The BCN Podcast

Not all AI is Equal Part 1

• BCN

Share episode

Artificial intelligence is changing the way businesses operate, but with so many different types of AI, it's crucial to understand that not all AI is created equal. Knowing which type of AI to use and when can be challenging. In this episode, Fraser Dear, Head of AI and Data Innovation at BCN, and AI expert Andy James, dig into the nuances of AI, breaking down its various categories. They explore practical business applications and provide actionable insights to help you start your AI journey.
 
 Tune in as we define AI and its diverse applications, categorise AI into generative AI, data science, and process automation, and discuss how machine learning enhances user personalisation and decision-making. In this episode our experts also explore real-world examples of AI in business, emphasise the importance of selecting the right AI model for specific tasks, and introduce AI Kickstarters to help businesses get started. This episode provides essential knowledge to effectively implement AI the right way. 

Peter Filitz:

Hello and welcome to the BCN podcast. This is our second installment, all Things AI, and today I have the pleasure again of Fraser and Andy joining me. Hello, gentlemen, welcome to today's conversation. Hey, hi, good to have you. So I guess, leading on from our episode earlier this month when we looked at essentially the past trends in 2024, and we talked about what we're likely to see this coming year, specifically really from Microsoft, we really want to elaborate a little more around what is AI and expand on not all AI is equal, as well as really providing businesses with a better idea of where they can start. So, with that in mind, I guess one for you, fraser, do you want to talk us through what do we mean when we say AI and the different types of AI?

Fraser Dear:

Yeah, sure. So if you've come from a previous one, some of this might sound a little familiar, but the reality is AI is everywhere. Ai labels are appearing on the front of most things to do with anything software, anything mobile, anything kind of biz apps or anything kind of consumer. It's all got AI on top of it. So what do we really mean? Well, here at BCN we define AI as the capability of a machine to imitate intelligent human behaviour. This involves the use of algorithms and models to perform tasks that typically require human sort of intelligence, such as visual or speech recognition, decision making, language translation, that kind of thing. Ai is kind of integrated across all sorts of different platforms and services. So from a Microsoft perspective, we might be talking about something like Power Platform or Cognizant Services or Azure to provide that kind of seamless, secure communication and data sharing between different kind of software components. But, as we mentioned at the headline, not all AI is equal. We kind of believe AI generally falls into one of three main categories. Category one is called generative AI. Category two is data science. Category three is process automation. So category one, generative AI.

Fraser Dear:

This kind of involves creating new content such as text or images or music or models, where you're actually asking for a model somewhere to do something on your behalf. And when we talk about a model, an example might be GPT. Most people have heard of GPT, but there are so many out there, so we've got models like Llama, Bloom, Falcon, Mistral, Claude Bert, Gemini, Dali. They're all over the place and they all provide very different, specific things, so they'll be tuned, they'll have different sets of parameters or they'll be designed for different use cases. Lama, for example, which is the model by Meta, has 405 billion parameters and is very much targeted for that kind of high-end performance research. However, GPT-NEUX 20b has a mere 20 billion parameters, so it's much, much, much smaller, but it's specifically targeted at general purpose text generation and that's what we mean by all of these different types of models.

Fraser Dear:

They're targeted for a very specific use case, and that's what we mean by all of these different types of models. They're targeted for a very specific use case. So that's generative AI Data science workloads. This category contains kind of the workhorses of that AI suite. So here we're talking about things like machine learning. It's the core of AI, where algorithms learn from data to make predictions and decisions about what's going to be coming up next.

Fraser Dear:

We've all watched those streaming TV services. Insert your choice here. But you've just finished watching one program and it says, oh, you might like this one because it knows what everybody else who watched that program then goes on to watch. So it's giving you an insight as to what might come next for you based on your preferences. The other sorts of things that you might see inside this kind of data science workload would be things like natural language processing, so it's things like that translation, that sentiment analysis, computer vision applications which take images of visual facial recognition, looks at maybe, autonomous vehicles. These are all data science workloads that have been put together to provide a kind of AI experience. And then the last category is kind of around process automation. So here we start to talk about things like robotic process, automation, document processing, have different functions or different elements, which we can then stitch together to try and create that experience that the consumer or the customer, or the business or the trust, or whoever it may be, is looking to achieve.

Peter Filitz:

That makes perfect sense. Thanks so much for articulating it in that way. So it comes down again to making sure that you select the appropriate technology and model for the appropriate business use case. Andy, I know you've obviously done a lot of hands-on work in this space. Might be good to sort of reference some examples that we've applied some of these different models to 100%.

