AI and Customer Data Strategy: A Deep Dive with Victor Garayar

‍

Victor Garayar is a Computer Science and Systems Engineer and an MBA from IMD in Switzerl and.Victor has 20 years of international professional experience in technology, including Consulting,Artificial Intelligence, and CIO and CTO roles.

Currently, Victor is the Director of Cognitive in Amelia, an Artificial Intelligence company based in New York, named by Gartner as Leader for Enterprise

Conversational AI Platforms. Victor leads implementation, growth, and expansion of Amelia for South Europe and Latin American customers, managing teams in India, Spain, Netherlands, US, Peru and

Mexico. Additionally, he is the worldwide leader of Amelia’s voice service, as +80% of his customers are through call centers, and his team manages 70% of Amelia’s conversations, which surprisingly are in Spanish.

‍

Victor Garayar: 

Hello, Chris, how are you? I'm Andres, how are you?

Host:
Good, good, very good. Super happy today.
Where are you based out of, by the way?

Victor Garayar:
Uh, I am in Kill Plant. It’s a city close by to Seattle.

Host:
Okay, you like the area?

Victor Garayar:
Yeah, I love it. I am in front of Lake Washington, so it's quite nice, especially at this moment, you know, in the summer.

Host:
Yeah, in the summer. Very, very nice. Yeah, the few moments in Seattle that you can really appreciate the sun, the views, and so on.

Victor Garayar:
Yeah, enjoy your three months. [Laughter]

Host:
Well, uh, I’m going to sort of talk about you here, uh, and some of what you do. And if I was to do that, I would say that you are the director of cognitive AI, which, uh, sounds like at least these days like kind of the same thing, uh, cognitive and AI.

Victor Garayar:
But it didn’t always, it wasn’t always. Um, but, uh, we would love for you to talk to us today about, um, what you do and then what’s going on. And, uh, you do this work at Amelia, which with interestingly enough is the company working in this space of call centers and sort of that sort of ecosystem of, uh, helping customers get the answers they’re looking for. And I think that’s a very interesting subject and we would love for you to talk to us about that today.

Host:
Sure, sure. As you say, like, uh, the cognitive AI, it’s always something that what does it mean? It can come in anything at this moment. In our case, starting with Amelia. So Amelia is an AI company. So we are based in New York. Our main product is also called Amelia. It’s a virtual assistant. So she is created like this. She’s a digital employee. Digital employee.

Victor Garayar:
Oh, yeah. We love to, or we like to do that, to call her like that because at the end of the day, so, I mean, it’s going to be a higher wire company, then we are going to train Amelia on whatever the company needs to do, and then we are going to check her performance, review performance, and then probably everything goes well. So it’s going to be promoted and so on and so on. No? So that’s more or less the area. So we have a platform. The platform is created as a digital employee or as a virtual assistant, and this is how we really, Amelia is the place to be.

Host:
Fantastic. So, um, there’s a lot to unpack there. Um, I think for some people it might sound fantastical. So, uh, maybe we’ll do it in sort of small pieces. Um, let’s talk about, um, what exactly the problem is that you’re kind of trying to solve for. Um, so, or sort of how you would think about initially attacking, uh, the problem and then how Amelia emerged from that. Is that, does that make sense?

Victor Garayar:
Yeah, it makes sense. So, for instance, in my team, my team is the one who implements Amelia in the company. So, we start from the problem. What do you say? No. So, usually, companies want to increase their customer satisfaction or they want to reduce or save some money, no? Because they have huge call centers or a lot of human agents trying to respond properly to a customer, and they want to grow. And you know the human capacity is limited.

With Amelia, companies can leverage AI-driven customer insights to better understand their audience and provide a personalized experience at scale. Amelia doesn’t need vacations, and she's available around the clock to support customers whenever needed, offering significant advantages in handling high volumes. This capability, combined with predictive analytics of customer behaviors, allows companies to anticipate customer needs, further boosting efficiency and satisfaction. When customers come to us, they’re looking to save money, increase NPS, and raise overall customer satisfaction by optimizing the interactions and response times. The objective is to enhance the customer experience across all touchpoints.

