Interview: AI Can Change the Scale and Impact of your Digital Shelf Program Right Now, with Steve Engelbrecht, Founder and CEO at Sitation

AI Can Change the Scale and Impact of your Digital Shelf Program Right Now, with Steve Engelbrecht, Founder and CEO at Sitation

Transcript
Peter Crosby:
Welcome to Unpacking the Digital Shelf where we explore brand manufacturing in the digital age.

Hey everyone. Peter Crosby here from the Digital Shelf Institute. Okay, yeah, so you need to pay attention to this AI thing. Steve Engelbrecht, Founder and CEO of Sitation and his team have developed an AI tool that is capable of generating 15,000 product descriptions per hour for their enterprise customers. Connected with their PIMs and the workflows to keep humans in the loop. And that’s just the tip of the iceberg. What does all this mean? Where is it going? What are the risks? Is this a job killer? Can it write my podcast intros? Steve provides some answers to all this and more in this episode.

Click and Subscribe on these popular channels.

Okay, yeah, so you need to pay attention to this AI thing. Steve Engelbrecht, Founder and CEO of Sitation and his team have developed an AI tool that is capable of generating 15,000 product descriptions per hour for their enterprise customers. Connected with their PIMs and the workflows to keep humans in the loop. And that’s just the tip of the iceberg. What does all this mean? Where is it going? What are the risks? Is this a job killer? Can it write my podcast intros? Steve provides some answers to all this and more in this episode.

Peter Crosby:
Welcome to Unpacking the Digital Shelf where we explore brand manufacturing in the digital age.

Peter Crosby:
Steve, thank you so much for joining Lauren and I on the podcast today. We're really grateful to have you on.

Steve Engelbrecht:
Thank you very much for having me. Really excited to be here.

Peter Crosby:
Well, I mean, I'm very excited to talk about how robots are taking over the world. We are actually going to talk about really certainly one of the most buzzworthy topics these days, AI. It's top of mind for brands and retailers. We are in this moment where everyone wants to do more with less. I often describe the digital shelf as a hungry beast. It will always want more data, more content to get better and better search results, better conversions on product pages. All of that will take richer, deep content in addition to other capabilities. And we're all trying to figure out how to do that because we're not going to put a lot more money and a lot more people at it because we have to make it profitable. So that's the world we're in right now. And I think, well one, I want to ask if that resonates with you, what the possibilities are here for a business value perspective?
Steve Engelbrecht:
Yeah, I absolutely agree. I think it's important to set the context of how big this is of what we're looking at and how disruptive this is. This is not an incremental improvement in computing capabilities or in the integration of computing into how we're doing business. This is a major, huge disruptive step. I truly believe that this will have the potential to change everything about the way that we're working in the e-commerce world for sure, really, in any information business, anything that requires a depth of knowledge or anything that's looking at large data sets, I mean, this is a revolution as big as the internet itself or the smartphone in my opinion.

Peter Crosby:
And anyone who's listening also knows the dark side of that that we've been hearing about of it's only as good as the data it's accessing and I've forgotten, I work with some really smart people at Salsify and Adam, our Chief Technology Officer is one of those, and he says, "AI lies, but it lies really well."

Steve Engelbrecht:
But I think that's way underplaying just how big this is. So certainly there's a very easy access to this, and that's one of the things that makes this so revolutionary. For the first time ever, we can talk to our computing systems. I mean, truly in natural language, we can tell it what we want it to do and what we want that output to look like. Now, you're absolutely right, there's a sort of scary gap here between what these models can do in terms of creating very believable output and whether or not we can actually treat that as correct. And already I think we're seeing a really interesting movement where people are just trusting that output. And I think it says a lot about how good the technology is, but it also points to a pretty scary thing about this is that what does it mean when we no longer check our work or we're no longer applying critical reasoning or solid business processes behind generating these outputs from these models?

Lauren Livak:
And Steve, when people think about AI to that exact point, I think a lot of people see ChatGPT as a way to type something in, get something out and be like, "Okay, I'm done. I can use it. The work is done. It doesn't need, to your point, any editing or proofreading," but it's really not the case, especially from an e-commerce standpoint because they might not have all the information. They might not know what you know, you might have to input some of the information. So how are brands using AI then, to integrate it into their strategy in a way that is scalable and accurate?

Steve Engelbrecht:
Yeah, no, it's an awesome question and we should be clear too. So ChatGPT, and other generative AI models, these large language models or LLMs, this is one of many different applications of artificial intelligence and AI of course has been around for a long time. It's already part of systems that we use and see every day. Things like your Netflix queue showing you things you might be interested in, that is AI, those are predictive analytics that are driving those algorithms. This one's just different. It's different because it's so big and it's so revolutionary and it's so accessible to everyone that you can actually talk to it and interact with it directly. So I think the easiest thing way to get started with this is just to explore it a little bit and play with it. And when somebody tells you it's a chatbot, you go in and you say, "Hey, how are you?" This is how my kids play with it. "Hey, how are you?" And all of a sudden you're having a conversation with this thing.

