Jordan Craig: How a 9-person team uses AI agent orchestration to run like a 50-person operation

[00:00:00] Robert Varon: There's companies with 5,000, 10,000, 20,000 people, and a lot of those jobs are going to be replaced. And I think people have to learn AI, learn how to prompt, and learn how to make themselves more valuable in the age that we're at.

[00:00:20] Kailin Noivo: Welcome to another episode of The Ecommerce Toolbox: Expert's Perspective. Joining us today, we have Rob. Rob is the Vice President of Digital at Jordan Craig. He runs all ecom digital strategy. So, really, really, really excited to have Rob on the show.

[00:00:35] Robert Varon: Thanks for having me. Excited to talk about everything we have.

[00:00:38] Kailin Noivo: Cool. So, let's start off, Rob, a bit with your professional background. So, maybe talk to us a bit about your career journey, where'd you start, and how you ended up in your current role.

[00:00:49] Robert Varon: Sure. So, I started out agency side, working actually for a retention agency. So, very much at the time, it was email-only focused, then it became email and SMS focused, and then ultimately launched the paid social division and really grew that out as well. So, started agency, went to another agency, and at that point, I had already helped well over 100 brands. And I was thinking, I'm helping all these brands with 1, 2, maybe 3 hours max a week of my time. What can I do if I give one brand all of my time? So, after 6-plus years of agency, I decided to move brand side, where I started working for a health and wellness company. Went really well, focused strictly on the P&L and rolling that out. After a few months, I was recruited by Jordan Craig to take over their digital side of the business, and I've been here for the last 2 and a half years.

[00:01:56] Kailin Noivo: Cool. So, for those that aren't listening, like, what is Jordan Craig? Maybe give us a bit of the history of the brand.

[00:02:04] Robert Varon: Jordan Craig is a streetwear brand founded in 1989, so we're celebrating our 37th anniversary later this year. We're in the affordable luxury space. Most of our products range from the $30 to a $150 range. However, some of our jackets are going for the $200 all the way up to $500. So, we have two divisions. We have the ecommerce side, which I head up, as well as the wholesale side, and both channels are really focused on growth, but more importantly, profitable growth.

[00:02:40] Kailin Noivo: Very cool. Alright. Well, let's jump right into it. So, as we know, with a lot of D2C and omnichannel businesses, the online portion of the business is pretty lean when it comes to headcount. So, maybe talk to us a bit about when you got recruited, like, how big was the team? Let's maybe get into it. How lean are we talking here?

[00:02:58] Robert Varon: Sure. So, when I started, it was four others and me and a few agencies. Ultimately, very early on, I realized those agencies weren't adding much value. If anything, they were costing us money in the net long run, so we ended up letting go of those agencies. We replaced one with another agency, but the rest were bringing everything in-house. So, we're currently a team of nine. So, still for our size, pretty lean, but a little bigger than when I started.

[00:03:34] Kailin Noivo: So, you guys are nine, but no agencies? Or, like one agency.

[00:03:39] Robert Varon: We have one agency. So, we have an agency handling Google and Microsoft.

[00:03:44] Kailin Noivo: What did you decide to bring in-house and why, out of curiosity? Like, what was kind of the aha moment for you?

[00:03:52] Robert Varon: Yeah. I mean, email, SMS just made sense. It's where my career started, had a really strong mastery of it. And I know if you can get someone who can spend the time and really optimize it, that is going to be key. So, one thing where I'd say we beat out 99% of other brands is that we have 86 journeys. We're sending campaigns every day. We're highly segmented. We have millions of names on our list. So, we knew, realistically, we could do email. And I knew if I put the strategy behind it, as long as I had someone who could execute on the vision, it was just going to be more successful. And as a streetwear brand, we're doing 100-plus launches a year. Agencies were struggling to keep up with that type of demand. So, that was one that made sense. Paid social or paid media as a whole, we saw that our agency was doing good, but a common experience that a lot of people have is that their agency gets good results, not great results. And again, this was something where I knew, okay, if we can optimize, bring it in-house, really focus on content production, we'd be in a good place to succeed.

[00:05:17] Kailin Noivo: Very cool. So, as we really get into it, a lot of these conversations that I've been having are what I kind of call the pre-AI era. I'm curious, how are some of the more foundational tools that are coming out, whether it's Cowork or Codex? How are these things impacting your business? Or how AI-pilled are you guys, I guess, is my question.