In addition to improved satisfaction, real-time customer tracking helps us monitor interactions as they happen, allowing us to adapt responses in real-time, whether it's through call centers, WhatsApp, mobile, or other social media channels. We can analyze incoming requests, adjust to trends, and even explore which channels are more effective for engagement.

Furthermore, customer data security remains a top priority, especially as we handle sensitive data across various digital and call center interactions. Ensuring secure handling of this information is crucial to maintaining trust and meeting regulatory requirements. Finally, AI-driven marketing enables us to customize campaigns based on insights gathered from customer interactions, allowing us to drive satisfaction while aligning with business goals.

‍

Host: 

Is that in terms of, and you're speaking specifically in terms of, um, being able to answer customer questions? So, there's the call center channel and then there's the social media channel where you're also answering questions there and you're considering both parts in terms of what Amelia is solving for exactly?

Victor Garayar: 

Exactly. And also, it's very important because sometimes, depending on the customer, absorb the company, customers are more willing to share information through, I don't know, WhatsApp or SMS or call centers rather than social media. So, you don't share in your Twitter or your Facebook Messenger, "Hey, this is my ID, this is my passport," etc. So, sometimes, depending on the channel, you will be more informative. So, you will basically answer some questions, but then in probably in WhatsApp, SMS, or call centers, you are more in the automated part, which at the end of the day, you are going to reduce some humanity interactions in that site.

Host: 

Right. So, um, we've talked—so we've talked about channels, different ways that Amelia reaches out. Um, maybe we'll talk about, you know, some of the challenges with, uh... Um, first, we can—we can talk about Summer Challenge at this layer, right? And before we go into the sort of more technical layer of challenges, so at this tier, uh, in this—in this, you know, what is it like to have customers interact with this, uh, you know, at this level? And then, how are these—or how is this interacting with, uh, these other channels?

Victor Garayar: 

All right. So, um, but I mean, by that is, um, just like you said, there's going to be the expectations for each one that make things easier, harder. Uh, recall centers, the assumption is that it's as robust a person as you can get. When you have a call center for WhatsApp, there's a little bit more flexibility because people actually want to speak in less time. Um, and that's—that's my—this is from a customer's point of view, that's my perception of it. But is there something else that even might be more surprising or is more unique based on your experience that you've found? Please go ahead. But, uh, so what is it like to sort of solve for those two places, or what have you found, uh, issues in those two places or all these channels?

 This is very interesting because the channel has their own characteristics, no? So, in the invoice probably—so while you woke up or in call centers, you will have some problems, for example, with names or addresses, you know? So, it's like, uh, especially so in—in my case, I'm in charge of the South European and Latin American countries, so in my case, so the English names that go to Latin names and the combination and so on. So, right now, it's quite difficult because probably John can be written as like a G-O-H-N or Y-O-N, just as simple as that. So, how can you verify this kind of... So, that's the challenge that we find when you need to have verification based on spelling or something like that, or reservations with airlines, you know. It's like, okay, my reservation is A, B, C, one, two, three. No, but for example, if you go for CS, MNTD, so TD or MN is kind of phonetic, it’s quite similar. And if you, for example, right now, we are talking with good quality, you know, nobody is having some noise around, but if you're in the middle of the park trying to call your airline because you are running to the airport and there’s a lot of noise around, so these specific things are very challenging for that from a voice perspective.

But on the other hand, in the chat, the challenges that we find are like, uh, you can put more graphic things, you can put images, you can put buttons, you can put anything to make, in some sort, life easier for the customer. But is that intelligent enough? So what is the balance between being naturally speaking or going for too many options that you don't give the customer the freedom to express themselves? You know, if you find, like, uh, hey, I can help you in one, two, three, four, so people are going to think that it's one, two, three, four, so they don’t think about the option five, V8, or the 19. When it’s different, for example, in call centers, it’s how can I help you? And the customer says, hey, I need about this, this, with WhatsApp being balanced, or being with the written channels, so to speak. So, the balance between options or graphics compared to more freedom, or just write whatever you like, is something that usually we have as a challenge, you know?

