Transcript 0:00 AI will transform procurement. That's just great, but which kind of AI? Doing what exactly? If you're tired of hearing about artificial intelligence with no details, this one's for you. Let's get into it. 0:13 [upbeat music] Welcome to ProcureTech Unpacked. I'm Joel Conin de Merce. 0:27 This is the show for procurement professionals who wanna cut through the marketing and actually understand the technology shaping our function. 0:35 Today, we're unpacking the June ninth, twenty twenty-five edition of the Pre-Procurement Newsletter titled "Ten AI Sub-domains that Actually Matter in Procurement." 0:44 I wrote it because I kept seeing the same pattern over and over again. Vendor decks, conference keynotes, press releases, all drowning in the word AI without ever explaining what's actually happening under the hood. 0:56 That was a pet peeve of mine then, and twelve months later, it unfortunately hasn't gotten any better. When a word means everything, it starts to mean absolutely nothing. So let's get specific. 1:08 My working definition of AI is that it's any computer system that can perform tasks that simulate human intelligence, like recognizing patterns, making predictions, understanding language, optimizing decisions, learning from experience, et cetera. 1:24 So here's a quick test that I came up with to figure out if a piece of content or a conversation about AI is going to be useful. 1:32 If you can swap the word artificial intelligence for computer or technology in any sentence and have it still make sense, then it means that that piece of content is not telling you anything that you don't already know. 1:46 Keep that in mind the next time, uh, you meet the term artificial intelligence out in the wild. 1:52 I'll illustrate what I mean by detailing three types of AI or sub-domains that procurement pros should know about, plus something important about where AI is headed in my humble opinion. All right, so let's go. 2:04 Number one, natural language processing or NLP. This is how machines read and understand human language. So contracts, supplier news feeds, emails, unstructured text of any kind. 2:17 It, it handles text analysis, sentiment analysis, translation, name entity recognition, et cetera. 2:25 And modern NLP uses transformer architecture, which is something you might have read about or heard about in the context of large language models or LLMs. 2:33 And it's the same foundation that powers those very models like Claude, Gemini, Copilot, ChatGPT, et cetera. 2:40 But NLP as a category of AI has been around for much longer than since the ChatGPT moment of three and a half years ago. The strength of NLP is its scale. 2:52 So NLP can process thousands of unstructured documents consistently in no way that a human team can match. 2:59 However, there are weaknesses, and that's nuance, specialized jargon, uh, ambiguous clauses, domain-specific terms. These things trip it up over time. 3:09 Uh, it remains a statistics-based type of AI, which means that mistakes can creep into the results. So on the procurement side, what is this applicable to? 3:19 Contract intelligence, auto-extracting key terms, obligations and risks, uh, doing sple-spend classification from unstructured descriptions, and supplier risk monitoring across news and public filings. 3:32 If contracts are sitting in a shared drive and nobody really knows what's in them, NLP is a great answer to start or a great tool to start playing with. 3:41 Now, uh, NLP reads language, but what about the data that isn't text at all? For that, there's computer vision. That's our second one I wanted to cover today. 3:51 So this is the AI or the type of AI that processes and interprets visual information. So images, videos, documents, physical objects that you capture on a camera. 4:02 The techniques used in computer vision include, uh, convolutional neural networks, I always trip up on that one, uh, object detection, OCR, which I'm sure you've heard about, or optical character recognition. 4:15 So it can classify images, detect defects, read printed or handwritten text. The strength of computer vision is, is that it handles visual data at a scale and consistency that humans physically can't, right? 4:28 So I always use the example of, like, scanning multiple X-rays and being able to tell, you know, what a broken arm is versus what it's not based on training data. Uh, however, that's sort of its weakness. 4:39 It needs large volumes of labeled training data to perform well, and it struggles with images that it hasn't seen before. And so from a procurement standpoint, where is this, where is this useful? 4:50 Well, automated invoice processing, which is where you've heard about OCR and computer vision in the past, I'm sure, where it reads and extracts data from scanned documents, uh, to produce electronic versions of those invoices. 5:02 You could also use it in quality control for manufacturing if you're producing the same widget over and over again, and you wanna be able to, uh, to check them visually with a camera as they're coming off the production line or as you're receiving them from an, an external supplier. 5:16 You could do receiving as well, right? If you're receiving based on quantity and you can have a visual, uh, way to spot that quantity, you could potentially do automatic goods receipts. 5:27 So when you're dealing with physical goods or papy-paper-heavy processes, computer vision is worth understanding and taking a look at. Now, we've covered language and visual data. 5:38 What about catching things that go wrong? That's the third one I wanted to cover here, anomaly detection. 5:44 This, uh, is being able to establish what normal looks like in our data and then flagging anything that deviates from that significantly. 