Subscribe
Share
Share
Embed
IIIMPACT is a Product UX Design and Development Strategy Consulting Agency.
We emphasize strategic planning, intuitive UX design, and better collaboration between business, design to development. By integrating best practices with our clients, we not only speed up market entry but also enhance the overall quality of software products. We help our clients launch better products, faster.
We explore topics about product, strategy, design and development. Hear stories and learnings on how our experienced team has helped launch 100s of software products in almost every industry vertical.
If you are creating a model that isn't efficient or creating wrong data or just taking a long time, is that something that can obviously drive up costs from a multitude of ways? Yeah. You've integrated this, but it's not working optimally, and you have no idea because this is how you think it's done, but your cost is 10 x of what it should be. That's one of the things that can happen.
Speaker 2:You can have different models for different functions. So you have your natural language model that we're interfacing with your customer service, but then you have another specific model that would be able to communicate across APIs. Those 2 models can talk to each other and perform certain functions. So, really, once you've sort of identified what you require from AI, it's really about selecting those models and getting them to talk to each other.
Speaker 3:Let's understand the landscape of AI. Unfortunately, it's changing every day. When we talked earlier about all the different models that are out there, those are being updated. New ones are coming out every 2nd month. So at this point in time, you definitely have to keep your wits about you and understand what's better and what's changing.
Speaker 1:Hello, everybody. Welcome back to another episode of the Make an Impact podcast. I'm your host, Makoto Kern. I've got my other cohosts, Joe and Brinley. Hey.
Speaker 1:And so today on our, another AI podcast that we're, we'll be discussing is really more about let's say you're a company and you want to use AI. You know? Where do I begin? Where do I start using this? And, you know, we want to really help out those companies.
Speaker 1:There's so much information out there, and knowing where to start is really important because if you get started on the wrong foot, it's going to be an expensive, project or you can be really efficient, then, you know, that's what, you know, we can help with. And so Joe and Brilly are gonna talk to us a little bit more about how to get started and and all the details about that.
Speaker 2:So I think it's a
Speaker 1:good way to
Speaker 2:get started. Also about kind of demystifying the the the, you know, what is AI. I've chatted to a few, few people who've always said, well, it just seems like such a barrier. We need to get started with it. What's it all about?
Speaker 2:You know, I must need an AI expert or how can it even help me? I think these are questions that probably so many companies have but don't know where to start. And you even think about where, you know, we've tried to explore, topics that still dive into quite a lot of detail, but, you know, I I think we can extract in this podcast really just the high level items on, you know, if you're thinking about introducing AI into your into software, into your organization, what are the next steps you can take? What are some of the kind of practical examples, that you
Speaker 3:can do? So Yeah. And what kind of areas should you be focusing on, within your organization? And, yeah, we'll go through some examples of that.
Speaker 1:Alright. Joe, you wanna kick us off with some AI modeling?
Speaker 3:Yeah. Exactly. So the way
Speaker 1:we'll sort of somebody it's not an AI virtual, you know, bot that's modeling for you.
Speaker 2:Soon as you be, who knows? Maybe now it's it's a little bit more boring than that,
Speaker 3:No. Yeah. Because I'm not boring at all. Yeah. Once you do
Speaker 1:Not boring at all, but not as exciting as
Speaker 2:an AI model.
Speaker 3:So let's
Speaker 2:just make that clarification.
Speaker 3:True. Exactly. So, yeah, what we'll talk about a bit is I'll just sort of discuss kind of what AI actually is. You know, it's a big term. It gets thrown around with everything.
Speaker 3:AI is being labeled on a lot of, initiatives and projects and software. And, you know, what is AI? What actually is it underneath? And I'll talk a little technically about it, try to break it down a little bit. But after that, we'll still go into a bit more of use cases for AI.
Speaker 3:And Brittany will go over that, and Makoto will discuss some of the real world, integrations that you can look at in your company. And then I'll also talk a bit about, you know, how can we determine that? How can we look at our company, look at what AI offers, you know, once you've gone through the soul and kind of match the 2 together? So to talk a bit about AI, really, what it really is in the end is models. And there are many models out there, and they all do different tasks.
Speaker 3:And you can think of each model as its own little program by this specific purpose to it, and it does something for you. There are large, large models and they're tiny models. Like, you know, large models would be something like chattypt. Tiny models would be very specialized. You can't really communicate with them with words, but you can give them some data.
