Support Your People with AI

[00:00:00] Firstly, thank you for coming today. We developed a series of webinars that we’re going to be running every few weeks, purely as education to help our customers and potential customers figure out the right pathway for AI in their business. Before we get into things, though, I’m just going to start off something here. I’m using a tool called Claude, it’s an alternative to ChatGPT. And I’m going to ask, you know, what’s the best product to clean antifreeze off a workshop floor? And we’ll come back to that soon.

[00:00:28] I wanted to talk a little bit about the philosophy 3DN have adopted around AI, but also get into some practical applications. The one thing we really wanted to avoid, though, was making this an AI showcase. There’s a lot of content already, you know, of people talking about how they’re using AI to do X&Y&Z. And I’m not suggesting that that content isn’t useful or that doesn’t have value. However, I guess what I want to do is to temper that with some of the realities of using AI day-to-day.

[00:00:55] We’re seeing a few different narratives play out at the moment. Everything from AI is going to destroy the world, like in Terminator, through to it being another tech bubble, and it’s not all it’s cracked up to be. But the two that come up most often in conversations are, one, that AI will do everything for you when it’s coming for your job, and two, that it’s broken, and we can’t trust it.

[00:01:18] At 3DN, we really don’t believe that your job is at risk anytime soon. And we also understand why you might think it’s broken, but the approach that a lot of organisations take in using it is flawed. And that kind of leads to that broken concept.

[00:01:32] The reality that we’re aiming for is that AI can help you do your job more productively, but there’s a bit of work to get there. So in a way, it’s not really that the outcome is in doubt, but the path to getting there is. One of the best analogies that we’ve landed on is a concept of treating AI like a new staff member. When you think about taking on a new staff person, clearly there’s a training exercise that you go through. The last thing you would do is sit them in front of a computer and say, here’s our website, and they start answering phone calls. And the same applies to AI.

[00:02:05] I see so many examples of people throwing three or four PDFs into ChatGPT and then copying and pasting a customer’s e-mail message and hoping the response is right. And invariably, it’s not. And you may have even tried this yourself and quickly realised that the response that you’re getting from ChatGPT, it doesn’t sound like you. It doesn’t have the right information. It’s not taking into context who’s asking the question or when the question is being asked.

[00:02:32] And in some cases, you may have simply dismissed AI as an option for your business as a consequence. So let me frame this for a second. There’s a lot of ways to use AI. Here at 3DN, we use it for application development, but most of our customers have their own customers that they’re dealing with. And one of the best candidates for leveraging AI is enhancing the way that you deal with your own customers. So today, we’re talking primarily about interactions that your customers have with you and the way that you respond to them.

[00:03:02] The way that we see the AI is broken narrative playing out is where you might go to ChatGPT and ask a relatively detailed question, and you’ll get a relatively detailed response. But if you ask the same question again, you might get a completely different response. And in fact, every time you ask it, you’re probably going to get something different. Three out of five times, you might actually get an okay result, but the other two are a mess. You know, they include information that’s just made-up. It’s not grounded in any facts or it’s misrepresented what you were trying to get to in the first place.

[00:03:35] When it comes to a business needing to respond to customers, continuity and accuracy of the answer is really important. And that’s where training the AI comes into place.

[00:03:44] Now, I also just want to touch on the term training. The reality is that we at 3DN, and most likely you and your business, will never truly train their own AI model anytime soon. Training takes millions of pieces of content, hundreds and hundreds of thousands of hours of computer time, and entire data centers. And those are resources that we just don’t have access to. And so even though I’m using the term training, it’s technically not correct. What we’re really doing is we’re augmenting the information available to language models so that it has more context about your organisation and your customers and the question being asked.

[00:04:22] Okay, so remember that question I asked at the start? Let’s see what we’ve got. If I scroll down and have a look at the result, the answer is that the right product to use to clean antifreeze off the workshop floor is Renegade. It’s an alkaline water-based solvent cleaner specifically formulated for heavy duty workshop cleaning, including coolant and antifreeze spills. Now, just to put some perspective on this.

[00:04:47] We’ve recently been working with a chemical cleaning company called Agar Chemicals. The problem they had is that they weren’t able to quickly identify the right product for their customers based on the cleaning problem that they had. What you’re seeing here is the test system that’s been implemented that they’ll shortly be using internally for sales staff and will be rapidly moving into customer queries.

[00:05:10] While we look at this, let me just explain how we got here. What I’m going to do is I’m just going to find a PDF that we used as a starting point.

[00:05:22] So, I’ve used the term augmentation, and you know what does that really mean? Well, this is the PDF that was used internally, and it’s also on the website as well, and it has a lot of detailed product information. And so when we say augmentation, what we’re doing is we’re providing all of these PDFs to the AI, and it’s trying to find information in these PDFs to help answer customer questions.

[00:05:45] In reality, on the surface, this looks like a perfectly good piece of documentation. And you’ve probably got documents like this in your business that might be membership benefits or how CPD works or a constitution that explains a political party. And we took dozens and dozens of these documents and we pushed them through our process. And we quickly realised with the customer that the sort of questions that customers were actually asking weren’t answered by this documentation.

[00:06:12] So, a new task needed to be done, and that was to review every single one of these documents and classify it in a range of different ways.

[00:06:24] The result was this massive spreadsheet. And that enabled us to take the information that was implied in the PDFs and convert that into explicit knowledge for the AI. It allowed us to enforce consistent terminology, consistent categorization. For example, the AI no longer had to guess that walls were generally covered in paint or that a workshop floor was probably going to be concrete.

