In this episode of Mobile Interactions Now, ContactEngine CEO Dr Mark K Smith talks to tyntec's Jean Shin about how to bring about Twitter-esque experience of instant response to questions companies are asked by their customers.
Listen to the full podcast here
Jean: Mark, welcome to the show. I'm honored and thrilled to have you on the show. We gave an intro in the beginning, but I know people will love to hear more about you and your work from you.
Mark: My name's Dr. Mark Kingsley Smith, and I run a company called ContactEngine, who are based in London, Bletchley Park, and northern Virginia.
Jean: I've been postponing this big topic of AI, waiting for just the right guest to have on the show. Without further ado, now it's time. Let's jump right into it. Starting from the very definition, and I want Mark's definition, not some wiki's definition. What is AI?
Mark: For me, what we do is a subset of AI. Most of our work is around natural language understanding and machine learning. Machine learning has its origins in statistical learning. It's a maths technique. What it means for us is that we can become more human-like, in the way that we engage in conversations with our clients, customers.
Mark: For me, AI is about being empathetic in the way that you engage automatically with people, and knowing when computers are best and when humans are best, and trying to marry those two together in a way that makes everybody happy.
Jean: That knowing part peaks my curiosity here. How our machines are actually knowing, what does it take to let them know what they're doing?
Mark: Yeah, it's an interesting question. I'm not a great believer in singularity. I'm not a great believer in machines ever acquiring sentience. I think the Terminator was a film, and a work of fantasy. There are certain things that computers do that are remarkably better than humans. The thing that they do most brilliantly is dealing with mass datasets and making decisions based on more information than a human being could reasonably hold in their mind.
Mark: We have a vast, vast amount of data, a vast amount of information that we've collected from customers, which we can then label or classify if you like, because we begin conversations. That conversation might be around a repair journey. Going to somebody's home to fix something, or it might be to do with an unsuccessful bank loan, and intervening in that journey to ask a question around help.
Mark: What machines can do well is, if you have mass data sets, and you can label them, so you can classify the response types, then the more of that you do, the more that your algorithms that are underlying math can get better at matching what a human would do, and then, exceeding what a human would do.
Mark: Now, not because the machines are cleverer, but when the machine has an almost infinite memory, and the machine will work day and night and as instantaneously as you want. If you want in the modern era to have that Twitter-esque experience of instant response to questions, then machines are the only things that can do that 24/7.
Mark: For me, the magic of computing is the vast amount of information that machines can hold. Then, once you've trained your algorithms, the machines can begin to learn. They can see some patterns in data, and react to things in ways that are expected, but give a better customer experience.
Jean: Does it really take learning from machines end, or is it more of a case of giving them better how-to instructions? Are they actually learning in a human sense?
Mark: Oh wow. That's a very interesting question. My PhD is in biochemistry. I come from the natural sciences side of the technical track. I don't think we're ever so clear yet, how the brain works and what intelligence truly is.
Mark: There's something of a confusion, the notion of making something artificial if you don't fully understand what it is. But, the notion of learning, if you imagine an algorithm is a newborn baby, then it doesn't understand anything. When you feed it information like you would teach a baby to speak, then it can learn, and at some point, it can take the words that you're teaching it and form it into sentences, and start to advance in ways that you haven't taught it.
Mark: Now, that's not a bad analogy for a machine learning process, because it is possible once you've trained your machine to allow it to arrive at making a decision, with a degree of confidence, that is equivalent to a decision a human would make.
Mark: Let me give you an amusing example. I'm at present with my company, going through labeling a profanity dataset. We have a relatively modest number of quite rude replies. Now, because we're dealing with millions of pieces of communication a week, we have thousands of these. People are mostly very polite, but sometimes people just flip, and they say lots of rude things. We have hundreds of phrases and words that would correspond to profanity.
Mark: Now, sometimes, this can be very funny. Sometimes, this can be ironic or sarcastic. This can simply be to emphasize a point. Sometimes, this can be, that was brilliant, in which case that's not a negative. That's a positive. Now, if you've got enough of that data, and you analyze that data in context, you're looking at the question you asked before that resulted in the answer that was given, you can actually make some assumptions, and you can label that data, and then your algorithms can actually make a decent stab, a 95, 96, 97% confidence level, they can make a decent stab at responding to something that contains a profanity, but in a sentence structure you've never seen before.