Speaker 3:They can do something with that data, and I couldn't answer few, that you have on that. And so when we're talking about models, models already just created almost databases. They're they're way more advanced than that. They can think of it as just adding in a lot of data into a very specialized sort of container, which are called models. And the type of data that you feed into it and the way it's created will sort of create that outcome based sort of solution around it.
Speaker 3:So an easy way to think of it is, you know, we wanna create a model which can understand when you give it images, is this an image of a cat, or is it an image of a tree? You know, a very simple idea. And so the way that works is that you'll get millions of images of cats. And each time you add a cats to the model, you'll tag that image, and you'll say, this is a cat. And each time you add a picture of a tree, you'll tag it and say, this is a tree.
Speaker 3:And, overall, once it's all in there and all those images are in there, what we're really doing is then being able to feed new images to it that's never seen before of a cat or a tree or something else and say to it, you know, is this a cat or a tree? And based on that understanding it has and all the data it really has inside it, it can then make predictive analysis in it and go, yes. This looks like a cat. This looks like a tree. That's really what we're doing here.
Speaker 3:And so even something like ChatTPT is really just an expansion of that. What that stuff is instead of feeding it images of cats and trees, they basically fed up the entire Internet. You can always think of it that way in terms of words and human language, and it's looked at conversations about how people have conversations. When someone asks something, what's most likely answer is from that? So when you change something like Chatupt, which is obviously a large language model, really, what it's doing is it's trying to predict what the answer to that would be.
Speaker 3:Your question and is trying to predict what would the sentence be to answer this or to sort of reply to this. It has no understanding of it. It doesn't know what you're asking. It has no intelligence really behind it. It's just predicting.
Speaker 3:And that's really what all of these models are. And so when you're looking at a solution for your business and you're looking at, okay, I have this process. And right now, we have some existing either manual process or we have some very complicated IT architecture that, you know, takes a lot to maintain. That's costly. You know, can we instead offload that to one of these specialized models that can do it for us a lot easier, a lot cheaper, and often a lot more effectively?
Speaker 3:That's one way of looking at it. Otherwise, you know, these models are just bringing out functionality, which would be extremely hard to do before. Like, again, chat gpt. Yes. Technically, you could program and write an application that could pretend as chat gpt that has no air at all, but just sort of analyzing words and trying to come up with some crazy way of sort of responding.
Speaker 3:And in the past, there have been attempts at that, but nothing like Chat GPT. So, really, AI is also opening up new doors as well as sort of being able to improve efficiency of existing ones. So when we're looking at models, there's sort of a way to perceive them. Okay. We know models can do things for us.
Speaker 3:What models are there out there? There's really 2 ways of going about it. You can create your own models from scratch, or you can use existing models that are out there. And there are loads out there. And so, really, when you're thinking about AI models, you're almost sort of browsing a shopping center of models.
Speaker 3:That's sort of the perspective I would take in it and looking at what's out there, seeing all their existing models. And there's loads of open source ones. There's proprietary ones, like, again, chattyPT and open AI models. And you're ready to browse in the category of them out there and going, which of these are best for my use case? And you can grab a few of them.
Speaker 3:You know, this this might be for your solution, several models, some proprietary, some page of open source, And you could feed your data into it and sort of see, okay, how does that work? So a very simple example would be, you know, looking at a predictive model and going, okay, I want to model it just as sentiment analysis. You know, I'm gonna feed it some text, and I want to understand what's the sentiment behind that. Is this is a sentiment that, you know, this is looking good or is it looking bad? You know?
Speaker 3:So let's say we're trying to analyze financial data. We feed in entire news article about, you know, some company that we're looking into. We can gain sentiment out of that and get a score out of that, and that's all the model does. Doesn't do anything else for you. You can't text the model.
Speaker 3:You can just feed it information. It gives you a number, and it says the sentiments around this is looking bad. And you go, okay. Great. Then you feed a 1,000,000 different news articles, and you automate this process.
Speaker 3:And I can just every day look at 20 companies and get an overview of all the public sentiments around these companies so you can make financial decisions based on that. So very specialized models around that. And for that, they're open source models that you can use. So, again, we were taking a step back from this. We can spin up a model within a day that can start doing this, and we can build some solution around it.
