 Thanks, everyone, for being here. I'm aware this is one of the last shifts, so I appreciate everyone coming along. As Cronland said, in a very prestigious introduction, my name is Shane Lynn. I'm CEO and co-founder of a company here in Dublin called Edgeter, and at Edgeter, we developed software for customer contact centres, so centres full of people who are answering the queries that you and I make to large brands. To talk today, I'm going to briefly just look at the technology that we've developed, how we use it, and some of the obstacles that we've come across as we've built that out. That has challenged our assumptions. For those of you who run customer contact centres, this may be some repeat material. So to start off and set the scene, customer service is a gigantic industry. There's over 6,000 customer contact centres in the UK, and 4% of the UK population actually work as customer service operators. Demand is increasing. It's a key differentiator for every company. Good customer service brings in new customers, acts as a customer attention tool. Bad customer service drives people away and can really have large reputational damage. Along with that, the demand for this is increasing. People want faster answers, they want more accurate answers, and they want it right now across whatever channel they're using. The total cost of customer service isn't just the agent cost idler. It's bundled up in a wide array of the aspects of running a customer contact centre from equipment, supervision, there's a massive amount of attrition in the industry. Training can take six to eight weeks for someone who might only stay six months to a year. There's a huge range of costs that are built into running a full operation. And that's where efficiency is key. Ever since there's been customer contact centres, people have wanted to do better customer service with less funds, less people, just more efficient. And that's where AI, I suppose, has exploded onto this industry. So at every conference that talks on AI, at every industry event in the customer contact centre space, it's AI, it's data science. And there's a huge amount of hype, I would say, in this space. And it's a difficult one to define. So for us, AI has become, in these conferences, it's almost everyone talks the same, the sentences are the same from every provider. And AI has become an umbrella term in my mind for almost any decision that's made by a computer. And that's blurred the lines. So any one of the systems that you see here is all put under AI and that's shipped out to customers. What we've seen is an evolution of these terms. I think if I was speaking at a conference 15 years ago, it would have been a statistics conference, that would have evolved into a business intelligence conference, which would have evolved into a data science conference, machine learning, and ultimately now there's a lot of AI conferences. We might need to change the title for predict, actually. But really there's been three main advances that we've seen in our space. And the core ones here are, first of all, the ability to process natural language has improved. The key differentiator for us that we've seen is the advent of what are called word vectors. So word vectors are a way of representing text information in numerical format that computers can understand. And the context and almost the meaning of these words is encapsulated in those numbers. Secondly, there's been advances in neural networks and how they are trained and the topologies that we use. And finally, the access and availability of cheap and very powerful compute power via cloud providers. But really when I boil on AI, it becomes these three things. When I boil these down, the real impact here from a customer service perspective is the ability to classify text very accurately. And that's the core advancement that's enabled this huge boom. So text classification is the sorting of text into discrete buckets. It's very simple. I give a machine a piece of freely written text. That machine puts that into one of the buckets that I've predefined. And the accuracy of those classification techniques has increased dramatically over the last 10 years, going from 80%, 85%, up to 95%, 98%. So it's still not perfect, but very, very good. And this really, that's kind of allowed the explosion in along with text-to-speech, I suppose the explosion in voice control. So really the stuff that we see around Siri, around Google Home, around all that kind of voice interfaces, really it's text classification that's enabled this. So if I look at like a chatbot, which is the big boom in customer service, on the left-hand side, text comes in. It doesn't really matter what channel it comes in. Customers are talking on all the channels. Really what happens inside this is a system takes the words, converts them into numbers that a computer can understand, passes those through a machine learning model, any machine learning model, really, but the most accurate ones we've seen are neural network-based ones. And at the end, takes that bit of text and puts it into a bucket. So in this case, it would say, this is a cancellation. Loads of queries come in, they're all sorted into buckets. If I had a bot, call it a bot to make it kind of friendly, but really it's just a program that separates the text into buckets. Let's say this bot can handle two intents, cancel and add pickup. Obviously it's gonna just classify these sentences based on a machine learning model and previously seen examples into one of the two buckets. And once it's done that, the kind of AI bit is over and you're into a programming challenge. So there's two bot paradigms I would say from this point. One is where the answer is always the same. So if someone asks, I wanna cancel my booking, the answer is sorry, we don't cancel bookings. And I can churn that out to everyone. And the other one is, in that case, I would say that could have been answered somewhere else. And really it's almost like a complex FAQ bot and you see a lot of these FAQ bot type things which are useful and do deflect a lot of contact. The second type of bot then is more complex where it actually goes and tries to fulfill the customer request. So it says, this is a cancellation and now I need to go into my API, I need to look up my database, get that person's booking, identify the customer, make sure they can cancel, actually cancel the booking and then reply. And in that case, it has to do everything that you would expect if someone goes wrong, if you can't cancel the booking, all the forks that exist within that. But it works. So for instance, there could be two responses here. So for cancellations, the response could be, that's a pity, I'm gonna cancel your booking. Here's 20% off the next one, try and get the customer back. And for adding a pickup, no problem, if you've got a request, there'll be someone with you at arrivals, okay? So I imagine the bot can do that and it's a program that can handle all of the exceptions in that case. That type of technology had led to the explosion in there will be no jobs announcements, where automation is coming, customer service agents are no longer needed. This new technology is gonna take over the world. And the reality is that every leap in technology has kind of promised automation that's gonna take over jobs and replace people. But these clippings are from the 40s, the 50s, the 60s, the 70s, the 80s, and consistently the message is the same. Every leap forward, there's a, oh my God, it's over. Pack your CV in your bag. And the reality is here that we tend to massively overstate the degree of machine substitution for human labor. We put human characteristics