Ford Motor Company (F) Earnings Call Transcript & Summary

Hello, everyone, and welcome back to Digital CX LIVE. I hope you enjoyed the last session, and we did have some polls in that webinar. We've also got some polls in this webinar, so we really want to interact with you during this session. We're joined by [24]7.ai. It's a sponsor session. And we've also got guest speaker here, Shyamala at Ford Motor Company. And they're delivering a presentation on AI/HI blending. Okay, over to you both.

Thank you. Thanks, everyone. Hi. I'm Celene Osiecka. I'm the Practice Lead and Senior Director of Conversation Design here at [24]7. I also have my great, a good friend and colleague, Shyamala here. So Shyamala, if you want to do a quick intro there.

Yes. Hi, everyone. I'm Shyamala Prayaga, and I'm the Product Owner at Ford Motor Company. And I lead the autonomous conversational AI efforts here.

Great. Yes. Thanks. So yes, so today we're going to talk about what we call AI and HI blending, which is artificial intelligence and human intelligence combined, and how these 2 technologies can work well together. So I just wanted to start here. So as a quick poll. I wanted to show our first poll here to get everybody just engaged, and we're going to start asking questions. So a poll's going to show on your screen on the right there. So who knows? So this is Deep Blue. But who did Deep Blue play? Does anybody know? [Voting]

We see a couple in there. Any others? Any other guesses? Go ahead and make a guess. While you're doing that, so Deep Blue was a computer that pitted against one of the most famous chess players in the world at the time. And it's just -- it shows us an example of where -- we look at machines and we look at humans, and they -- often we pit them against each other. They're at war, right? This machine's going to take my job. And the AI technology is too smart. It's too powerful, right? So we see that a lot. And I think that's one of these first interpretations of how machines and humans, our relationship has been with them, right? This is one way. Yes, we've got a couple of answers in there. So Garry Kasparov was the answer, for anybody that needs to know. But yes, it did win against him, and there was a whole bunch of controversy theories that it was a human playing him all the time, and there's no way a machine could have won. It was really interesting. But that's one way to see this. The other way to see this is this way, where you have Microsoft Tay, who's never heard of Microsoft Tay? And she shows us what humans to support your AI. So this is the other side where this is, oh, well, the AI should learn itself, and it should be all automated. And I don't need humans in there. I want it to do all for me. I don't want to touch it. I want it to be completely self-sufficient and unsupervised learning, which Tay was and what's happened was she quickly started to learn from her tweets and became derogatory and racist and just a whole bunch of things that were being taught to her and she didn't have that filter to know this is right and this is wrong. This is what you say. This is what you don't say. So this is the other side of no human is involved at all, what can the ramification be on that side, too. And Shyamala has a really good example too of kind of another piece around empathy.

Yes. So GPT is -- GPT-3 is a national language platform, which is OpenAI's platform, which is basically used for generating the language and helping the users. And there were a lot of examples where GPT failed. It's a blend. It's not completely what Tay is, where Tay is completely unmonitored kind of thing. But the -- OpenAI's GPT-3 is more of a blend between the human and also between the machine and it learns from different things. And recently, a study was conducted with a health care company, Paris-based health care company, where they were trying to see if these bots can be empathetic and can be used for health care needs. And they did something where a person started asking depression-related questions and like health-related questions, and the person said like, "Well, I do not feel great today. What should I do?" So the bot goes like, "I'm really sorry. I can help you with it." And then the user says, "Should I kill myself?" And the bot is like, "I think you should." So this is exactly where the machines are failing to understand what the user really meant, and they don't know where the balance is when it comes to really understanding like, at that moment, if the user needs something, I should be empathetic to help them out and not provide a response like this, where it can significantly impact the mental health for so many people, if that were the case.

Yes, exactly.

Celene, over to you.