Mark: Machines are able to do that, mostly with profanity, it's an escalation issue. It's actually passing onto human being. Because one thing that I'm fairly confident computers will never do terribly well, is being empathetic. Again, that tradeoff between the carbon, the silicon, that the human and the machine, is very important. There are moments in conversational AI as we call it, where the machines back off and hand it to a human being, who can then have an attempt at apologizing for a poor service, or something else.
Jean: It sounds like it's beyond choosing the right words.
Mark: Yeah, it's much more than that. The challenges of human understanding of languages are, to a certain extent, we've cracked it. You only have to experience the use of Google Home or Alexa or Siri or others to realize that the machines are making fantastic strides.
Mark: The subtlety of responses, take an example, I saw a comment recently, it wasn't a channel of ours, I think it was on the news, when somebody was thanking Virgin Trains, a train operator in the UK, for the wonderful sauna that they had had that morning. Now, that's heavy sarcasm. They're basically saying, the air conditioning unit had broken down. It was a very hot day.
Mark: Now, getting a machine to understand that is an enormously difficult challenge actually, because if you were daft, the machine would go back saying, "Oh, I'm really pleased you enjoyed the sauna," and that would have totally missed the point.
Mark: You have to build in to natural language understanding, an empathy towards what people are saying. I'm not saying that we've cracked that, but we're on the way towards cracking those.
Jean: You mentioned some of the data that you are learning from, tagging, and you mentioned some of the companies you're working with. Taking all that, if you were to paint a picture for us, because our audience is interested in how tech is changing the way we live every day, if you were to paint a picture of a customer journey that you want to see, more of an immediate future, what are we looking at? What type of experience are we looking at?
Mark: I think, this is sort of a shorthand that I occasionally use. A couple actually. One is the notion of one voice of communication with a service provider. Quite often, businesses are very siloed. The information that the billing department has, might be different from the information that service department has, which it might be different from the information a sales or a marketing function has.
Mark: As a customer, you don't care about that. You have a singular relationship with whoever provides your telephony, or your television, or your insurance products, or your bank. Therefore, the notion that conversation should be continuous and evolve with time, is where the technology can let you go. The data's quite fractured inside businesses. This isn't an easy challenge for them. In terms of the notion of the other phrase I often use and have coined, is human/computer rapport, the notion that if we have a conversation, make it very human.
Mark: You meet your neighbor for the first time, and you discover that they like football, or they like cooking, or they have children. The next time you meet them, you know that information, and your conversation is likely to be nuanced and informed by that knowledge. Now, in a much simpler way, machines can do that too.
Mark: If for instance, you'd carried out a piece of work for a customer, and they had given you a very poor score for that piece of work using voice of the customer, or NPS, or these other quasi-scientific means of collecting ones to 10s, but if they'd given you a very poor score, then the conversation you have with them next time needs to be informed by the fact that they didn't have a good experience with you. If however, they'd given you a very good score, then that conversation should be different.
Mark: This notion of machines acquiring prior knowledge of engagements, is human-like, and not only likely, but is already happening. Now, machines can't step over into creepiness level. I don't think a machine can ever be your friend. How does it feel? It doesn't feel anything. I often, I have to ask the question about sentience and singularity, and I simply reply that, after that computer won at chess, how did it feel? It felt nothing.
Mark: Computers can never feel, but they can be informed and the conversation that you're having can be longitudinal, and learn based on your prior engagements, and make for a better, smarter customer experience.
Jean: That sounds like a continuum of this whole notion of what people are talking about in terms of customer data, being able to compose a single view of each customer. Is this what is required?
Mark: It is, and obviously, many CRM systems will make great claims, and rightly so, to hold this single view of the customer relationship. Hence, the CR bit of CRM. I think where it starts to get difficult, certainly for me, is what we do is essentially proactive outbound conversation.
Mark: Now, most CRM vendors are not that interested in that world, because they're selling their software on a per-seat license basis, and so they like to have a human being sitting in front of their software, that acts as a sort of mediator between what the machine holds and what the customer requires.
Mark: Now in the modern era, with the kind of sophistication and natural language understanding and AI in general, and the ability to cycle across all channels to communicate with people, that's not the future. The future has to involve proactive conversation as well, to improve that customer experience further.
Mark: That's where CRM needs to change. It needs to do a little bit like web hosting has gone. CRM needs to go more to a transactional basis, rather than a per-seat license basis, in the same way that your AWS or your Azures, or your Google Clouds are now transactional, not having to buy vast numbers of servers and stick them in racks with uninterrupted power supplies, and all the rest of the stuff that I did 20 years ago.
