 Hi everybody, welcome back to UiPath Forward. Six, theCUBE has been every forward except the first one, two through six. We're excited to be here at the MGM Grand in Las Vegas. My name is Dave Vellante. And we're really excited to announce the AI 10 award winners at UiPath. I'm here with Maureen Fleming, who's the Vice President of AI and Automation at IDC. Maureen and I were judges in the competition. Sid Shah is here. He's an automation leader at Intel Jeff Rittner, who's the Chief Trade Officer at Intel. And we're going to talk about their use case. First of all, folks, welcome to theCUBE, all three of you, and congratulations to Intel for one of the top award winners. And we're going to get into your challenging use case, but pleasure to have you here. Thank you. Thank you. Thank you very much. It's great to be here. Very exciting. Yeah, so Maureen and I were judges along with Adrienne Bridgewater, who is an independent journalist. And then Dr. Ed Chalice, whom you guys know, and Palak Kadakia, who's the VP of Product Management at UiPath. And essentially, Maureen, describe what the judging entailed. Yeah, what we had to do was we received a whole bunch of nominations. And we had to look at their objectives, what we were trying to get out of the automation, and then the details about the implementation so that we could understand what was being done. And then the metrics are the outcome of the automation. So we had to score on each of these different categories and the scoring was interesting because we had to compare it with everyone on almost like a relative scale. So it was a great, the use cases were all really great. Yeah, it was quite simple. You had this, it was basically the objectives, the implementation, and then the results. But then in a scale of one to five, and then as you know, Maureen, you would go through it and say, hmm, I actually think this one was more interesting so you'd have to go back and jury rig the scores. But- Well, no, it just changed them to weight them. Yeah, we did self-weighting. That's what I'm telling myself. But so the use case here was international trade, the international trade function. Yes. So, Jeff, given that that is your area of expertise, can you explain what that is and why it's so challenging? Absolutely, yes. I run Intel's global trade organization. We have 130 people around the world. And our job is to make sure that we get the freight, whatever we're shipping across the border quickly, with the least amount of cost, and efficiently. And it's a huge challenge if you think about on a given day, the number of transactions that cross a border in every country, every port around the world. And at Intel, we're moving freight all the time. And so I have a team that have to look at everything and figure out what's the right classification, what's the country of origin, the valuation, and so on. And so we have people that do this all day long and it's very manually intensive, working with our systems and so on. So the use case really was an attempt to automate what we spend so much time doing manually. And so that's the exciting part. And is compliance part of that? You know, no chips for you kind of thing. Is that part of the deal? Sorry, you guys, had to get that in. But yes, but that's an important piece, right? You can't be shipping things that aren't allowed to be shipped to various countries. Absolutely, there's the efficiency piece, but there's also the compliance and making sure that what you shipped, it's accurately declared. It said, so how does, how did AI and automation fit into this? So basically what we did is took our finished goods and non-finished goods products and we had historical data for what we call them as HS codes or it's a tariff code to import any merchandise into a country, right? And that's decided by the government. We took 100,000 finished goods products, 100,000 non-finished goods, put it in a machine learning model and then we started our metrics to answer your question. We set it at the bar really high to 95% threshold should be our minimum requirement. For finished goods, putting it that in the machine learning model, we got 99% accuracy and for non-finished goods, we got 97% accuracy. I think the key challenge here was the HS code keeps changing. So the government keeps changing the codes. The minor updates are two to three months, major updates six months and a complete overhaul every four years. So now I think I have historic data, I build a model, great. But the codes change, how do I take care of that? So we build a retraining pipeline where the model gets retrained, we have an ingestion layer where we put the new codes, the model gets retrained every quarter and that's how we maintain that accuracy. So the goal, the objective if I got it right was 80% of the products get auto-classified to 95% accuracy, so smart. You sandbagged a little bit, under promise, over deliver. You came in at 99% after 16 weeks and that was 56,000 products. The current numbers are, we've been live for 25 weeks and in 25 weeks we've done 35,000 products at 99% accuracy. So in a week about 1400 products is what we are classifying. And this is main transformation instead of, like Jeff said, instead of manual work, it's machines, we're transforming how we're doing the work, so. Now