Yes. Well, lots of examples of how, in the end, there's lots of strengths that bots have and strengths that humans have. So that kind of gets me into my next slide here. So this was the survey that was done, just if job bots were available and working effectively for the online services that you use, which of these benefits would you expect to enjoy? So the users all rated this 24-hour service, getting an instant response, answers to simple questions, easy communication. Those are really what we feel they're top-ranking ones, and that's what the AI's strengths are, right? Easy, simple, instant, around the clock, that's what we think of when we think of AI. There are some things right now, based on the technology, that just not strong at right now. We're getting there, but it's really the human strengths live into the complaints and customer experience and detail, complexity, friendliness, the human elements, right? So there are certain things that AI does well, and there are certain things that humans do well. And our job is to take the benefits of both sides and to work them into one blended experience. So this gets into what we're going to be calling human-versus-machine frameworks. So there's basically 4 that we've seen, that I've seen in my experience that -- different ways you can combine the human experience and the AI experience together. The first one, and we'll go into these in detail, but the first one is very common, right? I think everybody has had some exposure to this in some way. I'll explain these, and I'll do a little poll here to see what the exposure actually is. But the first one is human goes to AI. And that is like a knowledge base. If you have an agent, if you call into a call center and the agent needs to look up something, that's an example there. And we have a couple of other examples, too. The AI to human is where, again, the AI -- it starts with an AI interaction, but then for some reason passes to human, like an escalation. The -- now we're getting to some more complex ones where you have a human to AI to human, so it bounces kind of back and forth. And those get into some tricky ones, but things like where an AI is assisting a human more in the background, if you think of it like that. And then the AI, so human to AI that supports one, that's more of the human assisting the AI in the background. So that's kind of some tricky logic when you get into setting up those things, but those can be really -- they're more new age, I guess, and innovative, but they're -- we're showing a lot of promise in both of those. So I just wanted to do a quick poll here based on what you see here. Let me see if I can activate this one, this poll here. So which framework are you most familiar with? I'm curious to know. Are you more familiar with the human to AI one in your -- again, this could be in your personal life or in your professional life. So which one have you seen the most? And I bet I know the answer, but I would love to see this, and I'll share the results when everyone is done here. [Voting]

So I've seen a lot for the -- oh, AI to human is winning, just barely versus human to AI. Okay. Are we all done? Anybody else? Oh, one more. Okay. I'll share -- oh, one more. Oh, it's probably a fight between the first 2. Great. Okay. Let me share this what we're at now. So yes, so AI to human, 1, which we plan to -- oh, actually, sorry, yes, that's interesting. Yes. I thought maybe #1 would win. But yes, AI to human, 1. So yes, so for -- these are definitely the most common. These are very good. Ooh, I didn't expect anyone to say these really. These are -- I'll talk about these -- the advantages of these 2. But for the first 2, yes, I think we've all had a lot of exposure to those, so that makes sense to me, and hopefully to all of you. Okay. So let's get -- jump into these frameworks and what that means. Do you have -- Shyamala, anything else you wanted to add to this slide before I jump into them?

No. I think this is pretty much like as expected. Many people know 1 or 2 pretty much, and then that's what they had. And we'll be, anyways, diving deeper into each of these, so that's when I'll give some of the examples, so.

Great. Okay. So for the first one, the common one, right? It's the human to AI. So we already talked a little bit about this where you call into a call center, the agent doesn't know the answer, and then they have a knowledge that could be powered by AI that's fueling that. So they'll type in their question, the knowledge base will come back with their answer, and then they'll read the answer on the phone, right? That's kind of just an assisted experience for a human. And then Shyamala, you had another good example there, too.

Yes. So pretty much like what Celene said, even in the consumer side, like, let's take an example of the autonomous vehicles, people will have all sorts of questions related to the vehicle, right? Like can I smoke in the car or what do I do in case of an emergency and things like that. Now some of those things could be, again, something where human to AI model can be used, where the knowledge base is giving you all of that information and then the agent can help the users.

Yes, exactly. So yes, a very standard model. I think it's been around for ages. Not -- there's some efficiencies we can definitely gain with that, making sure that your agent is -- does have an AI with platform with natural language built behind it. That's a really important piece to make sure that it's quick and fast. But generally, this is a very common practice that I think everybody probably implements at some queries per day. What's becoming more, I think, more on the forefront is this model, right, where you have your AI to human. And there is a lot of learning where -- this is what, again, we do with the majority of our deployments as well, where, yes, you start with an AI bot first and then go to a human if you need it, right, lower cost to higher cost, right? That's the idea. And I'll talk a little bit about some best practices and examples of this. So just think of one, for example, IVR escalation. Oh, just I hear a couple -- feedback. If we could go on mute, that'd be just -- yes, through with the feedback. But there's the IVR to escalation piece, where, again, call an IVR, you try to self-serve within the IVR. You could have some transactional journeys. The IVR might respond to you, might give you that answer. If it doesn't for whatever reason, if they need to talk to a human, didn't give the right answer for whatever reason, that's when you have that ability to escalate. So that's one example. Shyamala, you want to give your example as well?