Jean: I'm not going to ask you about that.
Mark: No, it was great fun. I did it. I'm so old.
Jean: Let me break that down a little bit. Basically we started out facilitating a better experience, and talking with a, say, customer service agent and a customer, now, you're beginning to talk about this more of a proactive, initiating a conversation or triggering some kind of action because you know these things about the customer and whatnot.
Jean: Now, is it good for the customer, in terms of their service provider knowing that much about them?
Mark: I think your question is, how Orwellian this can become. I think there are a number of examples out there where things get a bit ... We sometimes refer to it as passing a red face test. We don't want to get too irritating or too presumptuous.
Mark: I think however, customers like to be informed. They like to know what's going on. Customer service is probably the most important new frontier for communication, because if you look at so many services, let's take a bank loan or a mortgage or broadband or mobile telephony, all of the services in many countries, my own included, not the case all over the world, but my own country included, they're almost identical. Whether I get my TV service from company X or company Y or company Z, I can use the same kind of over the top software to watch whatever programs I want to on Netflix.
Mark: My broadband, I just expect to work. My mobile phone, I just expect to work. I've got a credit rating. The mortgage offer I get's going to be the same from one bank to the next. Therefore, what makes it different, is if customer service is perfect.
Mark: Now, you can never employ enough people for perfect customer service. It's not possible to have a one-to-one relationship between your 10 million customers and your 10 million service agents. If you have to make a call, and let's face it, most times when you're engaging with brands via the telephone or web chat or whatever else it is, it's because stuff's gone wrong. Where's my order, why is this broken, what's going on, why haven't you turned up? If you can proactively communicate to those people to tell them those things, we're going to be late because of the weather, or we can't turn up today because of this, or we are going to be there on Friday at 9:00, is that okay, would you like to reschedule? That is the kind of communication that everyone benefits from.
Mark: Just being told what's happening, even when it's a bad thing, being told is much better than just staying silent and then waiting for the calls to come in. Call centers cost a fortune. In the UK, there are a million people employed in call centers. 6% of the working population of Scotland are employed in call centers. In the US, it's just under four million people.
Mark: Now, the churn of these people is enormous. I know of one company who have 10,000 people in a particular call center, do a particular thing in the world of telephony, and they get 100% churn every eight months. So high, that they don't train them anymore. Just think on. They stop training them because they leave too quickly. Now, it's a job that humans don't want, it's a job that machines do better, and a job that humans want is the thing that humans do best.
Mark: Solving a problem that's very complicated, or dealing with an issue that's very tough. I'll give you a lovely example. I was talking to an insurance company recently around life insurance products, and they say we typically just get the one call. The reason for that is, let's say it's me. I've taken out life insurance, and let's say just after this recording, I drop down dead.
Mark: My wife would then look through all my paperwork. It will take her several weeks, because it's scattered all around the house, and eventually find the number to ring. She'll ring it up. She does not want to speak to a computer. She will probably, or she might be celebrating, but let's pretend she's upset. She wants to speak to a human being. She wants empathy. The conversations that first happen with this insurance company can take up to two hours.
Mark: Now, machine's absolutely not required there. Absolutely, be categorically the worst possible thing you could do. However, after that, machines are best. Let's make sure that we've got the death certificate. Let's make sure that we're telling you how long it's going to take before the payout's going to come. Let's communicate the correct bank account details to make the payment to. All those things, machines are best at doing that.
Mark: It's always horses for courses. Yes, I think there is an Orwellian challenge towards these things, but at the end of the day, the differentiation between most brands is going to have to be customer service, because pretty much, not always, but most services are very, very similar now. It's the human bit that makes a difference.
Jean: Given that kind of interaction, and how it happens along many, many stops in that customer journey, let's say it's inevitable, there sometimes, the handshake between machine and human agent, does have to happen. Where are we at in terms of making that work together? This whole machine to human agent, and what is being better performed these days currently by machines?
Mark: Yeah, it's a good question. Exactly when and where, is open to experimentation, is the truth of it. It's a little bit of a brave frontier actually. It's a bit of a new frontier. There are some moments where it's just obvious. In the example of the life insurance first call, a human every time. That's obvious.
Mark: Let's say, let's go back to profanity filtering. Let's go back and look at a customer journey that's broken in some sense. Now, that I'm breaking the customer journey might be because of terrible customer service, full stop, in which case the moment someone starts to get quite cross, having not been crossed before, that's probably their moment that human intervention needs to take place. There is cross and there is not cross.