Maureen, you and I were talking off camera and you were saying that you really like this use case because you actually had some domain knowledge about this space and so your degree of difficulty radar was on. So explain why it's so difficult and why this resonated with you. So first off, it was more tied to volume, the scale of what was going on. The other piece was the difficulty of clearing customs and then on top of that, I also was thinking about sort of lead time, you know, the metric, you know, I don't think that that was part of the thing but I was thinking about lead time. Anything you can do to speed up and get as many goods to the destination as possible on time has an impact on the business, the cost of the business. So I felt like, you know, that was a pretty good use case. And are there other sort of examples or patterns you're seeing in your research, you know, trends in AI that are sort of relevant to this discussion that you can share? Yeah, another winner, I think it was Omega Healthcare, they were doing this, it's the same sort of thing but for medical recording, medical records coding. And so for them that had, again, I gave them a really high score because they were able to improve the speed of doing the coding by I think 97% and improves accuracy and they also improve customer satisfaction. So to be that whole idea of taking a big reference database that has to be done manually because you have to have knowledge and suddenly it magically is done in an automated way. It doesn't lend itself to rules, it has to sort of be AI centered. So, you know, that I'm seeing over and over again are like ingesting documents with regulations and then tying those to financial performance. There's a lot of use cases that are tied to complex codes, complex regulations, that sort of thing. Let me read off the winners. Banco, Progetto, hope I'm saying that correctly, Cox Enterprises, Equifax, Expeon Health, Flutter UK and Ireland, Intel, of course, Lazard, Omega Healthcare, Omers and USI Insurance Services. So quite a mix of industries that probably all Intel customers. So, Jeff, where do you want to take this? Expand the number of countries, products and what's next for you guys? There's several areas that I think we could really utilize this technology. One of them is in the area of export controls. So because of export controls and sanctions, every single customer, you have to make sure they're not on a list, that they're not sanctioned. You have to sometimes even look at who owns the company. And so we do all that through human glue, looking at lists and so on. So being able to take this technology and have the machine do it, we know it'll be accurate and it'll go quickly. And so that's one area. Another one is export licenses. We submit export licenses routinely. And to have a program that would take all the data, put it together for the actual license and submit it, I think we might be able to take advantage of this to do something like that as well. So those are just two examples of things we can do. And so where are you guys, thank you for that, Jeff. Where are you guys in your automation journey? It's not like you started in November of 2022 when ChatGPT was announced. You guys have been doing AI for years and years and years. But automation and AI are sort of converging. Explain sort of where you are in the maturity model. So in the trade group, we initially started with just doing unattended automation. So our core is SAP GTS. So this is Global Trade Services. And of course, people are doing resources of performing transactions. So the first go-to, the easy one was hey, replace a human by virtual human. So we started there. Then we started expanding into AI. So we did machine learning project to predict the HS classification. The next one is more on digitization. So we have a lot of supplier invoices and we need all that data. So using UiPAD document understanding, we are extracting that data and we've done a pilot where we've extract that data and accuracy of that is also 86%. And we tend to use that data for customs declaration. After that, our next journey is on a GenAI POC. So we are looking for newer use cases for GenAI. And where we tend to see a rare use of, I wouldn't say rare, but not so known use of UiPAD is in audit space. Because like Jeff mentioned, we are so heavy on compliance, all post transaction audit, we tend to use automation there. So it's in different zones that we've done and it's a journey. So we're still evaluating different use cases. And the classification prior to this was done predominantly manually or all manually? All manually. Wow. So, give you an answer. I get it now, I'd like to change my score. I still had you guys a high score, but it's off the charts now. To classify one request, a trade analyst would take five minutes and we get around 1400 a week. And there were 11 people sitting manually and doing this and think the volume's increasing. Now that this machine is there, even if the volume keeps increasing, we don't need to put more resources to do that work. It's the machine doing it. So that's, I think, was a real transformation for us. But so, Jeff, I got to ask you, were you like the cobbler's kids? Because there were certainly ways to automate classification. Admittedly, I mean before