How Smart DTC Brands Turn Returns Into a Marketing Advantage

AI Operators.

Welcome to episode 4 of the AI Operators. Guys, the world is changing. Things are moving quickly. We are here to help you cut through the noise and apply what matters. We are here because of our friends at Richpanel and AfterSell. Ben, It has been, as it always is, quite the week. How you doing, man? How's the move going? How's your daughter?

Fill us in.

Oh man, it's been so busy. So many things happening in my personal life. And then there's all this AI madness. I can't wait to dig in. But as always, can't complain. Living the dream. How are you?

Living the dream. I'm wonderful, dude. The highs and lows. Michigan wins the national championship last week. Miami Heat lose in a thriller last night. So we're, we're dealing, but we do have a lot to cover today. Here is what is coming. Retail just lost $850 billion to returns last year, and AI virtual try-on is finally getting good enough to fix it. Ketches, backed by LVMH, just went live. Zalando cut returns by 40% in testing, and Google is embedding try-on directly into search. We are breaking down why this just might be the most important margin play in e-com right now. Plus Reddit's CEO says that he is hiring new graduates over experienced workers. Why? Because they're more AI native. We are sitting down with Joe from KetoneIQ on AI for creatives, Stephen from Cumulus Coffee on running finance entirely through Claude, and Matt, the chief of staff at Fulfill, on how an AI native ERP is changing the way that brands run operations and finance. Let's go. AI virtual try-on has hit its inflection point. Retail's $850 billion returns problem gets a real fix. Okay. So CNBC reported on April 5th that AI startups are finally cracking the code on what the industry calls the silent killer of retail returns. The numbers are staggering. $850 billion in returns in 2025 alone. That is 19.3% of all online sales coming right back. Companies like KETCHES, backed by LVMH, just went live on Amiri, and they are projecting a 10% lift in conversion rate and up to 20 to 30x ROI. Okay. Maybe we'll believe it when we see it. Google is embedding virtual try-on directly in their search product results. And Macy's and ASOS are getting into the game too. Why wouldn't they? Look at Ben. As we watch this wonderful tool that Maheen has mocked up for us, and we see you in that beautiful outfit, I wanna flag before we dive in one of the potentially biggest concerns with AI, which is catfishing. I know you, bro. You don't look like that. Um, Ben, walk us through what is going on right now.

I should probably get this outfit. Yeah, so look, returns is really a huge headache, not just logistically, but it just, it destroys your margin, right? You pay to ship it out, you pay to ship it back, you pay someone to inspect it, repackage it, restock it, and then, you know, most of it you can't even sell at a full price. It needs to go to discount or liquidation. So the all-in cost of a return for most DTC apparel brands is 30, 40% of the original sale price. That's substantial. Now this new tech is using diffusion models, which is the same underlying model being used for Midjourney, for example, to map a garment onto your actual body shape so it looks a lot more realistic, and that increases confidence with shoppers and as a result reduces returns dramatically. We're talking about 40% reduction in returns for apparel. That's massive. So, and what's more exciting for me even is that Google is embedding Tryon directly into search results starting end of the month. So you can imagine what that would do to add conversion rate. Now, if you do the quick math on what that means for a real brand, say you're doing $5 million with 25% return rate, that's $1.25 million in returns. Cut that by 40%, you get $500,000 back in your revenue. That's pretty incredible.

I agree this is a big deal, but not for the reasons we're talking about. So let's use my beautiful wife, Ellen, as an example. If she can buy, you know, or see herself in 20 different Hill House dresses from our friend Nell's company, like she's just going to buy 20 dresses, right? I know that. And so the reality here is the story is not in returns. It's in ads, as all things AI can be, right? You know, you're going to start to get mass personalization at scale, the ability to see yourself consuming a product, wearing a dress that's going to make you buy buy, right? And so the future of this is mass personalization of advertising at scale. I don't necessarily know that the solution is in returns. All right, on to the next. Reddit's CEO is hiring AI-native graduates over experienced workers. Steve Huffman says that new graduates are so much more AI-native than their older peer— peers. So Fortune just ran a piece on March 23rd where Steve Huffman, Reddit's CEO, said that his company is going to go heavy on hiring new graduates. His reasoning is simple. They are so much more AI native than anyone who's been in the workforce for 5 to 10 years. He's not just talking about engineers. He means across the board— marketing, operations, finance, content, analytics. Ben, what's happening?

So here, here's the thing. If I'm a 22-year-old and I grew up prompting ChatGPT through college, I use Claude Code before I open a spreadsheet, right? I don't see AI as a tool I adopted, right? It's just how I work. Now compare that to non-AI natives, that's, there's a real advantage here. But what I think Huffman is really saying here is that experience without AI fluency is becoming less valuable, but it's not the same as saying that hiring new grads over experienced people is the best way forward. The best hire is still someone experienced that's also an AI native, and these people exist. They're just harder to find right now. For anyone building a team today, the takeaway is pretty simple. AI fluency is super important. It's baseline for hiring the same way Excel proficiency was 20 years ago. You should be testing for it in interviews and don't just ask, do you use ChatGPT? But ask something like, show me a workflow you built, or what are your favorite AI tools and why? That kind of stuff tells you everything.