on what are inherently non-human systems. And the reality is that chatbots and all of these things, they're fairly stupid. Like they take words in, they convert them into numbers, they process those numbers, they very accurately put them into buckets. But similar to the way I could have a music box here that would play beautiful songs. The music box knows nothing of notes, harmonics, dynamics, emotion. It can play beautiful music, but we're a long way from replacing composers. Similarly, a chatbot, in this case, has no idea what a booking is, what a cancellation is, what an arrival is. So that contextual information is maybe more important than we think. Because in reality, when humans are talking, it's complex, it's nuanced. Language is massively ambiguous. Words of loads of context. The Chicago Bulls is very different to Beware of the Bulls. Things are lost in misspellings. And humans, they're incredibly robust to errors. So humans are very valuable here. With the basic bit of training that I gave you three slides back, nearly everyone in the audience here would realize that these two requests have the same intent from customers, but are nuanced enough that they weren't a different approach. So for instance, the top one, that's probably a salvageable booking request. They want to cancel, they heard something from a friend, giving them a 20% off their next voucher. Probably isn't the right move here. And secondly, down below, this person is adding a pickup to their booking, but they've added a bit of context. And without explicitly giving a bot instructions around grove detection and what that means, that grove is slang for food, that food takes, everyone has this context, that food takes about 45 minutes to an hour in an airport, that all of that context and information for every topic would need to be embedded in the bot for it to reply and take all of that kind of nuanced communication into place. And there's a reality here too that by the time someone has actually sent this message to a customer service agent, they've gone through a number of steps, typically. And there's actually a plethora of options available to customers now before they get, because we've tried this before, we're deflecting customers all the time. They've gone to the website, they've gone to the FAQ, they might have gone to a self-service portal, they may have used the app, they've gone to an IVR, it's hard to talk to a customer service agent because all of these technologies are there in place. And if there's a button, if the bot can cancel bookings as a program, that button should have been available at all of these steps for that customer before they get to an agent. The reality is that agents are working on more complex things the majority of the time than we thought. So when we started with this, we were like, we were on the bandwagon, we were, you know, let's automate. But when we went into customer service departments, the things that get through all of these nets are typically complex. Yes, there are things that are simple that can be answered by bots, but when you actually look at what an agent is doing, and you sit down with a stopwatch, the process of answering a query involves quite a lot of activities. So they receive a query, they read it, they understand it, they look up that customer in a CRM, they understand what the problem is, they take any context into account, maybe something's changed, they copy a template in, they edit that template, they copy and paste data from the CRM, they send the response, and then they save it maybe in a CRM afterwards and record the record of the communication. If we look at what they're actually doing here, the key bits are the bits where they're adding the most value. No one really values the ability to look up information, the ability to copy and paste data, the ability to save things in a CRM. The key bits are the most human parts. So reading and interpreting the complex requirements from the customer, as well as customization and putting the human touch on the response to that customer. Ultimately, what we've seen is, the best parts of customer service are the most human parts. So understanding complex requirements, negotiation, humor, the ability to customize a response to really empathize with a customer in their situation, that's what makes customer service great. Whereas the other side of things, the looking up information, the copying and pasting, that's not helping the customer have a better day. And that's the focus that we have at Edge Tier. So we look at systems that use AI to do the bits that computers are good at, but keep the human in the loop to do the bits that the human are good at. So our system essentially reads emails as they come in to an AI system or a text classification system, if you will, then reaches into the API for the clients that we work with, gets all of the information that they need to display to the customer and to answer the question following best practice, and then presents all of that to the agent. So the agent's job becomes one of read the inbound, read the suggested response, customize it, maybe change it a little bit and then send it. So you can see how their job has focused to just doing the bits that humans are really good at and adding the most value. Briefly, the system works way better than we expected. We've seen up to a 5x improvement in handling times with corresponding improvements in agent satisfaction and customer satisfaction. So all the important scores are up. This approach, we believe, it is the future of customer service. It is the future of a lot of industries. Capturing the nuance of human to human communication with the complexities or the speed and accuracy of AI. This will be a given. This is the way software is going. We won't talk all of AI, it'll just be expected. Not having AI helping your agents in a contact center in 10 years' time will be like having a shop without a barcode system, an accountancy firm without Excel. It just, that'll be the way you do it. That's the way we see the future at Ed Sheer and that's the future that we're trying to build. Thanks very much for your time today. Very nice, Shane. Excellent, quick question. So, yeah, it does come back to then giving the agent power but also the final say maybe on how they communicate but helping them expedite that response, I suppose. Yeah, and I would say that one of the key things that comes up for us when we're talking to new potential customers is, when does the system take over and when does it learn for itself? And we're just, the technology, we're a long way from, you know, that example of understanding what going for growth means. We're a long way from covering that completely off. So I think keeping it in the loop is definitely the best part. We had speakers this morning trying to crack the artificial general intelligence problem and that's what you're describing there. That'll take a while, yeah. It sounds like we're a bit away from that. Quick follow-up. So each customer you visit or contact center you visit must be quite different and you have to deploy this capability into their process. Is there a bit of customization and understanding their business before you can actually implement these things? Yes, yeah, there has to be. So we've focused a lot on building a very generic system that's 95% deploy it and 5% build the system and the little piece for each client. And that's a little bit of machine learning models. We sit with the agents, we work out what best practice is and we have to integrate with their APIs. But it's a four or five week job for a big improvement. Very nice. Thanks very much, Shane. Pleasure, again. Okay.