ad-free. And it allows us to spend time researching cases like VITAL instead of just reading headlines. So if it's within your means, even a small contribution helps. Luna: Okay, back to inventory. One thing I find interesting is how this ties into sustainability. Less excess inventory means less waste, fewer products that end up in landfills. Lucas: That's a huge point. The fashion industry is responsible for a massive amount of waste, and better forecasting is one of the most effective ways to reduce it. VITAL found that their markdown rate — the percentage of products sold at a discount — dropped from 30 percent to 18 percent after six months with the AI system. That means less product was overproduced and eventually written off. Luna: And that's a marketing advantage too. You can talk about 'made to order' or 'low waste' in a credible way because you actually have the data to back it up. Lucas: Exactly. Some brands are even starting to use the model output as a customer-facing signal. They'll show limited stock levels on product pages based on predicted demand, creating a sense of scarcity that's actually grounded in real supply constraints. But you have to be careful — if you cry wolf too often, customers stop believing it. Luna: So what's the next frontier? Where does this go from here? Lucas: I think the next step is fully automated replenishment — where the model not only predicts demand but also triggers purchase orders to suppliers without human intervention. Some larger brands like Gymshark are already doing this. For smaller DTC brands, I think we'll see more integration between inventory AI and ad spend optimization. Imagine a system that sees a predicted stockout for a popular SKU in two weeks and automatically decreases ad spend on that product to avoid disappointing customers, while increasing spend on a similar product that's overstocked. Luna: That's basically a closed loop between marketing and supply chain. It feels inevitable. Lucas: It does. And the brands that get there first will have a real edge. Because at the end of the day, inventory management is not just about operations — it's about customer experience. If you have the product when the customer wants it, at the right price, without waste, you win. Luna: Alright, I'm going to look at my own inventory of leggings and think about this. Thanks, Lucas. Lucas: Thanks, Luna. Talk to you next time.