Host: 

Right, so just to summarize the, uh, expectations on voice calls are, it's very open-ended. Expectations on written channels is usually, you ask users to select a bunch of options, but that fifth option that doesn’t exist might be the issue that they have, so you have to give those options a very well-bounded... And you could also do some chats where, um, you know, you’re just having a conversation on chat, but that has its challenges as well. And then, just like sounds and identification of people has its challenges. So, um, what I sort of—so, so basically, how do you solve this? How have you guys solved those problems with that? Like, how all those different layers, um, the identification problem, the open-ended question problem, the bounded question problem, how have you guys attacked that? Uh, because those are very general, and then you guys also have to apply it to every possible industry, so you know, which is going to be the next question. So, uh, just, uh, you know, show us how you, like, talk to us, talk to us about how you’ve been able to bound that, you know what I mean?

Victor Garayar: 

Now, that's a very interesting question because, in the case of a call center when they need the verification in terms of street names or what we were talking about, what we are doing is to be a real omnichannel platform. So, in this case, for example, it’s like the customer is calling, and then we have the number, and we just verify that this number, for example, comes from a mobile. Okay, if that’s a mobile phone, then we say, okay, we are sending right now an SMS, and please send me your full name or give me the address where you are. No, so then you receive the SMS, Amelia is waiting in the code until you send the SMS, and then once, I mean, the receipt will say, okay, I received that your name is Victor, so then please, let’s continue with the next questions of the call. So, that’s everybody that we have right now, the balance between having a real omnichannel platform in which we’re still in the call, and then we can go through, I don’t know, WhatsApp, SMS, or whatever it is, and then continue the call with the customer without any issues.

So, that’s one side of the problems, and on the other side, for example, with written channels, what we start is like—we make phases of the implementation because right now the customer is used to the interrupt, is basically with buttons, with options, and so on, and usually, what in written channels or in social media, WhatsApp, SMS, what they expect for is to start with a few use cases. So, we usually start with one, two, three, four because you only have the one to three, four options. But then, what the customer is getting used to are, and probably we open some, and we open the questions in order to get what other kind of requests the customer is asking for. Then we select or we can analyze all the information, say, hey, the fifth option can be this, and then it will start opening. So, we'll start closing the implementation with one, two, three, four, but then we start opening based on how the customer behaves, what the requests are. So that's more or less how we address these requirements, at least for the written channels.

By leveraging AI-driven customer insights, we can understand customer behavior and optimize our processes further. With the help of predictive analytics of customer behaviors, we can anticipate what additional information a customer may need before they ask for it. Utilizing real-time customer tracking helps us monitor interactions and adjust in real-time to ensure smooth transitions across channels. Throughout all of this, we prioritize customer data security, ensuring that the sensitive information collected during interactions is kept safe. In addition, AI-driven marketing allows us to tailor responses and services based on customer preferences and previous interactions.

‍

Host: 

You're doing something, I think, is unique, which is a sort of generative questions in the sense that it learns in the call what questions the next question to ask, but it's not necessarily a pre-existing list; it's a list that is something that's emerged from the conversation. So how do you... is there a way to make sure that it's always... it doesn't hallucinate, and the question is trying to ask? Or another way to make sure, but try to bind... is there a way to bound... because these models have a tendency to hallucinate. So if you're allowing them to generate questions based on the context of the user or at that particular moment, it can hallucinate. So how have you been confident that every emergent question is... you know, it doesn't take much. It may be even a word, a word that, you know, that's presence or past tense or some sort of like grammatical addition that can make it mean something else and lead the customer in a different direction, even though it's mostly correct, you know, in terms of just even like identifications of things? So, is there... that's a very hard problem because the hallucination is actually a feature of AI. So how have you guys been able to deal with that?

Victor Garayar: 

That's a very interesting question. So what we usually do is like we separate it into three layers. Okay, the first layer is the intents or the requests that we know that exist. Okay, even though we have the four options that we were talking about, we know that usually customers go for that fifth to the fifteenth. So, we try to train the model good enough to capture those requests and say, hey, I'm not trained yet, but I would love to attend this kind of request, so let me transfer you to my human colleague.