It does. And what Steve is talking about, like, this is real. Let's take a step back, right? If you don't know how to use AI properly, that's a threat to your livelihood. Candidly, we saw Block. Just announced that they were laying off 4,000 people, Oracle, 30,000 people, Disney, 1,000 people. This is a paradox right now where you'll go to a corporate company because you think it's safe, but candidly, I'm not so sure that that's secure anymore. And also these companies might be restricting your ability to lean into the tools aggressively, which is a problem in your future career development. I talk to college grads. And, and recent grads all the time. And one thing that I'm really interestingly hearing now is people who are getting offers from Fortune 50, Fortune 100 companies, I'm talking about like best-in-class engineers are saying, these guys move too slow. They're not going to let me use the tools in the way that I need to. I need to go somewhere faster. So ultimately, you know, Reddit CEO Steve is, he's half right, right? Like, yes. AI proficiency is critical, but Ben, it's everything you just said and we talk about here. This is not about just let the AI do it. It is about how can I, as an individual, imbue my expertise, my perspective, my skillset into the tool so that I can be better, faster, and do more. And so the solution here, Ben, is exactly what you said. It is not hiring a college grad who knows AI well. It is making sure that your most senior employees, the best of the best, are leveraging this technology to imbue their perspective every step of the way. Okay, onto the next story. Anthropic has launched Claude managed agents. From months of work and infrastructure down to days, Anthropic wants to help you as a non-technical operator run AI agents. So guys, the AI wars are heating up. A couple of months ago, OpenClaw launched, then it was acquired by OpenAI. Now Anthropic has launched their latest response to that threat. They are making it even easier to help you build agents in Claude Co-Work. Ben, what's happening?

Yeah, so the way I see it, this is Anthropic's answer to OpenAI o1, right? So we spoke about Computer Use in previous episodes. Computer Use sits on your desktop computer. It helps you manage tasks, create files on your computer. And it runs autonomously. So managed agents is the next level up. So instead of using Claude on your machine, this is running on the cloud and integrating into your existing SaaS software stack, right? So it's working autonomously for hours if needed. Let me give you some practical example. You can have your inventory system run an agent every morning, checks stock levels across channels, flags low stock risks, and drafts the PO. Another one is having your returns platform analyze every return reason and send your product team a weekly report on what's broken and how to fix it. So while Computer Use makes you faster at your desk, Managed Agent makes the tools you already pay for smarter, and that's really the difference here.

Ben, thank you for that overview. Thanks for keeping us up to speed. Onto the next. Lastly, Shopify just gave AI agents the key to your Shopify storefront. The AI Toolkit turns Claude Code, Cursor, and Codex into Shopify operators. Another week, another Shopify announcement. Claude Code can now plug directly into your Shopify store. Ben, what does this all mean?

I'm actually really excited for this one. Look, before this release, if you're a developer working on Shopify, you would use Claude or Cursor or other tools to create code. You then copy, paste it into Shopify and hope it will work. Many times it wouldn't. And the reason is that Claude and other models are being trained on stale documentation. When you now connect Cloud Code directly into Shopify, you know you always get the updated documentation. So when you create code for your store, it's going to work perfectly. That's for engineers. For non-technical users, this is also exciting because if you run a store on Shopify and you want to now change some information on your products, if you want to update some SEO meta tags, if you want to reduce all prices by 8%, instead of going and doing it manually on the store, you can tell Cloud Code to do it for you and it just works. I love it.

Of course you love it and you're excited by it. You are an engineer who's bootstrapped a business to a multi-billion dollar valuation. I'm not. I think, cool, big deal, engineers can now work in Cloud Code, instead of working in Shopify's storefront. For me, the breakthrough here will come when non-technical people can be leveraging the tools with access to the storefront information to make changes, to do cool stuff tied to their unique role. Ben, are you seeing anything in that world?

Yeah, so I do think this will help non-technical users. I will warn against vibe coding on your Shopify store because what you do in Cloud Code that's connected to your store is being pushed to production. So what you want to do is you want to work with a developer or an engineer to set up these workflows that you want to do. Say you want to decrease the, play with the discount schema on your products or change some SEO tags. You create these workflows with an engineer first, on a dev store on Shopify. You test it out. Once you see that works, you push it live, and then a non-technical user can do those every day, every week, every month, and save a ton of time.

[Sponsor Content] Amazing. Thank you for keeping us up to speed. In closing, I'm gonna repeat something you've heard from me often. You can start to see now where these large language models are connecting to outside datasets to get smarter and more personalized for you. We are increasingly seeing this becoming an operating system for white collar work. We are brought to you by our friends at AfterSell. Most brands have deployed AI across acquisition and creative, but the confirmation page, the highest intent moment in all of commerce, remains a dead end. AfterSell is the design system for the entire post-purchase flow, unifying cart checkout, post-purchase, and confirmation into one intelligent system. AfterSell brings real-time AI decisioning to the confirmation page, powered of course by Brain, which analyzes 1.95 trillion data points across 7.5 billion transactions annually to determine the next best action. The result? Instead of generic offers, customers see timely, relevant partner offers from 500 different brands like Hulu, HelloFresh, and Venmo, generating pure profit per order at scale with zero ad spend and zero customer data shared. Brands like Sephora, Nordstrom, True Classic, and Jones Road have generated over $1.5 billion in pure profit this way. Activate AfterSell at aftersell.com/operators and unlock AfterSell's full optimization suite, cart checkout, and post-purchase completely free.

Thanks for having me, Craig. Happy to be here. And clarify, I'm part of a team of 4. We're all badass creators, but it's, it's the full team that makes it go. And, you know, I'm a creative marketer there. I'm a flex player creatively, primarily doing editing, but working with partners and shoots. We have tons of partners at KetoneIQ, so I'll be bouncing around traveling to shoot photo, video, but I'm mostly making the ads for Meta, Facebook, Amazon, Google, CTV.

Amazing. And I love the way that you've talked to us about how this is an accelerant to the team. This is part of the team and infrastructure you're all, you're all working towards. So walk us through what sort of a creative workflow at Ketone might've looked like 12 months ago versus how things look today. What, what's changed?