[00:05:40] Robert Varon: We're very AI-forward. It is ultimately the future, and that's not even true anymore. It's the present. So, we really use two main AI tools. We use one called Relay.app and another one called Hazel, or from Hazel.ai. And both of those serve different functions for us, but both are fundamental to our continued success. Relay.app is ultimately a tool where they work with all the main Anthropic, ChatGPT, Gemini, etc, Perplexity, and it is very workflow-based. So, do this action, ask the AI, do that action based off that, do the third action. So, we have a lot of tasks that are being run in the background that just run through Relay.app.

[00:06:32] Kailin Noivo: That's your agent orchestration layer then?

[00:06:35] Robert Varon: Exactly.

[00:06:36] Kailin Noivo: Okay. So, Relay.app is your agent orchestration layer. It plugs into all your systems of record and all of that. Very cool. What are some of the workflows that you went after? I mean, I talked to some people, and they're like, yeah, we use AI to generate product descriptions. I'm like, okay. Yeah. That makes sense. But, like, how kind of in the weeds have you gotten on the workflows? Are there any workflows that you can share?

[00:06:56] Robert Varon: Yeah. I think one very basic one is just getting a full daily recap that breaks down every element of the business. So, what products oversold? What products undersold? What was the total revenue? What percentage came from new customers, returning customers? And that we get it at 08:30 in the morning. So, you wake up, you check the results, you look at it, and then it writes a full-on summary. That's very basic. That's entry-level. Then we have ones that are looking at our AEO and GEO to see how are we showing up, then it's giving recommendations on how to improve. So, it's literally teaching our team what needs to be done. So, those are two very simple ones. Another very basic one that we like is we run sales. Sales start at 12 AM Eastern time. Well, do we want someone else to go check to see if it's live? No. The AI is going to check and see if it's live. And then it's going through the site scanning, and if it's not live, it's going to shoot us an email at 12:01 AM saying, "Hey, guys. You need to fix this really quickly." So, while our team has to be online to be able to make the change if needed, it doesn't mean you have to go and constantly be checking and making sure everything's in place. So, those are three really quick ones. A few others that I like, just make sure you have everything. Influencer identification. So, it looks at all the influencers we're using and identifies them once a month, and it gives us a list of 500 influencers to outreach. So, then our head of content creation takes the list and DMs them directly. While we could have AI write and do the DM directly, we do like that human element. We still think there is a level, especially with influencers, people being able to feel. Another one, and so far, I've been speaking strictly on the digital side, but it's also impacting our wholesale side as well. We put in all of our stores that we're working with. It's analyzing those stores and seeing what stores are similar and then pulling who's the buyer for those stores, the location, then assigning reps based off whose territories our reps have and giving that list to the team. So, those are just a few. We have probably another 50 or so different tasks being created by AI, and that's just on the workflow side. Then we have Hazel, which is from Hazel, and that is for the strategy side. So, this is a totally different element, and we run different, basically, analyses through them. One example is figuring out how much to charge for products. We're uploading all of our data into Hazel. Now, they're connected to Shopify, Okendo, etc, but we are taking other data that we're collecting from people, putting it in, and then having it analyze what should we be charging for a product. What's the difference between $40 and $42? $2 in incremental profit if we're gonna sell out the product anyway. So, it's identifying what products can we sell at a little more money. Inversely, it could say, "we project you to sell only 30% of your inventory if you're at your current price." We recommend lowering the price by $5, $2, $3, whatever that number may be. And then on the buying side, identifying, okay, based off all your historical data, you should buy 2,000 pieces of this item, 1,500 pieces of this item, and really analyzing that. Especially in apparel, inventory is the biggest risk. If you overbuy, that's deadly to your P&L because you paid to get the product and you didn't move it. One last example is figuring out what type of products to bring in. As I mentioned earlier, we do a 100-plus releases a year, and we have to make those guesses. So, what we're doing is we're describing the products we wanna release. We put it into Hazel, and it says that's a good release, or that's a poor release. And, obviously, if it's a poor release, okay, we're not going to invest the energy to, a, fully design it, b, work with production, and then, c, take the risk of buying that item.

[00:11:46] Midroll: If you're listening to The Ecommerce Toolbox, you're entitled to a podcast-exclusive website audit. Go to noibu.com/podcast-audit for a free scan that uncovers the hidden friction blocking your conversions and shows you where you're leaking revenue.