Speaker 3:Otherwise, we didn't have that model, you know, to try and do this sort of predictive analysis manually or to try and do it with experts or or, like, an extremely sophisticated software would take years to probably build out with experts, and it would be an extremely complicated thing to do, whereas models really simplify this. They're like a shortcut to getting to be able to spin up really interesting and powerful solutions. So, yeah, looking at that and so understanding, okay, we got all these models out there. Okay. What can they do there?
Speaker 3:Let's dive into that a bit more. And so I'll just hand over to you, Brendon, and you can sort of go over some of the
Speaker 2:Sounds good.
Speaker 3:Yep. Practical applications. I think
Speaker 2:interesting points, I think, bringing more context if someone was wondering what a model is exactly. So Yep. Yeah. Getting to to what really is possible. So we have these these models.
Speaker 2:As Joe was saying, you can choose between the number of different varieties of different varieties of them, and they have these, you know, different possibilities or different capabilities almost. There really are a vast array of of different options, but it really is important to, you know, stay practical. If your interest if your business is in a specific market or has, you If your interest if your business is in a specific market or has specific, functional requirements, you know, just, you know, focus on those. So if we look at some of the common, AI applications and functionality, we've got quite a few different ones. We'd start with probably what a lot of people think about when they think of AI.
Speaker 2:We've got these chat completion services, and, it's really excellent for things like multi term conversations or just maintaining kind of context. So even, you know, being transferred between different departments or online, sort of chats. These AI models are really good at being fed context and continuing and completing a conversation. So applications are things like customer support, really providing more real time assistance by completing conversations with customers, answering common questions, and even helping customers resolve basic issues. So you really allow your customer service department to be much more efficient and allow the AI to do a lot of the heavy lifting with the more basic issues that you may run into.
Speaker 2:So massive cost saving opportunities there.
Speaker 3:Yeah. Exactly. And just to expand on that one, the the power behind, check completion is, again, when you're having a conversation with AI, it almost feels like you're talking to a human. And outside of that, just sort of talking to it, the next step is actually asking it to do things for you. So when you're looking at an AI solution, you're looking at not just as what the model can do, but what can we connect up to that model to give us no extra functionality around that.
Speaker 3:So let's say you want to have your AI do your online ordering for you. Instead of a user having to log on to website and add things to their cart, they could just say, hey. I want to buy some large socks in in in this brand, and it'll go, sure. I've added them to your cart. Is there anything else you want?
Speaker 3:And you go, that that's it. It'll go great. Created your account or whatever the case is, and, you know, you can add on that functionality on top of these models. They can be a really powerful layer between, services that you want to offer.
Speaker 2:And this is something that we've been looking at recently as well, Joe, where, you know, looking at all those different model types, you can have different models for different functions. So you have your natural language model that'll be interfacing with your customer service, but then you have another specific model that would be able to communicate across APIs. Those two models can talk to each other and perform certain functions. So, really, once you've sort of identified what you require from AI, it's really about selecting those models and getting them to talk to each other. Yep.
Speaker 1:I've got a quick question around that. Just around if you don't set up your models correctly or if they're suboptimal, I assume that you it's almost like when you code something efficiently or you don't. If you are creating a model that isn't efficient or creating wrong data or just taking a long time. Is that something that can obviously drive up costs from a multitude of ways? I don't know if that's something to explain where, yeah, you've integrated this, but it's not working optimally.
Speaker 1:And you have no idea because this is how you think it's done, but your cost is 10 x of what it should be. And that's what that's one of the things that can happen.
Speaker 3:Yeah.
Speaker 2:And I think there's sort of specifics, and, Joe, you're probably gonna go into this as well. I think there are any specifics which we probably don't wanna dive into looking at this, but I think we've covered on some previous episodes just around even things like token cost. And to answer your question, Vikram, I think absolutely, if you don't set up things correctly, if you don't have the right people to implement this, you know, you could have a whole array of issues. Cost is one of them, but it's also incorrect answers or, again, there's biases that can come through. So I think there's a there's a lot to it there, but I think to sort of keep it in the scope of what we're talking about today, it's almost, well, what are the what are the main things that that you would want the AI to to do?
Speaker 2:And I think just jumping back to, yeah, back to the next. So we looked at chat completion. If we look at another, function or capability of AI, it's text generation. And if you've ever played with any of the, the natural language processes, you'll see one of their abilities is amazing generation. So, you know, being able to generate these sort of coherent pieces of of text or poetry or creative writing, anything like that can be amazing for content creation.