Steve Engelbrecht:
What we need to be careful about is confusing output, which is just spitting out what it thinks you want to hear versus how correct that is. So getting into applications of this, where I'm fascinated by some of the applications of this technology are two areas. One is where we get into computing and getting into looking at external data sources, basically taking this trained model, but then applying it to real world problems. I think that's really fascinating.

And the other is about scaling. So anything that we want to do on a repeatable process that we want to try to generate consistent output at enterprise scale, you get into some issues here. And this is, I think, a really important point. The allure of ChatGPT is that you can log in and you can type to it. You can type whatever you want. You can even talk to it now with some of these models, with the Whisper technology, they've integrated into it. You can't do that at scale. So no enterprise merchant or merchandising operations team is going to sit at ChatGPT and one by one, grab product records and pull them in and say, "Okay, make me a description, make me a description." I mean, you could, but that's not going to save any time and that's still a pretty miserable approach.

Steve Engelbrecht:
What I'm fascinated in is how we can take this core technology and we can structure how we're interacting with the technology to make it consistent and make it performant and make it grow and scale with us. So not to bury the lead here, but the application of this that Sitation is focused on is generating product content. And in particular, we're starting with the basic information that we have about a product, your master data, maybe a few key talking points that would come from the marketing team or a product team. And from that, we're generating long form content. So this might be, for instance, we've got a product title, we've got a brand, we've got a category, maybe we have a few specifications about size, flavor, whatever it might be, depending on the category. And then maybe we have a couple of features. So sticking with food, maybe we have organic, we have gluten-free, we have non-dairy, have talking points like this, or maybe something about the flavor, flavor combinations.

Steve Engelbrecht:
Just from that, those would be the same inputs that we would take for a human copywriter and say, "Okay, we're going to create a piece of content that is for this particular audience, for this channel, it should be approximately this length, it should use this information. Maybe here's some sample output from previous iterations that we want you to write to a similar style." But those are all the inputs. And then from there, a human writer is going to take those and generate a creative output. That's the point of content for us. That's where the rubber meets the road for our technology, which is to take the same inputs and to make available to the model, all of that information that we need. Here's what good looks like, here's some guidelines, here's some sample skews, here's the tone of voice and the target audience, and to let it generate that first draft of what that content looks like.

Lauren Livak:
I'm putting on my old hat here and thinking about my small digital shelf team and what we could have accomplished if we had AI to support that from a content perspective. I mean, we had to prioritize the skews we could create content for, to your exact point, but we weren't able to do more and this would really enable us to create more content at scale so it's just super exciting.

Peter Crosby:
We've been spending so much time recently talking about this next decade of the digital shelf and the headline is every one of the players in this industry need to figure out how to do what they do more profitably than they've been able to so far because of all the investments they've made in innovations to get to this place where digital and omnichannel is starting to be a tightly integrated part of the business rather than separate silos that they've been during this more experimental phase of e-commerce and digital, but now the experiment's over and money's no longer free. It's figuring out how to do all of this A, more profitably and B, drive more growth out of every shot you have at the consumer or the buyer than you've been able to to date. And that's what excites me so much about the possibilities here.

Steve Engelbrecht:
Yeah, there's a lot to unpack there. I think you're right on of what the future holds for this. So even right now, you can see the truly global excitement for this new technology and people are embracing something that really, frankly, is in its infancy, I think., That this is the first time that this has spilled from the academic world into the business world in any meaningful way that's accessible to everyday people to start to see it. But you can see the way that people are thinking about how it's going to change, again, virtually every aspect of our daily work and home lives. There's good examples out there already of how personalization at a massive scale can be really effective. I think Facebook's a good example. Some of the personalization or localization around Google results are another example.

Peter Crosby:
So Steve, I would love it if you would now bring us to the reality of what's happening. You've been working on this for, and tell me if I'm mistaken, but several months now and have brought a product to market that's integrated with a PIM that rhymes with malsify. I won't mention which one it is because we try and avoid that, but you have customers that are using this today and bringing it. So walk our listeners through an example of what you're starting to see, how things are going, what value people are seeing, what are the hiccups and what's next? Ready? Go.