It is a loaded question because it still looks the same. Like if we're, if we are, you know, going to shoot with a partner, it's still generating standard concept, you know, pre-production, um, coordinating, putting that side together, going to shoot, bringing back the assets, editing in Premiere or Photoshop. Standard across the board, probably for most, most brands if they're doing anything. But, you know, where, where we've seen the process amplified is, is one on the pre-production side, coming up with concepts, you know, like we're always brainstorming as a team. We're throwing ideas around, throwing things at the wall, seeing if it sticks, but now I think on that front end, in terms of the research with Perplexity, with Claude, we're like, we're, we have what we need to like battle test our concepts and strategy before we even go into the shoot. So it's, it's kind of like strengthening our intuition going into production on the front end. And then, and, and more context on that, like You have the ability to search anything. You have the ability to ask any question. So it's, for me, it's like the biggest shift has been how creative can I be in my thinking with, with asking agents or whatever you're using to help you and support you. And you have access to so much information, like go and research the top trends in Meta ads. What's converting? What's converted historically? Who's talking about it? Are there TED Talks on it? Are Are there, you know, Adweek presentations?

Are there—

find any kind of, any kind of publications that talk about it. So it's like you can run it through that lens before you even establish concept. I think that's, that makes us much more confident. And then, you know, on the, the shooting side, we're still getting stuff. We gotta go do interviews, we gotta go get product shots, we gotta go get the B-roll. But I think where I see the support the most with AI is on the post-production side. So it's stretching assets, it's putting photos in different scenarios, it's, you know, background removals, background changes, lighting changes, you know, B-roll enhancement. And, you know, if you guys are tapped into the, to the newest video tool releases, it's like Seedance 2 is, is pretty crazy. So it's allowing you to move at the speed of imagination. So it's really no bars held on what you can do creatively in the post-production process.

I love the way you talk about amplifying your creativity because of everything you now have access to. And I love the way you talk about stretching the asset in post-production, getting more out of it. So when you're doing that, when you're stretching the asset in post at Ketone, what tools are you using in the AI space day-to-day, like take me through a real example. Let's talk about the last campaign you shipped.

Yeah.

Yeah.

I think, you know, for example, we've got a, we've got a sale coming up. We've got like a free gift and 30% off subscription. Right. And we have a, we already have a bunch of product shots, but what I might do is if I saw an ad win in the past that has a specific product shot, I might bring that initial imagery, no copy or anything, into, into Perplexity I've been using a lot actually, because I've got it connected to a model and I'm dropping an image. I have a skill built into my Perplexity. People do it in Claude too, if you've heard about skills and markdown files. So I have this massive document that walks my model through what makes a good photo. And basically I'm just, I'm yapping with it. I can do it on the phone driving in my car. If I drop a PNG in and say, hey, I wanna show this product in 10 different ways. And I'm just talking through each of those 10 settings, but then it's gonna run through my lens of proper prompting, enhanced prompting, realism, color grading, masking, you know, high-end commercial photography. And it's going to spit out those images for me. So I don't, I don't really need to go into like a tool or Freepik or anything when I want to generate imagery. I'm just doing it there, chatting with my phone. From that point, then I have a set of, I have a batch of images and I have copy that I know worked. Now there's two ways to go. You could try to generate the copy prompting and sending it into the model. But for me, I go with more of a fine-tooth comb so I can adjust things more. So I will bring that into Photoshop and, you know, handcraft that stuff for myself.

When you talk about the skill that you're bringing in, one thing I love about that is it's your expertise, it's your perspective. You've trained the skill or the markdown file on how you work and it is an accelerant for you. So to sort of close, uh, my, my section of questions, what's the difference here, Joe? Like, are you producing more content faster? You know, is it better content different than you could do before now that, you know, you've got AI in the creative process? What's changed?

Yeah. Yeah. I mean, it's, it's, it's all of those things really. It's, it's faster for sure. More nimble. And I think in the D2C world, it's really important for creatives to be nimble because we're testing so fast. You know, not everything works, not everything always gets spend. So you have to keep throwing things at the wall and not being married to a specific concept that you think might work or might not work. Just, you gotta keep going and keep moving. So I think it's allowed me to just like think like a goldfish creatively, have no ego and just keep it moving. And it's a numbers game at that point. If you make enough creative, something's gonna hit. So that's why I've felt like this, these tools have been expansive for me creatively because I can just talk about what I'm seeing and it's gonna come to life very, very fast. So, you know, for, I don't subscribe to the, to the You know, the phrases that we hear, hear that are just so played out over the last couple of years, like editors are cooked and Hollywood's cooked and blah, blah, blah. But in a, in a brand setting, you know, just because these tools are available, it doesn't mean the CEO is going to use them. Doesn't mean the CMO is going to use them. Doesn't mean the shipping team or product team is going to do it. No one's going to do it. It's the marketing team. So it's like, it's literally in the hands of the creative. It's a creative tool. So yeah, I think it's allowed me to move faster, think differently, and yeah, it's allowed me to reach creatively in ways that I didn't really have accessible before.

Really great, Joe. Super, super inspiring to hear. And, you know, hearing you describe your process, it sounds very different than what we hear from other brands. We, hear that the expectations many time is to put in a prompt, whether it's an image generation or video generation, and get a perfect ad, 10 out of 10 every time immediately after. So in reality, what happens is a lot of brands get AI slop and it's unusable. They don't know what to do next. So how would you help DTC marketing teams today to get started and make sure they don't generate AI slop?