[00:12:00] Kailin Noivo: Out of curiosity, are you using multiple models, or are you using a single model in your layer?

[00:12:06] Robert Varon: Yep. So, for Hazel, it's run on Anthropic, and it's specifically trained for business, which I think that's why it's able to be more efficient than just going into any of the Anthropics by yourself; they did the work to train it to work for ecommerce businesses. Where with Relay.app, different tasks, we use different AI because we test it. We say, okay, run with Gemini. Run with Perplexity. Run with all three of the Anthropics. Run with ChatGPT. We see which one gets the best result in the trial run, and then we take the winner and apply it for future and mainstream goals.

[00:12:47] Kailin Noivo: Have you had issues with conflicting systems of record when it came to automation? So, for example, some people, their ROAS or their ad spend lives in Meta. For other people, maybe it lives in Shopify or a third-party tool. Like, where, yeah, like, have you had any issues with, like, conflicting systems of record, if that makes sense?

[00:13:08] Robert Varon: Sure. Very simple. So, we use WorkMagic. We're at client 6, give or take, and that is our source of truth. We use their incrementality model, and whatever they say goes. So, if they say the incremental new customer ROAS is 1.6, then it's 1.6. It doesn't matter if any of our other models say it was 2 or 1.0 or 3. We have one source of truth, and we also take that information from WorkMagic and pass it into our models so that it knows, a, this is the source of truth, and, b, that it can learn and optimize on top of that.

[00:13:52] Kailin Noivo: Have you guys had to implement a data lake to make any of this work? That's something that I've had to hear from a few clients. I'm curious about how you guys got around that.

[00:14:00] Robert Varon: Sure. Not really. A lot of it is just passed via API, and then integrations with Google Cloud Storage.

[00:14:09] Kailin Noivo: Okay. So, you guys do store and timestamp data in, like, Google Cloud. Okay.

[00:14:14] Robert Varon: Yeah. But we're not making, like, it's made for us. We're not going out of our way to make anything or make any adjustments.

[00:14:21] Kailin Noivo: Gotcha. That's really cool. No. This stuff's all pretty fluent, so I'm always curious to hear. It's interesting that you guys went to a third party for the agent orchestration layer. That's been, like, a big kind of question that I've had because we've seen, and some of our customers, they do it out of their model, and then some people do it in their apps. Do you do any agent orchestration at the app layer? What I mean by that is, like, log in to the third-party tool and create the apps there?

[00:14:44] Robert Varon: No. Ultimately, we focus on these other platforms because they're trained with just all the data for business. And one thing that always comes up is privacy and security, and that's the reason why a lot of them do build their own apps and work from there. Ultimately, we have trust in these platforms, whether warranted or not, but I always say, is it any different than someone getting the data from your Shopify? It's probably just as easy to hack Shopify as it is these multibillion-dollar companies.

[00:15:21] Kailin Noivo: No. This is really, really interesting. What have you wanted to work on with AI, but, like, it's not there yet in your opinion?

[00:15:28] Robert Varon: I still think model generation, we're probably 70% of the way there, so just for content creation. And it's different for different brands. You take a basic tee brand. I actually think it is there already. But if you go to Jordan Craig and see our products, there is a lot of nuance to our products. While we do have scalloped tees, which are single-color tees, for most of our products, we have patches or we have text. And with the model generation, it's just not there. And when we get to a point where we can significantly mass-produce content, that just helps everywhere. And it's not only static; it's video as well. We are on a lot of advertising channels. We're on Meta. We're on Google. We're on Reddit. We're on Pinterest. We're on TikTok. We're on Snapchat. We're on Quora. We're on Taboola and Teads and all these other channels, Spotify, AudioGo. We are really diversified with our marketing mix. And if we could have a thousand creatives, that's just going to make it more and more likely that we're successful. So, ultimately, I'd love to see AI where it's truly at that level that we need. We're less than 6 months away. Where we were 6 months ago, to where we are now, and we've tested with a lot of vendors. And again, on basic tees, when we have them just do our scout tee, they're there. They can do it. But we need that extra piece. So, I think we're 6 months away, if not less, especially with how quickly AI is growing. But once we get to that point, it is just going to be massively scalable. While we do love AI and we are very much AI-forward, we're still a human-first business. So, we do wanna support real people, real employees. So, do I think we're going to fire people for the sake of AI? Absolutely not. Ultimately, that's just the history of our company. We've been very human-forward our whole 37-career-years. But is it going to be okay, maybe instead of hiring an additional two people? Yes. That's definitely the case. So, I think it is meant to enhance everyone and make it so that everyone should be performing at 200, 300, 500% more efficiently than we are now. But we do really care about people. So, is the difference of going from 9 people to 7 people worth it? No. We'd rather just be the 9 people.