Yes. So similar to how the IVR escalation is, it's pretty much, we call it, like handoffs from a bot to the human. And some of the use cases or examples where this could happen is, for example, if you look at someone wants to order pizza and then they are going to the bot experience, and all of a sudden something happened where the bot is not able to help because the payment is not working anymore. So now they need to hand off to a human. The other example, which you could think of is the user is an autonomous car and says like, I want to pull over. Now this is a high-risk situation where the bot understood what it needs to do, but it will not do it because there are stakes involved. So it will hand off to a human to take care from thereon. So those are some examples where the handoff from the FAQ could happen.

Exactly. Yes. Thanks, Shyamala. So I'll get into -- just an example. We talked about the IVR example just to see what that looked like. So again, this is a very familiar model, I think, most everybody, well, generally knows. But yes, what you're looking at is this one-way path, right? So the users are coming in, in this example, do IVR, but it could be the bot, a digital experience, it could be anything that's routing it to 1 or 2 things. So typically, in our experience, what we do like to do is we try to route to the intelligence first, the AI first, which is the orange. If that didn't resolve, you can go to agent. Sometimes they can bypass. You have routing or you have -- if somebody just says agent, we'll route them right to agent. But you see, I mean, the lines here kind of tell a story, right? We are forcing a user down a path when -- sometimes you don't need to follow that path. You don't need to force them down this one-way street. I'll give you an example. So for anybody that drives on highways, I live near Toronto, and we have kind of a massive highway system here that -- it's not very good. It would give you signs and directions. So very quickly, you could easily be on the collector lane, when you wanted to be in the express lane or express lane when you wanted to be in the collector. And once you're stuck in the express, you cannot get off your exit. You're stuck. You're stuck there, right? The same idea, I think, applies to this, right? You're forcing them down a path, and there's no option, there's no help. If you get stuck on an express lane, you can't really easily get off and get help if you need to or come back, right? You can't turn around and come back. So for anybody that's been in that experience, you missed your exit or whatever, that's what you're doing -- we're doing to the users now, right? We're setting it on a path and we're giving them no control. So that being said, there are some best practices we can do with this model to give them a little -- and make that experience better, right? And then I'll talk about the other frameworks as well that's going to assist that. But in this case, I mean there's lots of things they can do to give, yes, that user the key, so to speak. So I'll just cover a couple of these. Shyamala will cover a couple. But I wanted to talk about, one of our use cases here is live now on our website, is Bed Bath & Beyond. We do this again for a lot of our different customers as well, too. So it's just general best practices we use, but you can see a lot of these things being applied to this and other clients we've deployed. So first thing is around expectation setting. Users will come in at any time of day. They're going to come in at any location, for example, a Google entry point. They're going to come in to your website. They might call in through your phone. They might call in through your phone because you force them to. And then they want to actually talk to chat or vice versa. Maybe they want to call, but you've forced them into chat. So the key is to just make sure you're meeting them where they are and what they want to do. Transparency is around informing a user so -- it's a bot. With the AI and HI blending, it's just really important to be very precise of who you're talking to at what point and also what it can do. Like we talked about before, the bot's there for a certain reason, again, we're talking IVR, digital. But the bot there is there for a certain reason, and it can do certain things. And it's our job to educate the user what it's capable of, so that if we escalate them, they're not surprised, right? We're telling you, I'll provide you an answer or get you in touch with an agent. That's its job, depending on what your escalation strategy is. So flexibility. Don't force the user to use the bot. If they want an agent, let them talk to an agent. But it's our job to make the bot experience easy so that they can talk to that bot. You can initially ask the bot to triage that response for you. But if the end result is, they'll call a number, your goal is to not keep them in that loop of using the bot because that's going to work -- again, one-way path, restricting, not giving them the control that they're going to be needing. Shyamala, do you want to cover the other 3?