Yeah. Well, I think I mentioned skills and for me, that's probably been the biggest unlock really only over the last couple weeks because generating good imagery through AI, something that's realistic, that stops you and evokes something from you is really rare. First of all, because you really have to know how to compose an image. So you have to give it very, very tight latency terms. So your prompting is very, very accurate and precise. A lot of people don't really know how to direct. A lot of people aren't photographers, so it's like different language. But I think that, you know, I've gathered so much information over the last 4 years taking different courses. Learning different perspectives on how to prompt, how not to prompt, how to set up systems to generate good imagery that I've taken all that information and put together Claude skills. So again, like I said, when I'm, whenever I'm going into a Perplexity or a Claude and speaking an idea, it's running through my, my knowledge base. And, and going through checkpoints, checking itself, double checking itself. Quality checking, and it's just, it's a 200-page document that I have it run through at all times.

Wow.

Before it prompts, before it spits out a prompt. So, wow. That's a recent unlock for me. And, you know, I, again, I'm going to say Perplexity is, that platform is sick. Like you heard all the hype about ClaudeBot, right? And like having different agents go and do things for you, that's pretty much what's built and wrapped inside of Perplexity. So you have the, you have a lot of power in that program to help you get from point A to point B. But yeah, I think that's the biggest unlock is understanding and knowing how to utilize skills for whatever it is you're doing, whether you're writing copy or generating imagery, generating video, doing general research. Um, if you can build out a skillset for the model to refer to as it goes through the process and then gets you something, that's the unlock 100% for me.

Incredible. Incredible. You know, that's really refreshing to hear about perplexity from creative individuals. I think it's probably the first time, uh, Craig and I have heard that, so that's really cool. My last question to you is, you mentioned your process with images, with the 200-page knowledge base and brand voice and your preferences. Is that something you do for video ads as well? Is that the same process?

Well, I think I look at that as like a two-step process. I do have a video prompting skill that I built over the last week specifically for Seedance too. I'm still learning that one 'cause it's pretty fresh, but basically the most important thing to me is getting good imagery first. So I want to, I want to break that process up 'cause I don't want to prompt, say, make an image and send it through Vidyo. Like I, I want to be able to intervene in that process direct again at the Vidyo stage and then get my output 'cause I'm generating a bunch of images, quality checking those images, and then sending it through my system to get prompts for those. For those images to put into the video platform. So yeah, that's a two-step process for me, if that answers your question.

No, it does. It does. Last question for me is, have you worked a little bit with the node-based programs like Weavies and Floras of the world? I think the open source version is ComfyUI. Have you checked those out?

Yeah, I'm, I've dabbled a bit in, in Weave. Uh, I'm learning Comfy, but I think at this point it, it doesn't feel like a, uh, like a skill that I would tell another creative that hasn't gone into it yet to do.

Mm-hmm.

Because you have, um, platforms like Flora. right? And Flora is, um, similar to Weavii where it has this node-based system, but there's an agent attached in that, um, the, the, the board, uh, the canvas, I, I think is what they call it. And instead of you having to go in there and if you wanted to do a product photography shoot, you would have to, in Weavii, set up all the nodes by yourself. And build that workflow by yourself. In Flora, they released Fauna inside of their canvas, which is an agent that all you do is you talk to it and tell it what you want to do, and it goes in and auto-builds those node-based workflows for you. So it's, you know, for someone like myself that hadn't dug too deeply into the node base, it's a nice tool. And for someone that's just getting started in that and it's, you look at those nodes and it can be very overwhelming. Um, that's where I would go. I would go to Flora.

Yeah.

Yeah. I hear it's extremely, extremely technical to understand all the different nodes, connect them, and it takes a ton of time. And then, uh, not really sure the juice is, is worth the squeeze on that quite yet. But yeah, that's a wonderful perspective. Uh, Joe, thank you.

Joe, I want to close with two questions. One, Let's say there's a creative who's listening, who's behind you. They haven't necessarily gotten started. What's the one thing that they should do to get started? And two, why is that one thing following you on TikTok to learn your tips and tricks? Only half joking. What, what should they do? And tell us where, where we can find you.

If you're in the stages of trying to figure out how to unlock this AI content stuff and you're in someone who's in a similar position to me, like making stuff for brands, even making your own content. You know, focusing on the brand work, the most important thing to me is good imagery and building out a lot of assets. So figuring out, you know, do your research. There's, it's really easy at this point to like scan the internet, scrape the internet and see like, hey, who are the 50 top leading voices in understanding how to prompt for imagery? What are, what's the average themes between all of them? What should I take and leave? Um, let's build out our own skill that follows the, the themes that run through all of these people. And again, that research is easy. You could do that on Claude, you could do that on Perplexity, but I would build that skill out so you have a system for yourself in, in building out imagery. And I think again, that's been the biggest unlock for me is, is learning how to use skills in agents. 'Cause now I can really make, you know, I made a Jeep ad last night just for, just because I was hanging out and making it and like did it in an hour and it looks like a, it looks like a big budget commercial. And again, that starts with imagery. Like you can learn the video stuff after. It's getting so intuitive and so easy to use, but the image part, I think people overlook it sometimes. That's the biggest unlock. Image realism is a really, really sought-after skill in this, in this world for sure.

And Joe, I think building on that, right, you don't have to start with an AI-generated image. You start with, like you said, ketones, taking a ton of product shots. That's your starting point to then build on. So thank you so much. Everyone should go follow you on TikTok, Joey Bag of Content. Dude, you are a wonderful friend and I am greatly appreciative to have had you join us. Thanks so much. We are so lucky to be joined by my friend Steven, who leads the finance department at Cumulus Coffee, which makes the most delicious cold brew coffee on the market. The reason I am so stoked to have Steven on the show is because I honestly think he might be one of the most AI-forward finance leaders that I've ever spoken to. 5 months ago, uh, Steven was an AI and a ChatGPT skeptic, and now From what I gather, he lives in Claude, saving a ton of time operating as a lean team. So Stephen, welcome to the show, man. Tell us about yourself.