Andy James:

Well, whenever we do that because, like we we say, not all AI is equal and there's so many different things out there, but AI and GPT has become such a buzzword it's been really important to have a kind of go back to basics with people. So all of those projects will start with a conversation or a kind of a workshop. In fact, I'd run one yesterday and we were discussing this exact thing and I likened it to saying I need transport. So saying I want to put AI on it, or I need AI or what can AI do, is a great conversation to have, but it's like saying I need transport. It's like cool, what type of transport do you need? Do you need a pushbike to do a paper round? Are you moving house and you need a lorry? You're going to Australia and you need a plane? They're all different types of transport. They all move you from one place to another, but they are very specific in what they do. So it's really important to pick the right one. You know you could do a paper round with a plane, but it's probably not the most cost efficient way of doing it. So when we're looking at use cases, we then want to look at all the options that are available so we can pick which one's the right one to do. So that could be really simple things like a power app where you're at a networking convention and you want to take a picture of someone's business card to be able to capture their details. That is AI. And there is a Microsoft business card reader model that goes straight into a power app that you can do that and it extracts the text and you can save it. And then there's I mean there's loads you could do sending emails, adding numbers, contacts and systems and whatever it might be. But that ai model is there, ready to go? Brilliant, um, it might be that you want to use uh ai to you know a similar process to extract information from invoices Same kind of thing. There's a model ready to go to do that and help things along.

Andy James:

It might be that you need an agent so that, actually, when an email comes in to a complaints mailbox, instead of just sending a person back a generic email that says, oh, ever since you've had a problem, you'll hear back from us in the agreed SLAs. Actually, let's put that email through some sentiment analysis to see what is it you're upset about. Where is it that the problem is. Let's identify those, but then let's give that output to AI to say these are the problems this person's had, who needs to know about it? Let's direct this. Let's do the triage who you know? Let's direct it to the right team. And then what?

Andy James:

Actually a better response to the person instead of just some random templated response would be well, let's look at our core systems about this person. What is it they're doing? What business are we engaging them with? What else have they got going on? If they had a problem with project one, let's let the people of project two and three know so that we can say this issue over here happens. Let's not repeat it. We can give some assurances back to the client about it. Talk to them about in our response.

Andy James:

Thank you so much for your continued loyalty with us over this amount of time. We've done all this, this really great stuff. We're looking forward to this new stuff that's going on, but we'll absolutely deal with this problem. All of that can now be automated and use of AI. But we would have sentiment analysis in one place. We'd have text generation in another, we'd have agents picking things up. So it's the right tool for the right part of that function to make sure that the process is as efficient as possible but also as cost effective as possible and manageable. We can maintain it without kind of boiling the ocean or going overboard. So that's kind of the first bit that we look at on these AI engagements.

Peter Filitz:

That's really great overview and I really like the analogy with the transport. I think that really explains what we're talking about. Yeah, I guess, fraser, one for you Talk us through the sort of Azure piece and the large language model creation, because you know, I'm speaking to customers on a regular basis and they all think that LLMs are the savior and that's exactly what they need. But I think we've again just made the point that, yes, it may well be, but there might be other options as well to do what you need it to do. So let's talk about that in a little more detail, if we can.

Fraser Dear:

Yeah, sure, I mean, I think when we talk about Azure, it really is the kind of the binding glue behind a lot of the workloads Zand is just described there. So kind of the binding glue behind a lot of the workloads Andy's just described there. So kind of connecting to services and connecting to some of these models. There are standardized pieces inside Power Platform. So Andy mentioned the business card reader, for example. It's a predefined, pre-trained model that's ready to go. Awesome, great, great, great, great, great.

Fraser Dear:

But actually in you know fictitious company A, actually they don't want to read a business card, they want to read something else and actually it's quite bespoke and it contains maybe some codes, maybe some images and also maybe a barcode and a data matrix and something that's a bit more sophisticated than just a business card. Actually we could do one of two things we could train something specifically inside the power platform to tackle that and build a custom model, or we could then start to tackle the entire problem in a different way, using azure elements to build up a structured engagement. So let's use an example. We have a business who has a, an audio recording service on their inbound calls. Make up a use case here. Let's say it's a banking scenario. So some people are going to be talking probably about their accounts, their finances, perhaps maybe some insurance information, and so they're calling up to make a claim about something. There's going to be lots and lots of potential for personal, identifiable information to be proliferated within an organization.

Fraser Dear:

That's probably not a good idea and actually the organization who's making those recordings should be tagging those calls to say this contains personal identifiable information of these sorts, because actually actually something that contains you know, bob lives in Street A and their postcode is X and their phone number is Y is kind of one type of personal identifiable information. It's important but it's perhaps not quite as important as perhaps that person's account number and sort code, maybe their national insurance number. They would fall into maybe a slightly different category of personal identifiable information and actually those recordings will need to be stored somewhere else. They need to be maybe moved to a secure location or maybe be redacted. And Azure can do all of those things so it can process the audio one of the workloads. It can then redact the audio and redact the transcription and move the file to perhaps maybe blob storage. That's secured in its nature so we don't lose any information. But actually we're very clear about why it's been moved because it contains a very specific element and, similar to what Andy was saying earlier on, those things can be pre-trained by another model or you can create something custom. So perhaps this, this organization that I made up, has a very specific, unique code that's important to them. You can build that into the model so again it can automatically look for that kind of text string or that kind of sequence of letters and numbers and classify it, as that is the special number for my business and I want to know that that recording contains that information. So that's one use case.