[00:06:48] This actually resulted in a bit of a penny drop moment for our customer as well. They had this realisation that even though they had good looking documentation, it wasn’t as useful as they thought. They actually found that they had more than 10 products that were actually miscategorized and classified as well.

[00:07:05] Now, this is where the new staff member analogy comes back into play. Let’s dig into actually what’s happening behind the scenes when we ask this question. You may have noticed that there was a whole range of different tools that were implemented. For example, there’s a thing here called Find Paths, and there’s another thing here called List Entities by Type and something called Semantic Search. These are all different mechanisms that we’re using when we’re answering these questions.

[00:07:35] Now, firstly, the environment that I’m looking at this in is slower than it would be in the real world. It’s stepping through each and every part of a lengthy process, one step at a time, and that allows us to evaluate exactly what’s going on.

[00:07:49] The next thing is to consider what information we’re looking at. Unlike how something like ChatGPT works, where it’s referencing text stored in a massive language model, here, we’re directing the system to use something called a knowledge graph. In simple terms, a knowledge graph describes objects and how they relate to other objects. For example, my name’s Jay, I work for 3DN, and in a knowledge graph, Jay would be an object and 3DN would be an object, and they’d be related to each other with a relationship called works at or is employed by.

[00:08:25] The benefit of using a knowledge graph is it allows you to more explicitly explain how things relate to each other. In this Agar example, every product that they sell is an object in the knowledge graph. But every type of surface that you could clean is also an object in the knowledge graph. And the relationship between them allows us to determine if that product can be used on that surface.

[00:08:50] So in our example, antifreeze is an object and workshop floor is an object. And so the knowledge graph allows us to quickly identify relationships between those things. And then related to that are all of the products that might be actually used to clean antifreeze off the floor.

[00:09:10] That’s why when we look at this, we see things like find paths. This is the AI trying to find relationships between two objects. We see things like that semantic search, where it allows us to broaden the scope of what we’re searching for. A really simple example would be that we could substitute the word floor with workshop floor, and a semantic search would be able to find products irrespective of the term that we’ve actually used.

[00:09:38] If we think back to the analogy of having a staff member, you wouldn’t just point them at a filing cabinet and say, good luck. You’d organise the information. You’d explain it. You’d show them where things are. You’d highlight things that are really important and play down things that are less relevant.

[00:09:55] And in doing that, you probably find in your own documentation or your own website that there’s implied knowledge and inconsistent terminology and information that is more important at different times of the year. And without all of that, the AI will struggle. There is no magic wand to fixing that problem. However, building a knowledge graph like this takes that implied information and converts it into explicit knowledge.

[00:10:22] Think for a second about the sort of information that your business deals with. You might run events or training, and all of those are going to have common things like a date and a time and a venue. But they’re also going to have more abstract information, like a title or a description and a speaker biography. And if you’re in the business of recommending events or training courses to customers, then trying to align that abstract information with a focus area that a customer might have without having clean relationships between all of those points is going to be nearly impossible for an AI to help you with.

[00:10:58] So, what’s the end result here? Well, for Agar, they have an incredibly powerful tool that they’ll be exposing to their customers shortly that takes a really simple question, like, what’s the best product to clean grease off the floor, and evaluates the best options available and gives recommendations back.

[00:11:18] Now in this example, though, we’re only dealing with product information. What if we could add location or inventory? What if there’s a promotion going on right now for a particular product? That probably should be part of the knowledge graph.

[00:11:32] What if the customer is in Perth? The supplier over there only has a limited set of products, and so offering them something that they can’t go and buy is a bit futile. Maybe one of the products is reaching end of life, and so they want to promote that product so they don’t have to stock it anymore. These are all examples of where context of that information becomes really important. We’re talking about more than just the text and images that are in the business.

[00:11:58] Now, thankfully, the exercise of taking all of this information in your business and converting it into knowledge that the AI can actually leverage only needs to be done once. Once you’ve got that knowledge base, you’ve got dozens of applications available to you. Really, what you’re doing is training once, but then deploying everywhere. Kind of like the best staff member in your organisation who knows everything inside out.

[00:12:25] Now, of course, 3DN have expertise in building out these knowledge graphs and then leveraging them through a range of different ways in your business. And of course, we’d want to talk to you about that.

[00:12:36] But building a knowledge graph is a really big project. And if you wanted to start smaller today, there’s something that you can do right now with your existing emails. If you’ve played with anything like ChatGPT, you’re getting it to develop content for you, like responding to emails or writing documents, and getting it to sound like you can be a real challenge.

[00:12:56] I’m going to share with you an approach that I’ve used and it’s something that you can do also. Now, AI is moving really fast. New things are literally appearing daily. And irrespective of what’s happening in AI, building out a way to structure your information into knowledge is critical and never wasted.

[00:13:15] So, let’s have a look at that example. This is a prompt that I’m going to share with you that allows it to go through and review emails that you’ve sent in the past and extract from that your writing style, the terminology that you use, the things you do and don’t say, the different preferences you have for things like how long your messages are, the tone, how you format your emails, how you sign off your emails. All sorts of different things.

[00:13:45] And what we can do is we can give any language model this prompt and a collection of our sent emails, and it would then be able to determine what your writing style is. You can then use that writing style in ChatGPT or other language systems to be able to make it sound a bit more like you. So, we’ll share this with you as part of the presentation today.

[00:14:10] Our plan is to run these webinars every two weeks. This was our first. And I also wanted to thank you for coming along. And I’m going to make myself available to answer any questions, if anyone has any. Not about just this, but anything AI related, really. Thanks again.