Yeah, absolutely. Appreciate you guys having me. So, uh, yeah, my name's Stephen. I am the head of finance here at Cumulus Coffee. Uh, joined about a year ago. Um, before that, spent my career in, uh, investment banking, private equity, and investing. So, uh, yeah, excited to be here and, uh, talk about AI.

So, so let's dive in, right? When you're using, let's say, Claude for finance, like, what does that mean? Like, walk me through a typical day or week of the role that AI Claude has in your day-to-day.

Totally. And I will say I, I do use Claude, but you know, OpenAI and ChatGPT, you know, Craig, you always say they also create, you know, fantastic models and tools. And so I think you can use all the tools that I'll, I'll be talking about with them as well. But the way I think about using AI in kind of my daily, weekly, monthly workflows are kind of in, in 3 different buckets. I think the first one is just chat, which I feel like with all of the advancements that have come out the last 3 months, it kind of gets forgotten about, but I still think, you know, it, can, you can use it to really, you know, provide some really great insight, um, using some of the tricks and tips, obviously, Craig, that, that I've learned from you. I mean, I think this week in particular, right, there's a couple things, you know, we're working like a lot of companies on how do we get refunds on some of the tariffs given the recent Supreme Court decision. Um, you know, doing some interview scheduling and structuring for some, some new roles we're hiring. Uh, you know, doing some kind of vendor management and billing negotiation, things where you can really use the, the, the tools and the power of the models to, to, to really kind of uplevel, uplevel yourself. I would say that's one. I would say I do a lot of work in Claude for Excel, which is— I think, you know, Craig, we've talked about this— it's— I would say Excel is really what I'm using the most on a, on a day-to-day basis. Um, I would say it's somewhere in the middle between kind of one-off projects and kind of repeatable workflows. So for example, this week I've, you know, updated our long-term model with our kind of March actual results. That was something that would take me a long time before using these tools. It was really helpful. I used it to create some kind of one-off ad hoc reporting for our, for our sales team. That was, you know, something again that would take me 4 or 5 hours. I was able to knock it out in, in 30 minutes. So that is definitely one where, you know, I've been able to use with a lot of success. And I'd say the final thing is just in, in kind of Claude Code, right? And I'm by no means an expert in, in Claude Code, but, um, you know, I've been using it to, to really, uh, automate some real manual work, you know, repeatable manual workflows. So cash flow Cash flow forecasting, I think, is a really good one. I built something that kind of, we analyze some parts of our kind of return data all the time, was able to connect through an API through our kind of our software there. You know, I'm building something to automate some of our inventory forecasting and things like that. So those are kind of the buckets in which I tend to use AI on. Again, you know, every day is different, but on a daily, weekly, monthly basis, it's starting to become a pretty big part of, of my toolkit.

I'm so proud. I'm like tearing up in the sense of like, you're clearly using it for everything, right? And what it sounds like is you got to your breakthrough where it was like, wait a second, right? Of course I can use superintelligence and AI at every different touchpoint of my role. I'm curious, Stephen, right? You're someone who hadn't really used AI, you know, much before 4 months ago when we spoke last, you were at, it's not that great of a tool, right? It doesn't help me. So clearly there's been some. Transformation, what was the moment where you had your breakthrough of like, oh no, this is different?

Yeah, great question. I would say like, you know, we've talked about this, like the chat was always helpful, right? Like you saw the benefit of the chatbots, right? Especially as I'm over the, over the last couple years and being helpful. But you know, as a finance, as someone in finance, my day-to-day workflow was largely in Excel, right? And it was very clunky to get something out of Excel. You'd paste something in the chatbot, it gave you the formula back, you'd paste the formula in Excel, and it was just, it was just clunky. And so, you know, Claude released their Claude for Excel, kind of made it as an agent, I think sometime in December. And that to me was really, really the unlock, right? Really bringing the power of that, of those agents into an environment that I was very familiar with. And, you know, it was one in which I could very, I think, judge its capabilities very well. You know, Claude code, right? It's hard for me to look at, you know, some Python that Claude code has written and say this is good code or not. But Claude for Excel, I know pretty well that it's be— whether it's helpful or not. And so, I really used it for the development of one, our long-term kind of long-range planning model and our 2026 budgeting cycle where I used it to, you know, kind of revamp our, our kind of input cost model. I used to create some really useful outputs for both board members, for investors. I use it to audit formulas across the worksheet. And it was a, you know, a relatively complex model as a bottoms-up, you know, 5-year build. So I was able to kind of you know, audit how everything worked across the sheets well. Um, and you really, and you just really pressure test some assumptions. And I think it unearthed some things that I eventually would've gotten to or, or understood, but I'm not sure, I'm not sure it would've taken me a long time to get there. And it unearthed them quickly and allowed me to quickly kind of iterate and, and, and make the changes needed. So that was really for me the, the, the, the big unlock.

Amazing. And I, I don't know that people understand when you talk about using AI in Excel. I mean, I've literally worked with you and 75 other finance leaders where it's actually building the spreadsheets out itself. You're watching it build the spreadsheet. You're watching it analyze the outputs of what it created. So I'm, I'm curious, Stephen, as you lean in so aggressively as a finance leader, obviously you care about numbers, you care about accuracy. How do you think about the accuracy of the tools and your role in QAing them?