Speaker 2:So let's say whether you want creative or not, you can use the AI to automatically generate articles. So for instance, you've had some customer service conversations or you've had, you know, a request on a new feature product feature from your customer that could be almost built out automatically into a fully spec document. So what may have taken a product manager, you know, an hour to to write out can be generated in a few seconds. That's a massive some power feature. And, you know, this is generating blog posts or product descriptions or, you know, keywords for marketing, helping with SEO, all these kind of text generation opportunities are offered through the AI.
Speaker 2:And then we look at another, use, which is forecasting. So if we look at something called time series forecasting, it really can help sort of predict future values. It can be applied in, you know, industries like financial, weather, resource planning. A good example is with sales predictions as well. So sort of predicting future sales trends, looking at historical data, and, just helping businesses manage their inventory and even their marketing strategies.
Speaker 2:So, you know, some really, really, valuable features there as well. Going on, another one, and we'll look at another another three potential uses, is with translation. So you can look at, things like text to image or optical character recognition or OCR, as well as computer vision, which kind of covers things like object detection, image classification, facial recognition. All those are, you know, capable through these, models that assist with translation. An example is really imagine something like a multilingual website and, you know, just being able to create your your content in one language and have it dynamically generate all the content in a number of different regions that you, you know, you operate in.
Speaker 2:So, you know, instead of going to someone and requesting that you need a translator for this information or relying on an online translator that may not be as accurate, you can literally just do these almost translations on the fly. I was quite amazed just working with, Czech GBT the other day, and, I think my son was using it, and he mispronounced something. But it obviously sounded like, well, we'll go into, you know, Dutch. And, you know, there are ways in Dutch, and then, you know, you could switch over to Spanish. And it that just highlights how dynamic these models are.
Speaker 2:And, you know, just by generating if you said this is my base text for all my marketing material in English, no reason that you can't do a a build, and within a few seconds, you have all that content in any language.
Speaker 1:Does it have a English accent for you,
Speaker 3:or is it a German accent?
Speaker 2:It does. It's actually interesting. When it when it went to Dutch, it it put on a Dutch accent. But it start it started off American Dutch and then went into full Dutch, which was in terms of, like, the third question. It was was full of Dutch, which was Was that strawberry?
Speaker 1:Were you testing?
Speaker 2:No. No. No. That was actually, 4 o,
Speaker 1:g
Speaker 2:d four o just with the the mobile, speech kind of engine that they've got.
Speaker 1:I've been using strawberry, and I think it seems to think more. I think it's just a smoke and mirrors thing. They're just making it look like something's happening. Well,
Speaker 2:I thought my my mathematics was pretty advanced. So, I was joking with Joe the other day about it, and I, you know, found this what I thought was a complex mathematical riddle. And, I I fit it into, you know, the what is story? One o, and it came back with this beautiful with writing out equations. I was like, this is beautiful.
Speaker 2:It's reasoning everything. Now let me go back to 4 o, and it did exactly the same thing. I was like, okay. So, obviously, I mean, what they're saying, like, in some of the advanced mathematic exams, I think the difference was GPT 4 o is getting, say, like, 15%, and, one o strawberry is getting 80 something percent.
Speaker 1:Mhmm.
Speaker 2:So, clearly, that's that's a whole level of mathematics that Yeah.
Speaker 1:I'm not exposing myself to. So
Speaker 3:That's funny. Yeah.
Speaker 2:So I was impressed with both of them. But, anyway, then we look at, another option, which is kind of a good segue. It's really speech recognition or text to speech, And, that's really just allowing users exactly like I was talking about now to communicate in their in their own language, and you're having this sort of dialogue that would usually be in a text interface, actually in a, speech to text or, you know, text to speech as well. And that's really good for doing everything like just summarizing. You can imagine, you know, just chatting to an interface provides an easy level of interaction and less of a barrier, especially if you're not good with typing.
Speaker 2:So things that we would have been exposed to and potential use cases would be if you need a voice assistant. So a voice assistant helps you convey text based instructions or any content into speech for, you know, voice assistants, similar ones that, you know, we would know things like Siri or Alexa or, you know, the Google Assistant, but obviously just enhances the user experience with, you know, better kind of auditory feedback. And then the last one we'll look at is sentiment analysis, and, this can cover things like anomaly detection, data, or brand monitoring. And, really, if we look at something like brand monitoring, it's sort of analyzing the, you know, customer reviews, social media posts, and any sort of customer feedback to determine, well, what are the customers actually, you know, thinking now? What are they saying?