Yes. So the next one is error handling. And this would happen often, whether you are interacting with a chat bot or a voice bot. And mainly any NLP systems would often have errors. It could be that the bot did not understand what the user said or it could be something else, like little information was given, needs more information. So no matter what, there has to be a clear error-handling strategy. If not required, don't ask the user to rephrase the questions. Only if required, we need to do that. So one of the examples could be that the user said, "I want pizza for 3 people," and you're not confident about that situation or the number, so you can always ask the user in that situation like, "Okay, was it 3?" Instead of saying, "I'm sorry, I did not get that. Can you repeat it?" So you could be a little bit smarter there saying like, "Was that for 3 users?" So you need to have proper error-handling strategy, when to use explicit, when to use implicit confirmations and so on. The next is humanization. We need to also make sure that the AI is more empathetic, personalized and transactional. So if we know about the user, we can use some of those personalized settings to make sure that we are creating a more personal experience. And it has to be empathetic throughout. So then maybe, when you are interacting with the bot, there could be the things like, I was giving the GPT example. When the user said I'm feeling terrible, should I kill myself? So now that is exactly where the empathy was missing, right? It said you should kill it -- yourself or I think you should. So that's exactly where like if someone is coming to you with some sort of thing, especially with the customer-related things, there will be a lot of similar situations. Make sure your bot has the empathy through words, through tone of the voice and all the different kind of attributes out there. It needs to mimic most of the human interactions, not exactly like 100% of it, but to some element which will help elevate the experience. And the third one is privacy, which is one of my favorites. We need to ensure that only if required, we need to collect the data about the user. If you don't need that, we should probably not be asking the user for unnecessary information. I've seen a lot of bot when the user says, "How can I help you?" they start with, can I get your e-mail address first? Okay, do you really need my e-mail address in order to fulfill the request? If not, do I have to give that? So think about those kind of things, like when and at what point do you need to collect data and then do be informative about how you are going to use that data or why you need that. So over to you, Celene.

Great. Thank you. That's great coverage, so. Yes. So lots of best practices we learned from this approach. This, again, being the most one of the more common approaches, we're learning stuff all the time. And things are changing with the introduction of messaging and things like that. So a segue into that is, yes, async messaging, for those who don't know, is -- think of things like more texting, right? You're in Apple Business Chat or you're in Facebook Messenger or you're in Google messaging, and you're talking in that platform rather than chatting online or on the phone. It's a different experience where there's a concept of waiting. You're waiting for your response. You might wait a whole day for your response, not from the bot necessarily, from the agent, but that happens, right? What -- the concept of no starting over, right? So our -- this starts to shift our pathway of the one line, right? Now there's no concept of I'm going to put you on a straight line and you're going to follow the human-to-AI or the AI-to-human path, right? It's not that anymore. Now you're into when can the bot help you, when can an agent help you, and at what point do they go to a bot, what point do they go to an agent as they jump from day to day to day, as they're having this conversation in the text window, right? So async messaging is just a new paradigm that's just becoming very, very common. A lot of the tech giants are pushing it. A lot of users are pushing us in this direction, too. So yes, definitely recommend async messaging, but the bot experience within an async messaging experience is just really important to nail down. So I just wanted to talk about some of the futuristic things of this, things working in the background to help us, right? So human to AI to human, some really cool things you can do there. Think of bot delegation and agent automation. So bot delegation, again, the human comes in and says, "Oh, I don't want to -- I need to collect this information, but I don't know how to do it, bot, you do it for me." The bot -- you could send out your bot, collect that information, bring it back to the human. Saves the human a lot of time. It also helps for privacy, right? So if you ever -- if you think about what you call the banking systems and the agent will ask you, please put in your pin number. The agent doesn't ask you for that, your -- the bot does, right? So it's that idea of you're delegating the bot to do certain tasks for you because of privacy or efficiency gains. And then agent automation, it's just the way I -- just looking at time, I'll just cover 1 example from each of these. So we have to make sure we're questions. But yes, lots of examples of the AI working in the background for humans. And -- oh, sorry, Shyamala, any other examples there you want to touch on before we go to the next one?

Well, you pretty much covered it, so...