Yeah, really, really good question. So finance is inherently a little tricky in this respect, right? Where You know, for code, for example, I've heard that I think the reason why AI was such a killer use case for code is 'cause it's verifiable, right? You write the code and you run it, it does what you want or it doesn't. And I mean, I'm sure some engineer will tell me it's, and I'm sure there's more nuance than that, but generally speaking, right? That's, I think that's why it's been, you know, been used so, so often in, in that respect. But from a finance perspective, you know, the difference between a 5% growth rate and an 8% growth rate, or, you know, a 24% margin projection and a 19% margin projection is a material difference in a forecast, but they're both the results of inputs that are very plausible on its surface, right? Which is— so that, that element of it is tricky. And so look, the way I kind of think about it, ultimately it was true when I started my career, you know, 13 years ago at this point. It was true before these tools came out 5 years ago. It's true now that you as a finance leader are ultimately responsible for the numbers and projections that you put forth. And so from my perspective, it's, you know, can I blindly trust a hardcoded output that Claude develops? No, but I probably wouldn't have if a financial analyst did the same thing 5 years ago, right? And so I think it's incumbent upon anyone as a, you know, as a finance professional to figure out what are the checks and balances and safeguards that you're using. You know, ultimately I have seen, you know, I have seen these tools create an output that was, you know, incorrect. Generally speaking, it is more the result of a lack of context that I have given it about a specific problem versus hallucinating a number. That still, that still does happen. And so you have to keep, keep an eye out for it. But I think the error rate versus, you know, a human creating the analysis, I think is candidly pretty similar. And I think, you know, when things are really complex, its ability to dig through a lot of complexity in one, go is, is very, is very valuable and worth the trade-off. But again, ultimately up to you and you need to verify, verify what you're putting forward from a finance perspective.

I love that framing, right? Like I think people expect 1000% accuracy, but you know, sometimes who are the most strategically influential, but also wrong individuals. It's like human beings, right? And so you have to kind of trust, but, but verify. Steven, I want to close with one question that isn't necessarily tied to your role as a finance leader, but more sort of strategic, right? I think a lot of CEOs and CFOs are saying like, hey, what's the ROI on this stuff? Where are we going to start to see some financial impact of the tools as, as we invest here? As you think about your role as a financial leader at Cumulus, how are you thinking about the role that AI will play going forward on, on the balance sheet?

Yeah, really, really good question. So I think, and I think this answer for me and Cumulus is going to be different than it would for a large company, right? I think you just heard, I think it was Uber's CTO who came out and said they blew through their cloud code budget in like 3 months. So I think that's, that's a very separate conversation than a smaller team like, like ours. And so I think for a company like at our size, you know, it's more about figuring out who is passionate about these tools and who can drive value with these tools, um, in ensuring proper education throughout the organization. Um, so I think for, you know, I think that is going to be the highest ROI activity that, that we all can do. Um, you know, I think from I think for us specifically, it's going to differ on a function-by-function basis. I think for the finance team specifically, like, I think we probably could go without having, you know, an FP&A head for a lot longer of a time than we could have before these tools came out, just because I think that is work that you can get to a baseline minimum with these tools much easier than you could previously. I think you will still get to a point where you will want that human in the loop to kind of drive the intuition forward and have the entire context of an organization to, to drive forward. So I think that for, for us, that's probably how we think about it. And look, I can say that, you know, since using these tools, it's not like my workday has shortened a whole lot. It's in fact probably gotten a little bit longer. And I think that experience is pretty common among people who, who use these tools a lot, that you're just able to get more deeper in your, you know, to-do list and the to-do and the ability of these tools unlocks more and more things that you can do that you never could do before. Um, and so from that benefit, from that standpoint, it's probably like an overall benefit to productivity. From an organizational standpoint though, I think you have to look at it on a, again, where is you, where is your organization in its, you know, journey from a maturity perspective? And what is the goal of the, what is the overall goal of the function and how can AI really drive that? Now, again, like if you're at a $100 billion company, you have to start thinking about, thinking about things like, token budgets and humans versus tokens and those kinds of trade-offs. But for a smaller business, I think it comes down to what can these tools allow you to do that you couldn't before?

Fully, fully agree with that, Steven, and really impressive stuff. I know very few finance operators that work with cloud code, so kudos to you. That's really impressive. Um, I just have one question and we can close this out. So if I am a DTC operator doing $2 to $5 million in revenue and I don't have a head of finance or a CFO, which is most of our audience, what financial workflows should I hand to Claude first and how would you go about it?

Yeah, really, really good question. I think for me, this is one where it's a really unique opportunity to use these tools, right? Because I think before, like, if you're, if you're an operator and it's a relatively small business, like you need some sort of long-range planning model, you need something that can guide your decision-making product, your decision-making process. And look, ultimately, you know your business better than anyone. You know the drivers and the inputs, and now you have a tool that can help build the scaffolding around those and really drive, um, drive forward that type of, tool and use it both for internal purposes and also external purposes, right? Investors, board members, banks, et cetera. They all expect this sort of, you know, relatively detailed long-term model. And I think now, again, you know your business better than anyone. Now you have a tool that can provide scaffolding around it. Is it going to be as impressive or as, you know, potentially detailed, useful as, you know, having a CFO? Probably not. But again, at this stage, it's going to get pretty, pretty damn close. And I think that's a really unique opportunity. So I think that's the one that stands out the most. And then I think just generally cash flow forecasting, right? Like any, you know, business on that side, cash is obviously king and having a keen insight into cash is really keen. And so, you know, figure out how to plug it into wherever you receive your invoices, plug it into however you receive receipts and cash from customers, get your bank account in there. And let it go to town and it's going to probably give you a pretty, pretty good 13-week cash flow forecast. And that's something that before would've taken a lot of manual work. And again, you know, is it going to be as advanced as, you know, one might be with a finance professional? Probably not. But again, you're going to get, you're going to be able to create a tool that will be very useful to you in operating your business. And ultimately that's the, that's the purpose.

Stephen, what a great, great advice. I think cash flow is the silent killer of most small businesses. And so I think that's really, really great way to use these AI tools. Thank you so much for being on the show. So if our audience want to find out more about you and Cumulus Coffee, where do they go?