Speaker 2:What are they talking about? What is their sentiment towards our brand? So it can be a really clever way. You know? Previously, you'd have to have someone that's monitoring this, and then they're still processing it through their own internal filters and saying, oh, we feel we're doing well or we're not doing well.
Speaker 2:This is an area that we can improve on, and now we actually, you know, have much much clearer data from, you know, the AI model that would run through all that.
Speaker 3:And, again And, again, the the real power behind all this is it can be automated so easily. So instead of having some really sophisticated, as you said, software, which, you know, gets it right maybe 80% of the time or having a team of expert humans running through all this data, trying to categorize it, you can just offload this to the AI. So it's like offloading all that sort of functionality. And this is so quick to spin up and to test too. You don't need to invest large amount to sort of try and understand, well, do I have to invest 1,000 and 1,000 of dollars to even see if this works for me.
Speaker 3:It's very quick to sort of add in some data and and get those answers and sort of understand if it will work for you or not and then decide it's, like, you know, properly implemented into your infrastructure. So it's really great at that. Mhmm.
Speaker 2:Yep. Yeah. I think that automation is the key. I mean, that's where, yeah, cost savings come in. But, yes, that's that's an overview of the possible use cases, but I think, yeah, there's so many practical options, which maybe, you wanna wanna summarize some and and just look at what are what are good use cases.
Speaker 1:Yeah. Definitely. I think, you know, we could come up these are 5 good examples of, of use cases for AI. And what you're starting to talk to already is automate repetitive tasks. Obviously, anything that is time intensive, customer service automation, email filtering, expense tracking, those are things that AI can really help with.
Speaker 1:Processing data faster. Image rec recognition. We're not talking about the big brother face facial recognition maybe. But, you know, some of that for sure if you're especially, like, you know, manufacturing, vision, things like that that you could just pick up really quick and you you wanna create things that will be very fast. Document analysis, something that, you know, could be obviously legal type of documentation that this takes forever to analyze.
Speaker 1:I'm sure that that has improved since the beginning where there was a lot of controversy around, AI analyzing legal documentation and coming up with fake data behind it. So
Speaker 2:And to add to this, I gotta I was I was chatting to someone, and it seems that it's more and more popular now for for these sort of internal document analysis tools. You know, if you have a massive library of digital documentation, you know, you put together a an internal only kind of AI chatbot, and, you know, it just helps you surface that knowledge. You could say, well, I remember let's talk about legal. I remember we had a certain case where, you know, this was the outcome, and you can quickly find that case and summarize what the differences were with maybe another case. Now that is just such a game changer because it would have taken you hours and hours to find that document, review it all, especially lengthy documents opposed to getting a quick summary and actually seeing any sort of concerns now of of things like, oh, well, could the AI be hallucinating?
Speaker 2:You know exactly, oh, this was pulled from this document. You can see this reference right here. You you got all the facts, but it's just pulled it up. Made your job so much quicker. Mhmm.
Speaker 1:For sure. And helping parents help their kids with homework. I'm not gonna say that I don't use that because I don't remember my fractions and how to add and subtract and divide those.
Speaker 2:I was about to say, who hasn't used that tool? I know I I certainly have, have tried that. Alright. This is the, this is the homework assignment. Yep.
Speaker 2:I'm gonna hand you over to my son just walking through it very basic while I can see something I'd rather be
Speaker 1:doing. But I
Speaker 2:Don't know
Speaker 1:where it would be. I have used it for my 8th grader for the last 3 weeks for our reading assignments, English assignments, everything, but I said you can't copy word for word. Let's you figure it out first, and then let's use it as a support.
Speaker 2:Yeah. So we'll
Speaker 1:see where that goes. But social media marketing is obviously that last one with the processing data faster. I'm sure depending on who you are, whether you're Facebook, Twitter, Instagram, OnlyFans, whatever. You know? You're trying to use, you know, AI to regulate that content.
Speaker 1:So that's that's interesting of how you program that. Improve uptime and reliability is another one. That could be from smart home monitoring, your vehicle diagnostics, health monitoring. So I think that's plays it's gonna play a really big part, especially with when you get to health is is something that is important for me and my family and understanding, like, when can we detect certain, you know, conditions or anything like that is, I think, is very key. And if if something is reliable more than me visiting a doctor, I'd rather use that as a first line of defense.