Okay. That's great. Yes. Just the last example is around the humans working in the background for the AI. So bot supervision is, again, just imagine, on the right-hand side of the screen, you see, again, the agent. They're not talking directly to your user. The user is coming in, they're asking the AI questions and then there's a monitor, somebody monitoring and supervising the bot in the background. And they can see, "Oh, the bot is stuck. I'm going to get the bot unstuck." Or "Oh, the bot gave the wrong answer. Let me intervene," right? So it's kind of like a lead manager where -- but instead of managing an agent, you're managing a bot. So that's a really popular thing that we're starting to see a lot of traffic with. And the cool thing about that is that, that data from that user -- or sorry, from the agent supervisor can be fed back into the bot so that next time, it learns, so it's a really good way of optimizing your bot that way too. Yes, so just one example of -- there of how you can use humans in the background for AI. So yes. So basically, in summary, yes, I just wanted to touch on the -- as you see here the -- we already -- we know what AI is good at right now. We know what humans are good at right now. And the goal is to take those pieces, the empathy, the personalization, the humanization of that, as much as we can put that in the AI, but where we can't and where we can advance the technology to do the truly understanding something versus truly not, that's where we can use that agent -- or sorry, the humans and the agent piece to funnel that in and to bridge that gap together. So yes, that's our quick spiel of what we wanted to cover around bot blending. And I would guess we're opening up for questions now.

Lovely, thank you so much both for your presentation. Yes, we do have some time for a Q&A. [Operator Instructions] Celene, if you didn't already see some bits on the chat there, quite a few other people voting number two, in fact, I believe was AI to human.

Right. So I'd say, yes, I do have one more poll, yes, sorry, we can talk about. So yes, we didn't cover that. Do we want to cover it now or jump to questions? But yes, sorry, forgot my last poll here.

Yes. Let's activate that poll.

Okay. Sorry. Yes. And because the best practices I feel are really -- that's a really good learning we can all take away from this too, so let me just share my screen, sorry. Forgot our last...

No worries. You don't need to share your screen. I've just activated the polls. But if anyone wants to -- if everyone wants to click to the Poll tab on the right of the screen there, and you should see Chat, you should see People, you should see Poll. If you click on poll, you can vote with your answers in there. So what best practice do you feel is most important? Celene? [Voting]

Yes. Yes. So again, I don't know if there's a right or wrong answer to this. I think Shyamala might pick privacy. That's one of her favorites. But yes, I mean, I -- again, just based on my experience that I've seen, all of these are important obviously. But if I had to probably pick, I'm going to cheat and I'll pick 2. The -- I've seen most -- the flexibility and error handling, to me, are in the -- all the conversations that I've read between agents and AI, and humans and AI part, it's really those 2, right? The flexibility can be very limiting in the experience, and they will be driven away from AI technology if we force them to use it. The idea is to invite them to use it and make it good. So if we want AI to stick, we cannot force them. So that's one thing I've seen. The other piece is around error handling. Again, it's back to that we're forcing them to stay in the bot sometimes. We're forcing them into an interaction, and that will drive them away. So if we want adoption in these technologies and you want rely on the bot and your AI, it's a balance between making it really good, so they want to use it rather than forcing them to use it. So then, otherwise, you won't be able to reduce those higher-cost channels. So yes, let me maybe share these results. So far, they're pretty cool. Yes, so for everybody, error handling was the winner. So yes, so good for that. But no, I mean again, all of them are very important, but I think, yes, it's really interesting to see what everyone else says as well. That was very cool.

Right. Yes, I would like to add, like although privacy is my favorite, error handling is very important. Transparency and all, like, is also important. If you look at what Google Duplex does. If it is calling on behalf of the user to someone else to book the appointment, it actually never acknowledges that it is a bot. And the way it speaks, the tone and all of those things actually can confuse anyone, and that's where ethics falls into place, like you don't have transparency. You're not telling the users exactly that hey, I'm Google's Duplex bot calling on behalf of a customer to book an appointment or any such things. So that is also important because you cannot keep the users at dark and say like, well, I'm giving you the best experience possible.

Exactly. Yes. And then again, it goes back to your forcing people to use the technology they may not have approved or want to do. And then it also, I think the issue of transparency is if you set it up to be a human, they're going to talk to it and act like a human. And like we said before, right, AI is great, it's not quite there to become that true Turing Test, right, where it can pass for a human all the time. Again, Duplex was very close. So that's where it gets to be -- yes, one, you don't want to set up for failure. And two, you don't want the user to feel they get an emotional feeling if they feel like they've been tricked, right? So yes, definitely agree with that as well. Yes.