5:51 It is also a statistical method that uses things like, uh, isolation forests, autoencoders. That's not important. What's important is that you know that this is what's happening in the background. 6:03 It can run supervised or unsupervised depending on the use case that you're doing, right? And that-- but that's also the case for many of these other sub-domains. 6:11 The strength of anomaly detection is you're catching rare but critical events that rule-based monitoring will miss entirely. Uh, and then the weakness is false positives. 6:20 So if you configure thresholds poorly, that can lead to a l-alert fatigue, and your team may stop paying attention to the tool that's using anomaly detection because it's giving them stuff that's not a problem, uh, you know, very often. 6:34 From a procurement standpoint, this could be applied to adverse media monitoring for your suppliers, scanning the web, pulling news stories that might be of interest, fraud detection on suspicious purchasing patterns based on, you know, either duplicate payments or things that are, are happening, happening in your workflows that don't happen usually. 6:52 Uh, and that's the s- the, the third piece which is-- or third example which is process monitoring to catch workflow breakdowns before they escalate, right? 7:00 If, if, you know, a req is bef- has to be created before a PO and somehow a PO is being created without a req and you have that business rule, right, that's an anomaly. 7:09 We could flag that, uh, based on your business process if we're plugged into the system that's, uh, and, and the database tables in the back end and the, the timestamps at which things are being created. 7:18 So that's a lot of how process mining works as well. The anomaly detection in process mining tools, to be precise. 7:26 So these are three examples of types of AI, and depending on the problem you're trying to solve, you'll see that they're either great or they're not so great. Uh, but that's why this gets interesting 7:37 because large language models like Claude, ChatGPT, Gemini, that's what most people picture when they hear AI today. And yes, they're genuinely powerful, but what makes them powerful is that they're not just one thing. 7:49 LLMs are built on natural processing, which is what we outlined in our first point here at their core, but modern AI systems layer multiple subdomains or types of AI on top of that foundation, depending on the problem they're trying to solve. 8:03 So if you give it an, an invoice, uh, to read, it's gonna call computer vision type skills, uh, at-- from the language model to try and do that task. 8:13 If you need to flag unusual payment patterns in a conversation, you're-- it's calling ano-anomaly detection, uh, to work alongside the NLP model in the large language model, et cetera. 8:24 And so the BA-- the best AI systems today are assemblers, right? 8:28 They're-- they route problems to the right method or subtype of, of, uh, AI, and then they synthesize the results through a model that can reason and respond to you in plain, in plain language. 8:38 Uh, but given that these systems can do everything, air quotations here, it means that they also won't give you the best possible performance on a smaller, more targeted use case, and that it's gonna be more expensive to get the same answer potentially. 8:51 So this is why A-- this is powered by AI tells you almost nothing when you're looking at technology. The real question is, well, which type of AI combined how and in which sequence and to solve which specific problem? 9:04 You could think of AI as every type of Lego brick ever created, and your job, our job as procurement people applying technology to a business problem, is to figure out which ones to put together depending on what we're trying to achieve. 9:19 So that's all interesting, but why does it matter to you as a procurement professional? Well, you don't need to be an expert in these AI subdomains. These are just examples. That's not your job. 9:29 But you do need to understand there are multiple types of AI, that each one has strengths and real limitations, and that when you're trying to solve a problem with technology, it's worth asking which method is actually being used, in which sequence, and why. 9:44 That's it. That one shift from is there AI in it to, well, which kind of AI is, is in it and what is it doing and in which sequence and why, right? What's the process that you're using to solve this problem? 9:57 That'll make you a sharper buyer, a better evaluator, and a lot harder to impress with a demo that looks good but doesn't answer your actual problem. 10:04 So if you wanna go deeper, my full article, uh, covers seven more subdomains, RPA, machine learning, predictive analytics, optimization algorithms, recommendation systems, reinforcement learning, and more. 10:17 Same format, strength, weaknesses, real procurement applications, and the link is in the show notes if you wanna access that free read. 10:25 The Pure Procurement Newsletter goes out every Sunday to thousands of procurement pros. You can subscribe at pureprocurement.ca. 10:32 And if you know somebody who's being pitched AI solutions right now, send them this episode. You'll give them the vocabulary to ask better questions and to get better results. See you in the next one. 10:42 [upbeat music] That's a wrap on this episode of ProcureTech Unpacked. If this one resonated, subscribe wherever you get your podcasts and sign up for the Pure Procurement Newsletter for weekly insights between episodes. 10:55 If something we covered sparks a question or an idea, we'd love to hear from you. All the links for the reports we discussed are in the show notes. We'll see you next time. [upbeat music] Resonate.