Yeah, cumuluscoffee.com. We sell our machine and our delicious coffee and and various other things through, through our website. So come, come check us out. Give us a try. I guarantee you it's, it's, it's, it's as good a cold brew as you're going to find. So highly, highly, highly endorse it.

And the reason that it's as good a cold brew as you're going to find is because you'll notice I'm chugging some Starbucks here. It was started by the former global head of innovation at Starbucks, the head of retail at Starbucks, our dear buddy Mesh. Stephen, thank you so much. We're going to have you back on the pod. We'll have to talk about agents and other things, but I will see you tomorrow, man. Later. Sounds good.

Sounds like a plan. Thanks, guys.

[Sponsor Content] Thank you, Stephen. Some brands are cutting their support tickets by almost 50% with AI. I've actually been recommending Richpanel to my community for the last 6 months, so I'm really happy they're the headline sponsor of AI Operators. What they're building is pretty different from most AI support tools. Most tools are basically just AI agents. You train them once and hope they behave. Richpanel built what they call a Boss AI. You just tell the Boss AI your values, principles, and desired outcomes. It handles the rest. The Boss AI reviews thousands of conversations every week, scores them, and continuously improves them to get your desired outcomes. Brands like Richpanel cut ticket volume by 45%, and Jones Road actually cut their support team size in half. and both of them were live in under 2 weeks. Richpanel even guarantees 30% ticket reduction in 60 days or your money back. If you want to see what that could look like for your support team, go to richpanel.com/demo.

Yeah, thanks for having me, guys. I'm Matt. I'm the Chief of Staff at Fulfill. We're everyone's favorite e-commerce ERP, handling all the fun stuff on the ops and finance side of things on the backend. And recently AI has been all the rage, both with our customers and internally in my day-to-day workflows. So excited to be here and talk through how I'm using AI.

Wonderful, Matt. Well, thanks for being here. Really curious, can you walk us through how you actually use AI in the day-to-day at Fulfill?

Yeah, I can give a couple tactical examples. One example is using Claude for Work to take information from screenshots and create a spreadsheet out of that information. So I had about 30 different screenshots. I had a bunch of people's names and email addresses, and I wanted to copy and paste it into a spreadsheet. And instead of going in, doing all of that copy and pasting, I just threw it into Claude for Work and it bat out a spreadsheet in a couple seconds.

So that's one tactical example.

Another tactical example is using MCP connectors. So for example, Indeed has an MCP connector. MCP is essentially think of it like a USB port for AI tools like Claude and ChatGPT to connect to other softwares. So Indeed has an MCP server that lets you do benchmarking salary research with it. So if you're doing any hiring, that's a pretty cool feature to, uh, to use there to help streamline that process. And then third, final example I'll share here, um, and one of the examples that first got me on kind of the, the AI hype, AI buzz, um, is creating a custom GPT. Or a Claude project for Claude users. So it's essentially training this chat with all of this context and information that other people in your team can then go and get questions answered from.

That's really cool. We also spoke earlier about you building a Claude project to help coach and ramp up new hires. How does that work exactly?

Yeah, so I find a lot of times when people are onboarding or coming into new teams, right? Oftentimes you're either sending documentation for people to read, videos for people to watch, or maybe they're sitting through live training sessions. And so what we did is we took transcripts from all of our live training session calls and created Claude projects for each of them and told the Claude project to train and prompt people having a conversation with it to help reinforce their learning from that session. So the Claude project would take the transcript. From the training session, devise some prompting questions to gauge the person's knowledge and retention of the information. And it was trained to not just take their answer and move on like a regular quiz or survey would. It was really trained to push them, ask them follow-up questions, make sure that their answer is going into sufficient depth that they've reflected appropriately on what they've learned before them moving on to the next one. So that's been a big unlock for both helping people reinforce their knowledge as they're learning new concepts, and then also freeing up time internally as well. And to have fewer, the need for fewer one-on-one calls.

That's really, really cool. I know Craig likes to use NotebookLM for that use. Craig, have you used Claude for that as well?

I have used Claude for, for that. And I, I do also just want to build on one use case that Matt shared, which is this idea of aggregating information for the fulfill team about contacts at new hires. But I've seen so many brands stand up similar instances where for legal claims. Hey, can marketing teams say this or that? And then it will spit out, no, we're not approved to say this for reasons X, Y, and Z. So Matt, I think it's brilliant, this notion of aggregating data to really just be able to query and ask questions relative to if something is and isn't okay or who's the contact there. I, you know, kudos to you guys.

So last question before we move on to some of Craig's questions. You mentioned you're using Claude to give weekly performance analytics. I'm sure a lot of people in the audience would love to hear a little bit about that. What does that look like at Fulfill?

Yeah, so this is more of a, a personal performance analytics exercise. So connecting, connecting Google Drive, connecting Gmail, connecting Slack, any tools that I'm using in my day-to-day workflows, turning on those connectors in Claude and asking it to reflect on all the areas of development or things that I'm working on personally or professionally, and then asking it to rank me based on my conversations, based on how I'm spending my time. Like, these are my priorities right now. Do you think I'm prioritizing correctly? And if so, where did I stray from those priorities? Am I sharing enough context when I ask people to do something, right? If not, flag those threads where I could have given more context. Was there anything that I said what, that was confusing or contradictory? So any of those things that you might be working on as a leader or Even an individual that, you know, you're working towards your next promotion or performance evaluation, I find that that's a great way to just prompt Claude and it has access if you give it access to your calendar, your email, your Slack, everything that you're spending your time doing all day. So I feel like that's a cool little, uh, personal development coach for you in your pocket.

We should do the same thing for the show. Craig, how do you think you're gonna do with Claude? How do you think it's gonna score you?