Mark: There might be a combination of profanities about the cost of a service, but actually, the majority of the content of the messaging is actually very positive about the work that was done. At that point, it's then a tradeoff, a decision that needs to be made by the brand, to say actually, our lifetime value of this customer typically is three or four years. If they have a bad experience, that might be two or three years. In which case, at this moment, we think we should intervene and talk to that person, and spend the extra money to make sure they're happy.
Mark: That has to be done experimentally. I'm a scientist by background, so every time an intervention is made, you need to make one and zero choices about that, to actually prove whether that was a good thing to do, or a bad thing to do. Actually, driving brands towards thinking about simple AB testing, and even testing that we do lots and lots of AB testing around formality, or informality, about short-form messaging, you have to do that with control groups. You have to test and test again. You have to look for improvements and trends.
Mark: It's actually quite complex to do, but it's the only way of doing it. You have to be very numerate in your thinking about these engagements.
Jean: Any of the examples, you're thinking of insurance companies and whatnot, have you seen some exciting results?
Mark: I think a lot of the stuff that we do in the telephony industry feels to us to be quite normal, because we were doing it for many years, but we have a ... If you looked at ... There's a product suite we have called Appointment Control. This is somebody who's going to have an appointment for the installation of something, or the repair or something.
Mark: Most organizations are quite poor at communicating what's going to happen to those customers, or they might go down a mono-channel approach. They might say, oh, it's in our app, the one you have them both download, or we'll send you an email to the email address that we haven't properly got, or we'll make a phone call, but we have our call center employee between 9:00 and 5:00, and you're at work, so they won't reach you.
Mark: If you do what we do, which is cycling across all channels available to us, a text message, an instant message, a phone call, all automated of course, reach someone, then you are more likely to have a successful engagement, and you are also able to time that engagement, because you know it's going to happen between 2:00 and 3:00. You can survey afterwards, and say how was it for you in the moment of service delivery.
Mark: If you do all of those things, then you get a very good and very clear ROI, because most companies will know exactly how much it costs to send a person in a truck to do something. Those costs vary in the UK. Probably the lowest we see is about 40 pounds. Probably the equivalent of $40 today. In the US, that can be quite a lot higher, upwards of $140.
Mark: On that basis, if you can reduce failed appointments by 10 or 15 or 20%, one of the big telcos in the States, we've just come to the end of a proof of concept, which is now going national from next week, the answer is 12.8%. 12.8% times 10s of millions of people, times 100 odd dollars, is a massive ROI.
Mark: Also, there's some environmental benefits to that as well, because you're not wasting fuel going to unnecessary locations. There's very significant customer benefits as well, because the customer's happier, because it happened right first time. We have a client who repair washing machines, who used to do it right first time, one in 10. One in 10 times, they turn up to people's homes, looked in the washing machine, they knew there were 98 drums in their range, and the van only carried two, and so, do the maths. About one in 50 times, they could do it first time. The rest of the time, they went, no, I haven't got that. I'll have to come back.
Mark: A washing machine. You're going to be quite cross about that. Launderettes don't really exist anymore. What do you do? You just start to smell, or you run out of clothes, or both.
Mark: What we do is we proactively communicate with them, we get the serial number back of the device. The customer's going to want to give it to you. They're going to want to be accurate, because they want it fixed, and then they fix it nine times out of 10.
Mark: There's all sorts of really obvious benefits beyond the better customer service, beyond the definite higher NPS scores that you get from being proactive. The ROIs are enormous. One of our challenges as a business is, we are hugely skewed towards working with very large companies, because we're a transaction model. We basically get paid per conversation. Hence the reason we work in telcos, or Whirlpool, the largest white good manufacturer in the world, and so forth.
Mark: Where I'd like to see more of this kind of work is, is lower down in smaller companies, who maybe only have 200, 250,000 customer engagements. Those people are just as worthy of communicating to. We just need to get the math right on that, so that we can work there as well.
Jean: Let's say the ROI is improving, and industry as a whole, we start actually seeing some numbers to, at least to benchmark against. In the boardrooms of the corporate world, do you have any other questions in addition to ROI that's nagging them, or ... You mentioned earlier on, about red face test. Is there common questions that you have to answer in terms of some of the ethical questions that might arise?