AI, admittedly it wasn't that great, things like SVM and probabilistic latent cement, math basically, and you chose to stay manual for a while. Why? Because you needed that level of accuracy, I presume, right? Yeah, accuracy is important. I want to be clear, we do have SAP GTS and there is classification that's not manual then. We also have what's called an integrated shipping memo. It's all the non-revenue junk, right? And that's got to be, you got to have the same classification and we are able to, we were able to automate that about 80% because you can pre-define descriptions and then attach a classification. But then there was always 10 to 20% that kicked out and that's where you have people spending time. Because it's an odd item or something that's not clear, so you have to go look at the book, figure out what it is. So I think this is the value of taking that percentage from maybe 80 all the way up to the 98%, 99%. And the astounding thing is the timeframe in which you were able to achieve it. And we're just, early innings is sort of the bromide, but it's true. I'd love to hear from each of you where you see the potential of this type of technology, specifically as it relates to sort of automating some of these, forget the revenue generation for a moment. I mean, this is profit generating as it drops right to the bottom line. But it seems to me, Maureen, that we're going to first see sort of improved labor productivity. You heard Bernie Olson today in the keynote. I was blown away by his stats. He said 3% productivity improvement in the relatively near term, and he'd be disappointed if it wasn't higher, 4% or higher. If that's true, there's hope. What do you think is the future? The future. I mean, I think that we're going to just see this evolution of capabilities. And so people will, in some sense, some people are trying to think through what can we do quickly? And other people are saying, how can we do something super high value right away? And so as it progresses, different organizations will be harnessing in different ways. And then the foundation models and all the things that you can do with this stuff will change. I think one other thing that we're looking at is the multimodal aspect where you do, you combine image and language, English or the text and things like that so that you can start having a more holistic way of looking at how you apply AI to different types of use cases. Right now everyone's scouring for use cases and proofs of places for their proof of concept. And so some of them are stumbling on amazing results. That demo was pretty powerful today. Basically reading some really illegible, because it was Kanji characters, they're illegible to just a mono language-speaking American. You just ingested the itinerary and said, this is the hotel you're staying at and had all the data in seconds. That was quite remarkable. But Sid, what's your vision for, specifically where you want to take this at Intel? I think more and more generative AI because we've been there, done that with automating operations. And that's where if you look most of the unattended automation, 75% of automating work is in operations. We've done that. And I think every company's doing that. But how do you take it to the next level with Genai AI? And I think that's where we're trying to figure out. And that's all dependent on data. Like unless you have clean data, you have modernized your legacy, I think even Genai won't be able to help if you don't have clean data. So focusing on where we can take these use cases in different areas of trade, going forward, that's where my vision is. Yeah, it's definitely one of the top use cases you see for Genai AI is helping improve data quality. Jeff, we'll give you the final word. Yeah, take us home. You know, I'd like to use an example of trade like a coin. And there's two sides to it. One side is all the work that it takes to get prepared to actually ship. And that's what Sid has been working on, automating that. The other side of the coin is you've got to get all that data to the government. And if governments aren't ready to receive it, then you're still going to be subject to paper. So my vision as I look at this, I'd like to see a day when from the moment it's ready to ship to the moment it gets to the actual end user, it's digital. And there are no fingers, nobody's touching it, everybody gets access to what they need to see, when they need to see it, it's powerful. That's what I envision. That's a fantastic vision. Guys, again, congratulations and thanks so much for telling your story and spending some time with us. And Maureen, thank you for all the work that you put in. I mean, as you know, when you saw the results in terms of the submissions rather, very impressive. You guys put a lot of work into this. I think, you know, I'm not a big fan of participation medals for kiddie sports, but everybody should get a participation medal in this one because there was some really good, fine work being done. It was hard to judge. Yeah, it really was. And I'd love to take another shot at it. Well, thank you again, guys, good to see you. Thank you so much. Okay, thank you for watching. We're going to take a quick break here, but we'll be back at UI Path Forward 6 from the MGM in Las Vegas. You're watching theCUBE.