Speaker 3:I've always wondered just on that point. You know, when everyone's wearing Apple Watches and they're gathering all that data, if they could somehow match that up with people's, like, hospital records and then sort of use AI to sort of understand those links between all the measuring that's doing on you and, you know, maybe like diabetes or something like that, and predictably say, you know, it looks like you're at risk for this based on 10,000,000 other people who have got the same sort of monitoring data as you. It's gonna be hugely powerful. I can see them going down the road. So
Speaker 1:I am curious though if you if you haven't haven't heard in the news 23 and me, the entire board walked out and quit the same day.
Speaker 3:Oh,
Speaker 1:really? So, yeah. Oh, so There's been controversy around them selling people's data to the to the government. Yep. Obviously, there's there's uses for good and there could be uses for, you know, evil with that.
Speaker 1:So the entire board leaving is, something's happening. They probably don't wanna get in trouble. Yeah. Yeah. Interesting.
Speaker 1:Oh, well.
Speaker 3:Yeah.
Speaker 1:Yeah. So the last 2, for good use cases are optimized resources, and that could be from energy management, inventory management, traffic management, things that, you know, are, again, that take a lot of data that you want AI to to scan and quickly figure out a more optimal way versus a human doing it, that's a great use case. And then the final one is to provide predictive smart insights. And so that could be financial planning, weather forecasting, and sports performance analysis. And I think, yeah, weather forecasting, you know, I'm I'm curious to see how accurate that is.
Speaker 1:Just watch the new twister. So I'm I'm I'm thinking now, can you can you predict with better reliability about tornadoes when that happens versus what they were using before? And is that is that possible? That'd be very interesting to see. And financial planning, I use that quite a bit actually for just to end just curious of, you know, the Fed just cut rates in the United States by, you know, 50 basis points.
Speaker 1:Do you know? Is that going to crash the market? According to history, it should have. And stocks and crypto are all up today. So it's hard to say, like, is it gonna is it gonna be able to predict those kind of things and help you out?
Speaker 1:And so but those are the five examples of good use cases. And, I think that's a good segue for the next one.
Speaker 3:Yep. I'll talk a bit about how you can actually then approach to determine in your own company, like, where I can actually help you. So the best way to kind of go about it, if you wanna do it yourself or or hire a team like us who has the, sort of lot of experience in understanding the capabilities. But you're really trying to take, what can AI do right now that's capable, that's, you know, actually feasible. There's a lot of, you know, experimental AI services out there that you can't actually very easily and reliably integrate into your actual organization, you know, in a scalable way.
Speaker 3:So it's actually important to understand, okay, cool. There's all these models out there and all these services. But what can we actually use in a really simple way that doesn't, you know, require a large amount of tooling to actually get it to work that we can just plug in and start using straight away. So the idea is to first go through an exercise of understanding what's out there. Let's understand the landscape of AI.
Speaker 3:Unfortunately, it's changing every day. When we talked earlier about all the different models that are out there, those are being updated. New ones are coming out every 2nd month. So at this point in time, you definitely have to keep your your your sort of, what's about you and understand what's better and what's changing because they're not small jumps. They're sometimes large jumps, and you don't want to invest into a whole AI strategy with certain models.
Speaker 3:That can be very outdated very quickly. You need to be able to build up your strategy and your services around that, that you can pivot to these new services as they come out without you requiring a whole load of rewrites or sort of architecture changes as you've integrated this AI. So it's really important to make sure that you're staying flexible. So it's an idea of, like, let's understand what's out there. Let's understand the landscape.
Speaker 3:And then as Mako just went through and Bunny went through, let's look at, you know, what the actual use cases are. And you sort of go through all your entire process flows, sort of ideas that you have, product offerings, trying to understand the market where you can actually see AI coming in or something that your voice wants to do, but it was always just seemed too complicated or too expensive to actually realize. Can AI actually help us with that? So you're taking 2 lists, understand the capabilities, understanding, you know, where we can actually see value in these and then matching those 2 together. And ideally, once you do that, you've made those matches and you can start looking into the cost effectiveness and, you know, understanding, okay, do we actually have the data and infrastructure to support the solution, which is another step.