Mark: Yes there are. There are, obviously for some years, the issue of data protection, is front of mind for all companies. Bear in mind, we're getting not massive amounts of data, but if you want to write a good message that gets a response, you'd need to know certain things. Somebody's name, where they live, if you're going to go there, and products and service and what have you.
Mark: GDPR is top of mind for us. I only employ 60 people, and it seems sometimes about half of those are involved in GDPR related matters, and rightly so. We have to take that extraordinarily seriously. I think the other thing that troubles me, and I think this will become increasingly difficult, and we have addressed it ourselves, but it's easy for us to do so, is AI as a service products, which are black box and don't explain what they're doing, I think that their days are numbered.
Mark: I don't see how that works in a GDPR world. A lot of probably 50, 60, 70% of the AI claims made in ... I work in what we laughing call Silicon Roundabouts in Shoreditch in London, and probably as much as three quarters of the companies that are talking about AI, are actually backing off that AI, to one of the big players, the usual suspects. Your IBM's, your Googles, your Microsofts, or what have you.
Mark: Now, the decisions that their AI is making, is not easily explainable. Sometimes, it's just too complicated to explain and make visual. I think there's a big backlash against that. I think companies will be thinking, why should I give all my data to you, so you can make my service better, and you make that improvement available to my competitor? That makes investment decision difficult for people to understand.
Mark: I see both sides of that argument, but I understand it. For us, we have one major advantage, which goes back to the simple coincidence of our starting the conversation, and it being shortfall. We're asking singular questions. It means that the decisions we're making are relatively simple. There's still some hideous maths behind it, but they're relatively simple, and therefore it's possible to ...
Mark: The whole issue of XAI, of explainable AI, is going to become more and more prevalent. I think you are on abound to hold your data for your client, and not share it with other clients, to train your algorithm of their data for their people, because it's theirs. Then also, explain what you did and how you did it.
Mark: That's the next big push. I'm certainly seeing some look of puzzlement on the faces of ... I was in a pitch with a American bank in New York not that long ago, where they were telling me how vastly complex their AI approval process was, and they'll have a board of experts. I was sitting there thinking, that's good, but I reckon I can outbid you with my AI board, which has some of the best prof in the UK on there. We're going to end up with some kind of intellectual standoff?
Mark: That's not wildly helpful. The notion of white box explainable is going to become more and more of an issue.
Jean: Just, not to belabor the point here, but knowing that you had a academia background as well, the collaboration you are talking about, sharing model that you're talking about, is that different from what you used to do in academia, or a more scientific field? Is there something that we can learn as an industry from there?
Mark: It's a challenge. It really is. It's getting worse actually. I know quite a few academic institutions, and I have quite a few pals who work in that world. They're being picked off one by one.
Mark: The large American software houses are looking at academic institutions, and they are hoovering up the talent there. I understand why, but, and there's a big but here, science is ... There's a cliche, a quote from Isaac Newton that, "I saw further because I stood on the shoulders of giants." What that's trying to say is, that academia is an advancement based on prior research, and any research needs to be published, and it needs to be peer reviewed, and it needs to be able to be tested by others to prove it works. Otherwise, you end up with cold fusion arguments.
Mark: 20 odd years ago, when people said they'd cracked the perpetual motion machine, and they hadn't, it couldn't be tested. It didn't work. It was nonsense. Now, if you've got large software houses hoovering up academics, the best minds, then putting them inside a client confidentiality wrapper, that will slow development down. That is a massive disadvantage. Now, what that means for a tiny company such as mine in comparison is equally problematic.
Mark: I have an AI board, which is, it's people like Nick Jennings, who's the AI lead at Imperial College, and the AI leader at Kings College, Michael Luck, and the guy that invented argumentation technology, the most complicated NLP challenge, is a guy called Chris Reed up at Dundee. Now, I am supporting them in certain of their research ambitions, and understanding that whilst my desire will be to retain the IP, that's not how it works.
Mark: There will be situations where we'll be supporting stuff. Our customers will be sharing data, where we will be stopping ourselves from being able to patent in Europe, because we will have to publish it in order to satisfy what academia is.
Mark: Now, that is a little bit different in the US, and we have patent pendings on various of our own technology. I'm an inventor on a couple of patients in the US. It's a little bit easier out there than it is in Europe, but it's a really, really difficult problem, and one that I think is going to get worse, because at the end of the day, most innovation comes from two places, academia and small teams. Innovation does not come from giant software houses. They buy it from small teams or academia, and if they hoover all of that up, then we will not advance as a species as quickly as we have in the past, and that's bad.