[00:18:20] Kailin Noivo: Yeah. Makes sense. I also think, like, you guys are probably in a place of privilege where you don't have 200 people. Right? It's different if, and we're similar. Like, we're 100 people, but we're a software company. We don't have 18 people on content writing. If we did, then it would potentially be a more awkward conversation. So, I think what you're saying makes sense. I think the other side of it is just, like, factually, if you're a part of a massive bloated org, it's gonna become tougher and tougher to justify over time. But, I mean, yeah. And at some point, 90% of the population worked in agriculture, right?

[00:19:00] Robert Varon: I've been to all the, I go to all the big conferences, especially the ones in New York, and in 2022, the messaging was, "AI is not gonna replace jobs." In 2023, the messaging was, "AI is not gonna replace jobs." In 2024, the messaging was, "AI is not gonna replace jobs." And I said, even before 2022, that that is a very nearsighted opinion. And like you said, yeah, there's companies with 5,000, 10,000, 20,000 people that a lot of those jobs are going to be replaced. And I think people have to learn AI, learn how to prompt, and learn how to make themselves more valuable in the age that we're at. 100, 200 years ago, I don't know how to farm; I would not have been valuable, but I think people need to know where we're going, how to adjust, and how to continue to grow. And I think that starts at the school level. It needs to be taught very early on, just like general business should be taught earlier on.

[00:20:04] Kailin Noivo: Couldn't agree more. And, ultimately, we're already seeing it with our product. As soon as you introduce an MCP, new opportunities kind of come up. Do you know what I mean? Like, okay, well, who's gonna build the agent orchestration layer? Who's gonna actually manage the agent orchestration layer? Where are you gonna do that? Is it in multiple things? If someone on your team leaves, like, are those agents centralized? Like, I think there's a lot of derivative opportunity that gets created off the back of new workflows. So, it's really, really, really cool to see that. Has AI impacted, like, rebalancing your ad spend? Like, is that something that you guys have kind of looked at?

[00:20:42] Robert Varon: Sure. Yeah. So, we are running the models especially to figure out what we should be doing, and what channels are we overspending. Now, it is a little easier because we have WorkMagic and we have a source of truth, so we still have that human element. Okay. Should I be spending more? Should I be spending less? But where it really plays a role is in predicting future spend. So, I'm thinking about launching this ad. What are your thoughts? And it will analyze the creative. We project this ad based off all your historical data to have a 1.2 incremental new customer ROAS. And that data is taken from WorkMagic and applied to their model. Okay. 1.2, that's not profitable for us, so we're not going to run it. Let's try this ad. We project 1.8. Okay. We're going to run it. If it got a 1.7 to 1.9, okay, the model worked. We don't expect perfection, even in predictive analytics. A lot of time, there's plus E for things we don't know, things that are impossible to foresee. But, ultimately, if it's 1.7 to 1.9 and it projected 1.8, great. If it is 1.2 and it projected 1.8, okay, take a step back, speak to the AI, see where it went wrong and try to readjust and recap, basically get the model to work.

[00:22:07] Kailin Noivo: Do you have any AI blunders? I'm happy to share one. But with every "move fast and break things," there's always a break thing.

[00:22:13] Robert Varon: Sure. We've been pretty lucky that we haven't had anything break. The main reason is that we have a lot of, we still do have human-in-the-loop, so occasionally, an AI will hallucinate and be like, "Going back, you should charge $300 for this product that we should be charging $50 for. Please go back, reanalyze this, and correct yourself." And then it'll correct itself very quickly. "Oh, it should actually be $47.50." Great. Why did you tell me $300? And it'll come up with this logic. But, yeah, if we would have just given it full autonomous control, it would have listed the product at $300 on our website. So, I would say, no. We haven't had any true hiccups. We've had the opportunity for hiccups, but they've all been caught before it's gone out.