I think we're getting better and better, but you know, I, I use that exact same use case, this idea of getting real-time sales feedback or coaching based upon the things that you're working with. It is a wonderful sort of like arbitrator or arbiter of truth and, and spitting things out back that way. I want to move over though to the customer side. So Fulfill has built MCP integrations, Claude Skills, and even micro apps directly into your ERP. So for someone who doesn't know what any of that means, like what can a D2C operator exactly sort of actually do with this today?

Yeah, so the concept behind the, the MCP connection there is that because of Fill is a source of truth for a lot of your operational and financial data, when you connect it to Claude or ChatGPT through our MCP server, you can then interact with your ERP data directly within Claude or ChatGPT. So you could ask Claude, hey, when was the last time Target placed an order with us and how much was it for? Right? Something as simple as that, that historically maybe would've taken logging into the ERP, click to the reports, add a filter, right? That's something that you can quickly ask even in the Claude mobile app when you're on the go, if you're in between retailer meetings at a conference, for example. So that's a, a cool use case to be able to get answers to questions. You can also build dashboards and reports within Claude as well using Claude artifacts that are generated from that connection. And then we also released the ability to create and edit and update ERP records from Claude in ChatGPT, of course matching your permission. So if you can't create a purchase order in the ERP, you can't do it from Claude. But, um, that's a cool way for people to say, hey, create a PO, update the journal entry, whatever kind of action you're wanting to take, you can now initiate those from Claude and ChatGPT. So that's been a big unlock for merchants, especially when they're on the go. And another thing that we rolled out as well is the ability to essentially vibe code custom micro apps. So if you're looking for a specific dashboard rather than building it in Claude and having it live in Claude, you can use our MCP and our MCP will actually allow you to deploy that custom interface within your ERP workspace. So you don't need to be a developer. You don't need to worry about where you're hosting the app or deploying it. Our MCP takes take care of all of that. And so it's been really cool for non-technical users who, you know, maybe have heard of live coding and they don't know what it is, or they want to, uh, want to show off their technical skills there, uh, to be able to go and, uh, live code custom, uh, dashboards within the, the ERP. So those are a couple cool use cases that our, our customers have been using AI for.

Amazing. You know, Ben and I spend so much time talking about this idea of Claude as your operating system for work and the way that you're Setting up these connectors and giving people access to fulfill information makes total sense. We heard a lot from Dan on the team at Topo Designs about the ways that he was leaning into similar things in the supply chain space. So just absolutely, uh, incredible. Um, I want to close with one question, which is you said that the world has changed a lot over the last 6 months. And for an operator who's still running their business, maybe in spreadsheets and are, you know, pretty basic Shopify reporting, like What is the first AI workflow that they should set up? Like maybe not the most advanced thing, maybe not necessarily Justin Fulfill, but what is the first thing that someone who's listening right now should, should start to do?

I would think through any software that you're paying a bill for, regardless of how cheap or expensive it is, how much or little you use it, and go and see if those softwares have NCP servers that you can connect into Claude. Even if it's something like Canva or Figma or Apollo or HubSpot or Google Drive, Slack, Gmail, whatever softwares you're using, try to get those all connected into Claude or ChatGPT because having context and everything in one place is gonna be a game changer for getting answers to quick questions. So if you are living in that spreadsheet world and jumping between different softwares, that would definitely be my first step that I would suggest would be to go and just get connected. That you, uh, every, everything that you can connect, get it connected.

Matt, thank you so much for being here. If people want to learn more about Fulfill, if they want to download the Claude skills that you've made available, where can we find you?

Yeah, check out fulfill.io. We'll, uh, we'll share the link in the show notes here so you can check out our, our ERP software and also the free Claude skills that we have to help determine free shipping threshold analysis and fulfillment optimization.

Amazing. Later, man.

Thank you, Matt.

Everybody's favorite segment, AI 101. This time we are going deep on context. This is the one concept that determines whether AI gives you a generic, crummy answer or a legitimately useful one. Ben, I often say that as human beings, we're now just context engineers orchestrating between different layers of context. What is context and why are we context engineers?

So look, there are 3 levels of context that matters for operators. Level 1, conversational context, right? This is just what you've said in the current chat. Every message you send adds to it. Level 2, project or system context. In Claude, you can set up projects with custom instructions that persist across every conversation. So you write your brand guidelines, your tone of voice, Your customer profile once and every conversation in that project starts with that foundation. Level 3, tool context. And we heard some of our guests talk about that. When you connect Claude to your data, your Shopify store, your analytics, your customer database through things like MCP servers, that's external context. The AI can now reference your actual numbers not hypothetical ones. What I've seen is that most operators are stuck between level 1 and level 2. They're having one-off conversations when they should be building projects with tons and tons of context. It's the difference between hiring a contractor for a day versus working with a team member who knows your business inside and out.

Ben, I am proudly non-technical, and so I'm going to ask you if how I'm about to sort of like understand this all is accurate. Let's say I'm talking to my, my son Coley, and I'm just telling him, hey, large language models are fancy autocompletes, and the output that you are going to get when it is autocompleting the sentence might just be a little stronger if you give it some background and clarifying details about yourself. The output will be a little better if you feed it the relevant context. Do I kind of have it right?

1000%. The more The model can learn about yourself, about your preferences, about your style, um, the better the result is going to be.

All right. Time to tell Coley to get off the basketball court and in front of a computer. Let's learn about some context. All right. And that is it for episode 4 of the AI Operators. Thank you, the listener, for listening. Thank you to Joe at Ketone IQ, to Steven at Cumulus, to Matt @fafil and to Dario. I am so sorry. We did not have time to get you out of the waiting room. Maybe next week. Guys, if you like the show, please like, rate, review, subscribe, follow. We are here to be your resource to cut through the noise and apply what matters. We will see you next week.