Then, what we do is like at the second layer, depending on what kind of documentation or information they have. So, we use generative AI in turn. So, depending on each documentation that they have, they have policies or they have some documents that we can use, or maybe information from the web, and saying, okay, we can answer those kinds of questions based on this specific information. Okay, then probably you will have some hallucination, but we try to encapsulate as much as we can so it's not like a bunch of documents. We need to really check that the information we have doesn't overlap so much among each other and so on.

By leveraging AI-driven customer insights, we can refine this process further to understand customer needs better. With predictive analytics of customer behaviors, we can anticipate potential questions or requests before they arise. The use of real-time customer tracking helps us understand how customers are interacting with our system and adapt in real-time. As we continue refining these processes, customer data security remains a top priority, ensuring that all the data we handle is protected. Finally, AI-driven marketing plays a role in understanding customer preferences and tailoring our responses and offerings to their needs.

‍

Host: 

Why is the overlap a risk, if you don't mind me asking? Like, why is the overlapping documentation becoming an issue?

Victor Garayar: 

That's a really good question because it makes you hallucinate more because you're doubling down on the data.

Host: 

Exactly, exactly. So, probably... I don't know, if you are calling about, I don't know, to block a card, but then you have the... in the banking industry, but then you have the block of the card, block of the account, uh, to block, I don't know, your mobile user, no, in the mobile app. Blocking can be anywhere. So it probably... blocking part can overlap.

Victor Garayar: 

I'm saying a very simple example, no, but the blocking of some sort can generate an answer that does not necessarily address what you're waiting for. So, you are giving for it to reset your password in the mobile app, and you're asking for the blocking of your credit card. Not so, this kind of thing probably may happen, so that's why we are very careful on what kind of documentation or what kind...Of information we are including into the model, so in B2B you don't, so you cannot afford so much those hallucinations, that's liberal, you know, right? So there's a lot of cleaning that has to happen. Uh, we don't know for this to work out, you know, and so this I think this is a good segue into the idea of integration because this is a very general solution. As general as it is, it still needs to massage quite thoroughly into a company's context because it’s essentially a generic model that once you massage into a company's context, it can have utility in the space of communicating with customers from apps. So, you know, that's quite the challenge. I alluded to some of the ways that you attack that challenge with deduplication of information, but if we were to start again and talk about in general how do you approach integrating into organizations and sort of make it, you know, plan out how this solution can sit in the right place and have conversations with customers, with there’s just a myriad of technologies that people use and a myriad of ways that they store it, there’s just so many opportunities for a problem. 

Host: 

So, how do you guys systematize that in a way that your outcomes are consistent?

Victor Garayar: 

That’s a very good question. So in this case, what we try to do is like, uh, we are very close to the company in what sense? To check the data. At the end of the day, AI goes to data, you know, goes through data. So, and we want to make an impact that should be for the requests the customer does and that has the more volume, so to speak. Okay, so what we usually do is either it be since call center or written channel, whatever we were speaking, so in terms of channels, all we try to do is get as much data as we have. So why? Because then we can answer some questions: what kind of requests the customer usually does, no? Then, whatever the complexities of those requests and finally, if this is something that as a company is able to solve with integrations, because the idea is not only understand, “Hey, I need to block my card,” yeah perfect, I understand that you’ll want to block it. So what is the utility here? So at the end of the day, we need to block the card and say, "Hey, your card which ends in one two three four is blocked right now and in 20 minutes, you are going to receive an SMS with a number so you can go to the bank branch and then you can ask for a new card or you have this link so you can ask for your new card." So they’re going to deliver it to your home. So, this kind of thing is something that is very powerful and impactful for the company and also for the customer.

So, this is something, these three layers or the three columns that we try to do. One is the volume of requests, so we go for the, what I call the 80/20. The 80% of the calls should go for probably 10, 20 or 25 requests. Then, of those requests, what are the ones that we really can understand and finally, from those that I can understand, whether the ones that I really can integrate and can really resolve. We do all of this by leveraging AI-driven customer insights to identify key patterns and trends in the data. Additionally, with predictive analytics of customer behaviors, we can anticipate future requests and adjust our strategy accordingly. By focusing on real-time customer tracking, we ensure that we’re capturing every relevant interaction and providing timely responses. As we integrate these insights, we also prioritize customer data security to ensure that all information is handled with the utmost care and confidentiality. Ultimately, this enables us to drive AI-driven marketing strategies that effectively target customer needs and enhance overall satisfaction.