Steve Engelbrecht:
Yeah, awesome. Thank you for that. Yeah, so we happen to be a major partner to that PIM that rhymes with malsify and that PIM is very popular with global brands who are exactly these users that we're talking about. So these are brands, they're manufacturers, there's retailers, there's distributors, but the common thread here is that there are lots of products. Those products have lots of data and those products need to travel to many destinations or channels. So just like everybody else that has product data to manage, there are problems that start to emerge and become exacerbated at scale when you think about anything you have to do to a product or to a category or a brand, hundreds or thousands or tens or hundreds of thousands of times, even millions of times with some of our larger customers.

And the opportunity is to think about where can we shave off some seconds and some time and some cost here in that value chain to help bring those products to market. Backing up for a second, working in the retail or distribution world, when we think about items set up for one of our customers, by far, the biggest element of time that comes in here is how much time is spent after we've decided to sell a product, but before we can actually get that product into a channel so it's visible to a customer. So broadly speaking, that item setup process needs to go through contract and set up in the ERP, it's going to get pushed over to an MDM or a PIM system, but then somebody has to get it ready for sale, and that's going to be importing information from manufacturers. It's going to be massaging that information, enriching it, making sure it's ready, all of the work to get that product data ready so that we can get it up in a way that is compelling and useful and consistent and true to our brand when that data is available to a customer.

Steve Engelbrecht:
So that is the problem that we've attacked specifically with our offering, RoughDraftPro. And what we've done with RoughDraftPro is that we've built a couple of different key technology offerings here. First is a wrapper essentially to the APIs that power ChatGPT, and other offerings from OpenAI, which allows us to add structure and consistency to how we're talking to these endpoints, and also to inject product data into those prompts so that we can basically say, "Here's a bunch of examples of what product data looks like at this institution. Here's the inputs to this and what we consider to be good outputs." Then we say, "Okay, here's some new fresh inputs. Show us the output." That's basically what it does, and it comes back and we're going to have it generate one, two, maybe three different options, present that into a workflow and help our users to decide which one they want to go with and do they need to make any changes to it.

Steve Engelbrecht:
But very importantly, we can do this thousands of times, very, very quickly. So the benchmarking that we do comes from the scaling of this technology. Because we're trying to build an offering that will appeal to a global audience, and in particular, to enterprise customers, we know that not only the accuracy and the data quality is important, but also our ability to do this quickly, efficiently, and affordably. So our latest benchmarks, these are fresh numbers from last week, is that we were able to do 1,000 products in four minutes, which is 15,000 products per hour. And this is taking all of the inputs from product data, like I was mentioning, the specs, the product title, the brand, whatever, passing it into one of our models, comparing that to all of the fine tuned data from that particular brand that says, "This is what data looks like here," generating the output and saving it.

Lauren Livak:
I just had a slight daydream of an executive going to their C-suite and being like, "I need 15,000 copywriters and $10 million dollars." Would never happen.

Peter Crosby:
But if you want to get the work done in an hour, that's the need. You need 15,000 copywriters.

Steve Engelbrecht:
What this would be is this is when the model is basically, it's saying something that isn't supported by the inputs, so it's going off the rails, again, in a very believable way. Remember, these models are meant to be believable not to be correct. So it's going off the rails in a way that makes you go, "Huh, I don't think I mentioned anything about free shipping. So where did that come from?" So we need to catch those things. To do that, we're doing this a couple different ways. Number one, we can actually guard the guards, so we can have another layer of the AI that's comparing the original input to the output and looking for anything that would not be substantiated by the inputs. But the other element of this is the workflow piece that you're talking about, and I talk about this in my blog as well, but it's actually exactly the term that you used, which is humans in the loop.

Steve Engelbrecht:
So we want a workflow that even if it is machine assisted or heavily driven by AI, there is a human in the loop element, at the very least, for spot checking, but ideally for reviewing the output for consistency and correctness before we're publishing that information into anything that would be visible to customers. And frankly, especially at our biggest customers, the legal and compliance team is going to demand it because again, that product liability aspect of what claims we're making about a product, in many industries, that is heavily regulated, it needs to go through a regulatory process. It needs to be signed off on by multiple partners. So we can't just change things and toss them out there casually and hope for the best. That's not a good business model. Instead, we want to make sure that this is part of an integrated process where we are replacing the copywriting function or maybe the first draft function, maybe the rough draft function of this.

Lauren Livak:
There needs to be people in place to guide this technology and to work with it, rather than just letting it operate totally independently with no oversight.

Steve Engelbrecht:
And we want to involve those people in the process, establish a model where we're doing all of those checks and balances to ensure the reliability of the information that is published.

Lauren Livak:
People need to be involved in ensuring that all of this information is accurate, and I think that's where the strength of the technology will be.

Steve Engelbrecht:
Absolutely, and it’s about helping users make informed decisions and optimize the overall output process.

Peter Crosby:
Fantastic. Thanks for sharing how AI can influence content creation at scale while still maintaining quality control.

Steve Engelbrecht:
Thank you for having me.