Fraser Dear:

If we start to think about kind of like that personalization and recommendation piece again same idea we're going to be using kind of Azure workloads to identify patterns in someone's buying cycle or perhaps someone's sorts of websites someone might look at. You could then track all of those and say, well, actually Fraser went to website A, website B, website C, and he's never been to any of those websites. We can't use that data to inform me, but Peter's done exactly the same sequence and the one that he went to afterwards was website D. Awesome, maybe we should suggest that website to Fraser because it's likely he's going to want to look at that website after this, based on that big set of data that we've got in the background in our business and that's where we start to then create these large language models.

Fraser Dear:

Take machine learning look at lots of data, predict an outcome with that data, apply test data to make sure we get the right answers and that's all of a sudden becoming something quite powerful and quite valuable to businesses to then pass to their consumers. I've just touched a little bit there on that kind of machine learning element. You can then start to get into kind of much deeper elements, so like deep learning workloads, and in here we're trying to mimic the human brain structure and actually start to think about, maybe like image recognition at a much deeper level, you know, identification of subtleties like patterns within an image, and that then starts.

Peter Filitz:

Fraser, thanks for those insightful thoughts and overview in terms of the Azure models. I guess, Andy, one for you. Businesses are obviously looking to embark on their AI journey. We've launched a number of AI Kickstarters to help businesses get going on that journey, so to speak. Do you want to talk a little bit about them and how businesses can get started?

Andy James:

Absolutely, and the reason we've got so many is exactly for this reason. Not all AI is equal, so we need to have different options of Kickstarter for different process, for different needs. So we have kind of a quite straightforward opportunity with an AI Kickstarter where we can come in and within a couple of days we'll have a co-pilot studio chatbot. And again, this is where terminology is important. I don't want to say that we're going to implement a co-pilot. That's a separate service that we can 100% offer, but we'll get a co-pilot studio chatbot up and running on a set of data that you might have that you can then use the bot to interrogate and ask questions of. So that could be anything from HR policies and processes around the organization. It might be that your organization uses a core economic impact report that's hundreds of pages long in a number of conversations and bits that you need to have and actually, how well are people going to read and digest that? Let's put it in the chat box so they can ask questions and interact with that data in a more natural way. You know the endless almost opportunities for where that could be implemented. The next thing we would look at is kind of having AI embedded into some kind of an automated process to try and, like we've said a number of times, remove that kind of administrative burden. So that could be like we were talking about, where we put an automation process in place to look at emails, get the sentiment analysis, do whatever we're doing there, sentiment analysis, do whatever we're doing there.

Andy James:

The third option that we've got for kind of an AI Kickstarter is where we're seeing repetitive kind of processes across all organisations, across all sectors and industries, and that's around extraction of information from purchase orders, sales orders, invoices, those types of documents. Currently, in a number of places it's a really manual process for that order to come in, someone to sit, read through and type it into your kind of core system, whatever that might be. So we have kind of the framework of a solution we can put in place that when that order arrives, an automation grabs the email, takes the order that's attached, runs it through AI to extract all of those line items and all the details from that order, and then we get a confidence score from that AI model to tell us how accurately it feels it's read that information. If it's confident enough that yep, I absolutely got this right, then we can push that automatically into your system. So we completely remove that administrative process.

Andy James:

If the AI is not sure, then there's a power app that you can just kind of manage by exception and you'll see all of the orders, the automations process, the ones that have gone all the way through and the ones that need some kind of intervention. You can have a look in the app, make any changes to the data just to the data that's wrong, click a button and then the automation will take it back on from there. So that's more automation than app. But it's because of that AI reading that information and extracting everything that we're able to kind of smooth that process over and make it a bit easier. So they're the kind of things that we've got as Kickstarters, but we can have a conversation over any process and see what we can fit in.

Peter Filitz:

I think it's fair to say that there is a plethora of options available to businesses today, and you know we as a business can certainly help organizations on their journey in identifying suitable business use cases, along with appropriate technology and services to meet their needs. Well, I guess that brings us to the end of today's conversation. I hope our audience has found it informative and useful. Please don't forget to visit our website, wwwbcncouk, for a wealth of information about AI and all the other services that we provide, and don't forget to like and subscribe to our podcast. We look forward to seeing you on the next one, and take care and stay safe.