‍

Host: 

Is there a, is there a, um, percentage that you guys try to attack on like, you know, at this level, this is where the technology has maximum effect? What I mean by that is you said 80/10 rule, which is great, but it's sort of if, of the 80%, we have to have a 60% accuracy, which is really high because you're almost cutting in half operation costs basically. But that you guys really sort of want to go for. So, you know, is there a boundary there? Because I can imagine this, like every industry, the complexity of the questions that they get in these calls is gonna differ. So, um, that even something like as simple as like, you know, the amount of data for airlines from the amount of data for banks, right, in the sense that.. In the sense of how much it moves the airline data disappears every time, right? And banking data just grows in the study and literally is kept in the safe, right? So there's a lot of that. So just speak to that. So these different kinds of, uh, these different spaces, how do you measure a consistent level of success with that in mind, with all this in mind?

Victor Garayar: 

Uh, because you have to make sure that Amelia improves in general as a general solution.

Host: 

Exactly, no, so that's why we are very close with companies to get whatever we need to go for because there are different kinds of strategies. So there are strategy… it’s an example that we have with one with some of the customers. Say, "Hey, I want, I mean, in the front just understanding any request that she has, and then a percentage is going to be resolved and then another percentage is going to be transferred to a human needs."

Victor Garayar: 

Because what we were talking about is like IPL have options, you’ll reduce the scope when you don't have options. The customer is able to express whatever they want, and then goes to the correct part of the tree that you’re... caffeine are called center, no? Either a human agent or with Amelia, so that's a good thing.

Host: 

So based on that, we make the numbers. The idea is that, for example, in every use case, depending on the request, on what we call the use case, we can go for an understanding around 70 percent at the beginning to 90 plus, uh, in three, four to six months, which is very high, I know. Uh, yeah, and it's really fast.

Victor Garayar: 

Yeah, and, uh, but this is something that we could get based on experience and the deal, how robust is the platform right now, and then, aside the, uh, the understanding on the accuracy.

Host: 

Then is how much of these requests we are able to really resolve, and that's a tricky question because then we need the company APIs to make the BF, you know? And the idea is that we usually, what we want to do is like, we make business cases with our customers. That's our difference probably with other platforms. Maybe they do the same or not, but this is what we try to do. "Okay, let’s make a business plan because you are going to put some X money and we want to return." Yeah, a while money, you know? And it's going to return in some amount of months or probably during the first year or during the beginning of the second year, depending on what the scope around, no? So that's unrealistic.

Victor Garayar: 

So looking for 70 to 90 percent of accuracy in the first six months, and then we got, let’s see that we can reach, depending on each case, between the 50 to 70 to 80 percent of them.

Host: 

So you are, at the end of the day, you are talking about 40 to 70, 75 percent of the resolution of each request. So we have customers right now that we are receiving 5 million calls, uh, in the use cases that we can probably cause, uh, a day or five million calls a month. So my is 150,000 hundred thousand calls per day, uh, and we... so from the ones that we are, 40 percent of all the requests, we are able to solve, and from them, we are more or less between 75 to 85 percent of resolution.

Victor Garayar: 

Exactly, and the accuracy of all that we have with my right now is about 9.3 percent in the first intent—in the first attempt, sorry. So we are more in 95-96 because we have two to three attempts. But that’s powerful. For example, in one of the customers, we have others that is more complex, as we said, because we are the insurance and we need to vote for the geolocation, and this is quite tricky because we need some Omni Channel, as I said. No, some of them are able to respond on the SMS with some APIs and so on, but funny things like, for example, the resolution. So we were measuring, um, the resolution of the human agents, and they were more or less between 50 to 55 percent.

Host: 

So from the 100 or 100 goals, they were able to register properly acquired in 50 or 55 calls from the 100. We are in the same level right now with them, and we are increasing, so that’s quite interesting. Or what we do, we are able.