[00:23:07] Kailin Noivo: Yeah. Okay. I have a similar story. We all at leadership write, like, a weekly update.

[00:23:13] Robert Varon: Yep.

[00:23:14] Kailin Noivo: And I pull from my recorded calls, and one of them was with my, like, business coach. So, there's some advice I was seeking from dealing with the situation, and that was in the update. And I pulled it before it went out, but that would have been really bad.

[00:23:31] Robert Varon: For sure.

[00:23:32] Kailin Noivo: So, there's random stuff like that, if that makes sense, where I feel like, with you guys, maybe it could leak your profitability or could do something really, some IP or something. So, no, it's important to have a human in the loop, but it's good. I mean, guys, with every move fast, there's always a break things. So, kind of comes with the territory. It's gonna come with the territory, in my opinion.

[00:23:53] Robert Varon: And going back to what we were discussing a few minutes ago with the new roles that are coming out, I think human-in-the-loop is going to, you're going to see people that that is their job title, 'human-in-the-loop', where their job is just to double-check AI before anything goes live.

[00:24:09] Kailin Noivo: Yeah. Exactly. Whether you call it agent orchestration, the human-in-the-loop, actually, it's interesting. It's gonna bring like one thing, and so the last question, because we'll wrap it up after that. Are there any new workflows that you always wanted to implement but were resource-constrained to do? For example, personally contact, or every person that writes a review, or I don't know, like, something to that range. But have you been able to implement any new workflows that are maybe a bit more enterprise given that you have, like, operational leverage?

[00:24:42] Robert Varon: Ultimately, I'd say there's nothing we've run into yet that we haven't been able to do, excluding the content piece, which I already talked about. And because, ultimately, these models are really advanced, and we're at like a crawl, walk, run stage. And I feel like we're in the middle between walking and running. I guess you could call it jogging. So, there's nothing so far, and that could just be because we have really great tools. With the exception of content creation, there's really been nothing that we haven't been able to do. Now, some projects have taken a lot longer than others, where we've had a project that took probably 60 to 80 hours of just continuing to work with the AI, and it would get 5% right, then 10% right, then back down to 7% right, where it tried to overcorrect. And like, legitimately, two weeks of my whole work week were only focused on a single task, and then it worked. But, yeah, that's not super sustainable in the long run. So, while I will say there's been nothing so far that we haven't been able to do, there have been tasks that have disproportionately taken up too much time to accomplish.

[00:26:04] Kailin Noivo: Okay. Very cool. I'll give you one anecdote, and then we can sign off. My brother runs about a million GMV Shopify store. It's actually growing quite nicely. And he's always wanted to leverage, like, a heat-mapping tool to optimize his pages and actually write abandoned cart emails. So, imagine you had an intern, and they just looked at the abandoned cart session. And then whatever they hesitated on or whatever went into the email. So, actually using our product, he built, well, I helped him a lot, but through our MCP, he built a skill that connects through Shopify, where basically every abandoned cart, it looks at our session replay, our heat-mapping data, our error data, our performance, it looks at all the marketing campaign and what they did on the site, and then it'll actually create a personalized abandoned cart versus, like, the generic one. So, that's kind of like a, I don't know if that's even an enterprise workflow, that's more of like a labor-constraint workflow that he was able to, which was interesting because that's kind of one of the first times I've seen an enterprise workflow being brought to the SMBs. I feel like there's probably tens or dozens of those use cases. So, interesting.

[00:27:15] Robert Varon: Absolutely. Like, one thing we do do is we take all of our Clarity, Microsoft Clarity, videos, and we run those through our AI models to see and learn where are people stopping and where are they getting, but I think it's very cool to then take it at the personalized level and have that applied directly to their cart.

[00:27:36] Kailin Noivo: Exactly. Alright. Rob, this was a great combo. Thank you so much for joining us, and yeah, this is a great episode.

[00:27:42] Robert Varon: Thank you so much for having me.

[00:27:44] Outro: The Ecommerce Toolbox Expert's Perspectives is brought to you by Noibu. To find out more about Noibu and how we can help you debug your ecommerce site and rocket your revenue, visit www.noibu.com. That's n-o-i-b-u.com. And then make sure to search for The Ecommerce Toolbox Expert's Perspectives on Apple Podcasts, Spotify, or anywhere else podcasts are found, and click subscribe so you don't miss out on any future episodes. On behalf of the team here at Noibu, thanks for listening.