Speaker 3:So, you know, we've talked about AI being really easy to integrate, but you do have to integrate it. Right? You still got your existing systems, your existing databases, your existing data. Do you have enough data to to support it? If you want to create a customer service bot that can answer loads of questions, do we have the data that you can actually use to answer those questions?
Speaker 3:If it doesn't, then it's not going to be very effective. So you need to understand, okay, do we have the data? Do we have the services that easily integrate into AI? Have we built some custom solution, which is gonna be very hard to match up to it all? Are we currently using loads of third party services, you know, that just, you know, have their own very narrow solutions that we can't easily integrate all these new AI functionalities into, which can be a big problem right now.
Speaker 3:Either large, software sort of sites and large software services, something like Salesforce, are slowly adding AI integration into software, but it's very scoped to what they wanted to do for their specific offerings. You can't easily, like, you know, add OpenAI into it and start using OpenAI, which has these amazing models. You have to sort your limits to what Salesforce can do. So it's important to understand, okay, what's our existing infrastructure? Can it actually be easily integrated into it?
Speaker 3:Once you've seen that and you go, yes, it can, you've looked through all of that and you go to a green text everywhere. Then it's looking at the cost effectiveness of this. Okay. Let's understand the compute costs around this because AI does cost money to run AI, to run the services around it. There's a cost to that.
Speaker 3:There's compute costs for some AI models or token costs, some AI services.
Speaker 2:You need
Speaker 3:to understand that and sort of do some sort of feasibility study to understand if it's actually worth it. Is it actually, you know, going to make sense, financially for you? And, of course, once you've done that, you actually need to implement and start testing it. One of the main problems with AI I've seen so far is you'll see a lot of really flashy demos that are doing something very narrow, and it looks amazing. And you can be like, wow.
Speaker 3:That looks like magic. And if I can get that into my company, it can book flags for all my users and, you know, for my, staff and will make automatically do all this for us. But, really, it can't. If that demo showed a very specific use case when you try and scale it out to do more for you, it actually can't. So you gotta make sure that you can test it, understand what it's capable of, and then test it as you go.
Speaker 3:Right? Pulled up the test plans as new models come out. You're not sort of adding these new models into the mix, and they're just not working in in reliable ways. You need to be able to retest them and make sure that when we add in new models and new services into it, that we're still getting the same accurate answers that we were before as we kind of pivot around between all the different technologies. So it's kind of a long winded explanation.
Speaker 3:But, really, again, you're just trying to look at all the capabilities that are out there right now. Look what you currently have. Look at your services, as Billy mentioned and Makoto mentioned earlier, like, look at your pairs of tasks. Do you wanna process data faster, improve uptime reliability, optimize resources, do some sort of prediction, all you want to accomplish and try and match the 2 together, do a sort of do a study around that, which can be extremely helpful. When you're thinking about what AI services can use, we've talked about models.
Speaker 3:We've talked about all these different services around that. I'll hand over to Brendan. Yeah. Brendan, you can sort of discuss that a bit if you want to. You know, where do you go from there that when you're trying to understand this landscape?
Speaker 3:What should you look at?
Speaker 2:So you have that approach, which, Joe, you're mentioning, and, yeah, of course, you know, now it's it's about, you know, lining up services or, you know, the the actual implementation. So there there are a few options, and you can really start again when we started this podcast looking at the models. And it's really a decision on, you know, do you use prebuilt? Do you, you know, do you have a specific niche requirement with small models and you train your own? You gotta factor in what that's going to cost to host them, to train them, you know, the the larger as Joe mentioned, the, you know, the larger the scope, more money, the more resources it's going to take to train that.
Speaker 2:It's also time consuming, and you're going to need that sort of expert knowledge. So sort of pick and choose, you know, do you want to go down that route of hosting your own model, even training your own model, or do you opt for third party services? There are a lot out there, I know we've evaluated a lot, and they're often expensive, and the problem is you can also be tying yourself into the exactly, Joe, what you were saying, the functional direction that they're going to best support their product, not necessarily your product. So it's general. You're going to have limited functionality.