‍

Victor Garayar: 

I don't want to say to replace what the human agent does but at least to support the human agent and the companies, and then the human leaders are able to do whatever goes beyond that because they have the creativity, they have the negotiation part, they have anything else. And we make companies grow in that way. Right, so I think one question is to go in the opposite direction, which is how do you manage with all the success of the platform in terms of its ability to alleviate issues? And when it comes from the customer side, you know, what if they expect a hundred percent? How do you manage that expectation? It's very hard to do, and it’s probably not possible for quite a while, but how do you manage that expectation? How do you show them where the boundary is safest to draw and then flip over to... because they will have some success, and they will be able to optimize or reduce operation costs, but they may have a gold number that they wanted. I’ve had that issue myself where there's a lack of understanding of the technology, which is normal because it's new, but yeah, there’s the expectation for it to replace completely. But it just can't, and then how do you have that conversation quickly? Because a lot of people don’t want to hear it. By analyzing AI and customer data, you can better manage customer expectations and provide a clear understanding of the platform's capabilities. Understanding AI and customer data allows leaders to set realistic boundaries and guide their teams through the inevitable challenges of integrating new technology.

Host: 

No, that's a very interesting topic because this is one of the things why I like to be in this industry. It's because you constantly need to talk with customers, to educate, to educate each other. Because I’m educating all the time, I’m learning all the time from the different processes that each industry has and every characteristic or particularity that they have. So I have almost 20 years of experience in implementing ERP, CRM, and now AI in different kinds of industries, and every time I go to a customer, I find something else, which is amazing. And in the opposite way, it sounds interesting because you need to educate the customers in a way, how to expect from the platform. Otherwise, if you don't align the expectation with the customers, what will happen is like I know what you're saying, "I'm expecting 50," and you know that 60 is already high, but I’m telling you that I am in 70 to 85 and they say, "Oh, that's interesting, but how do you know that?" For people that don’t necessarily know about the industry, 70 or 85 percent is a good number. They say, "It’s just 85!" No, that’s not possible. So I was expecting that you said 100 percent, why isn’t it 100? Yeah, why isn't it 100? And it's... and everybody in the whole world talks about how 100 is when you see these models, fake human beings talking and all that. It’s like they’re thinking that’s what they’re gonna get. And it’s... the marketing is powerful, but the reality is when you get into the weeds, it’s tough to meet exactly.

Victor Garayar: 

Exactly, and what I usually do is make it simple with humans. So, how much... do you have a hundred percent of accuracy? Not necessarily.

Host: 

Yeah, that’s true. So, if you want... every time that you’ll come back with a human agent, it’s like, "Hey, I’m sorry, my last name... I understand from my body here that I have already how to spell... So, G, of what is?"

Victor Garayar: 

So, because you have experience on that, and why you have that experience, because there are so many people around that cannot understand your last name, and that’s accuracy. So, that’s why the hundred percent is not necessarily possible because not even the human agents or human people are able to understand 100. So, that’s more or less what I say. You know, for example, when people say, "Why do you need APIs?" So, I think that you probably need them. So, just with FAQs, you are able to solve some of the questions, you know, and you can contain some of the calls that you have, which is true. But then what I say, for example, is you will hire a human agent just to talk with your customer, and they’re saying, "Hey, yes, the day was pretty good, and you know, I have this document that the policy said something, but I cannot resolve you right now." So, how do you feel about it? No, is that really impactful for your company to have a human?

Agent just sitting and talking and chatting with your customer not necessarily so you need to resolve the problems that your customer has because you are customer-centric, you know, so this kind of examples is how I try to tell customers how what to expect in these kinds of solutions. And having a baseline, we can grow. What I say is like, "Let's start very quick, small, quick, and then the idea is that we can improve in a very good way." So that's why in six months, we are able to go from 70-75% to 90% because we have a reduced scope. We do it quickly, we understand our customers, because every customer has a different behavior, even though they may be in the same industry. So when we understand how the customer behaves, we can improve as much as we can.

Host: 

When you were talking about the voice channel, how do you guys attack the problem of tone, communicating emotion in the language?