Speaker 2:You can't pivot necessarily to where you want to go. Building on what you're saying about Salesforce, you know, they'll they'll do that to, you know, often improve their service offerings, but, you know, when you want to do something custom or move outside, you've got to ask the question, do you want to go down a proprietary route where you're tied into a certain application, or do you need the flexibility to move where your users are going and pivot which a kind of customer approach will offer you? And then the last option is really using cloud services. So there are a lot of great if you look at AWS or, you know, Amazon's web services, you look at Azure, there are a lot of platforms that offer these cloud services with different models that you can easily tap into and you can integrate into your own software. So they've got the ability to immediately take advantage of a secure environment where, you know, you can change the technology.
Speaker 2:You have more opinion on, you know, what you could do and what you could use and the direction you go, but, you know, you are able to just harness these really quickly and select your models and choose which range, have that data security, and, you know, keep it within the sort of environment that your organization if you're a Microsoft shop, keep it with Microsoft, and, you know, use that. So, you know, those are some options on how can you practically, you know, go ahead and and improve this. And Yeah.
Speaker 3:And we just to sort of also discuss the last point a bit when we talk about cloud services, they really it's almost like model as a service. It's kind of like you're using an existing model out there that existing cloud services are hosting for you and making available for you. So you can pick and choose as you want, since it's really easy just to sort of connect up to them. So just sort of when you're looking at the landscape, it's like, okay. Do we build our own models?
Speaker 3:Do we use existing models but try and host them ourselves and try and use them ourselves, which is, again, really costly when it comes to computes and infrastructure costs? Do you sort of build out this to sort of, you know, go with a a package third party solution that does it all for you? Again, as Billy mentioned, very limited there. Or overall, yeah, do you wanna go with cloud services, which I think are which, in my opinion, are just a great middle ground because it gives you the flexibility to sort of pivot to any models that come out when new ones also change your direction. But it's still giving you a very reliable infrastructure to build upon without sort of being limited by the other 2.
Speaker 2:And I think we've we've kind of covered a a whole range of considerations and applications, and you may be wondering, well, I I I see some potential, but how how do we go about doing this? And I guess that's where it comes in. How can we help?
Speaker 1:Yep. I think that's that's the the selling the AI pen. Sell me this pen, but no. In in reality, if you're if you're more or less confused after listening to everything that was said, it probably stands to to think about really which direction do I go. Do I hire some AI professional, take the risk, try to integrate something based on what I think this one person feels or thinks is the best approach.
Speaker 1:You know, at Impact, we've been we've been launching products, software products, for almost every industry for the past 20 years, small, medium, large businesses, you name it. And our approach has always been around a holistic approach where we think about the user. We're user centric first, not feature centric, which is this is another thing that could be another feature. And antiquated companies will think about, oh, we just want this feature, that feature, and then I don't care what the user wants. They're just gonna get it, and we're gonna spend a ton of money to integrate it.
Speaker 1:That's the wrong approach. We approach it very user centric, but we also look into business goals, objectives, constraints of the system. We tie this all together, and we we have our what we I don't wanna say proprietary, but it's something that we do very uniquely that has really helped organizations understand and lay out a path of what is a priority, what do users really want, what how does it affect our business, and create a a rapid way in which we do this quickly with leadership to come up with a road map of how that is going to happen and create your MVP. And this is something where because we've we've expanded our expertise to not just do software products, but actually integrate AI into products, This is something that we feel is is very important for organizations getting into this type of feature into their products because it can be a very expensive resource, a very costly resource if you do it incorrectly. And it can make or break your company.
Speaker 2:If I could add in there, I really like the point that you mentioned about, you know, it shouldn't be feature kind of driven. And I think so many people are scrambling, on that to get on the AI bandwagon. And we're gonna have this, and and that is the wrong approach. And I think that's why Impact's approach is so valuable because it is looking at what your users need, where can the user experience be improved, and from a business perspective, where can cost savings be realized, where can new revenue streams be identified, Coming out with going through this process and coming out with an idea of, right, these are our opportunities. How do we tie features to those?
Speaker 2:Instead of, well, we've got to get a feature out, but, you know, it has not been validated from, you know, any sort of level of user need or, you know, revenue opportunity or cost saving.
Speaker 1:Yeah. Exactly. So I think I think that's a good, good time to wrap up. And, yeah, appreciate you listening in today on our episode. And, again, feel free to reach out.
Speaker 1:Let us know if you have any questions. Like, subscribe, and, thank you again for your time. Take care, everybody.
Speaker 2:Forward to the next one.
Speaker 1:For sure.
Speaker 3:Thanks. Go well, everyone.