Victor Garayar: 

I think one of the interesting things with AI is its ability to do that. We've had autonomous systems communicate sort of very digitally to us, you know, for many, many years. And everyone does everything they can to press zero as fast as possible. So I think Amelia is the difference, I guess. It's actually being able to solve those problems with a person feeling like, "Hey, they got the answer." Because there's a difference between getting the answer or feeling like you got the answer. A lot of customer service is the latter—feeling like you got an answer. So how does Amelia deal with that? It's a very difficult question, but I think the previous conversation should frame the problem and then see how you get to that point. How does Amelia do that? You know, because that's, I think— I might be wrong—but I think that's what's different from all these other platforms in terms of the previous generation of platforms. People feel like they got an answer, and that's how you get to the 60-70-80%. By analyzing AI and customer data, Amelia is able to fine-tune responses and improve the accuracy of the information provided, ensuring a better user experience. Additionally, leveraging AI and customer data helps Amelia adapt to evolving customer expectations, providing more relevant and personalized interactions.

Host: 

Okay, now that's very interesting because tone is something that is very tricky right now. And, you know, sarcasm is unseen. He's asking that probably when we get to that point to find out what is "orgasm" or not, like the AI is going to be on the next level. To be honest, what we try to do is basically check—not necessarily that tone—but what they say. What do people say? Well, so again, I have to tell you this. So why do I need to repeat or so this kind of expressions that give you a sense of frustration? So then what we try to do is, "Okay, let's stop here. I can transfer you to a human agent. Don't worry." Okay, and on the other side as well. Because what we try to do is like, for example, use cases of outbound calls for collections. You know, you call and say, "Hey, you owe me money."

Victor Garayar: 

Not necessarily is the greatest call that we will have because you owe me money, you know. So how can we address these kinds of things? Not, but in these kinds of states, what we try to do is not only say, "Hey, you owe me money, so when are you going to pay me?" but also, "Why don’t you pay me? What is the reason behind that?" I’m probably thinking, for example, Lincoln resolved that problem. "I don’t have my internet at home every month, so why should I be paying?" Then you’ll find out that not necessarily is it a technical problem that you may address, so information is key here. So sometimes, it’s thinking outside the box in some specific use cases, so you have more information, and sometimes you’re getting better results only with information that you have.

Host: 

All right, okay. I think that's very well put. And yeah, you know, basically as soon as you can divert it to a human person who can better deal with frustration until the technology has gotten to the point where you can really do that. And that's very challenging because you also have to know when it’s wrong. Because sometimes humans—like you said in the beginning—humans... someone.

Victor Garayar: 

So, sarcasm is something a robot might pick up that a human can miss. And what I said to someone is, if you have something that's more human-like, you get all the flaws as well, including hallucination, misunderstanding, all these things. The only difference is that it stays up at night, but it's just a person. You know, you're gonna get everything you wanted.

Host: 

So, let's talk a little bit about technology. We've really framed all the different kinds of problems that you guys are solving, but now let's talk directly about architecture and design. The obvious thing is you collect a lot of customer information, stitching it into the pipeline of your platform. My assumption is that it’s a general model that will take that into consideration, and then it’ll need to be trained. Your model will need to be cleaned. That's the standard for machine learning solutions. Can you talk a little bit about how Amelia is able to deal with these problems and make this happen? For the people who are interested in technology, who are watching, this is really interesting.

Victor Garayar: 

Yeah, in this case, we have a combination of a general model, as you said. When Amelia is able to understand basic questions, like “What is her name?” or “What is her creation?” or whatever we call “the social talk.” And then, what we have in another layer are pre-trained models. For example, when we talk about banking, it’s different than insurance, Telco, Airlines, hospitality, and so on. Based on the experience we have, we try to have some pre-trained intents, just as a support because, like we discussed earlier, every company has its particularities.

Based on that, this is a good thing to have, as it accelerates the implementation process. But then you need some specific phrases or how customers express themselves in order to include that on top of the model. So, we have pre-trained models based on the industry that sits adjacent to the basic Amelia general solution. Wake up the general solution, then we have the pre-train, and then we have the specific training, so to speak.

And another layer is the “tropicalization,” what we call. For example, a word in the U.S. might be “refunding” in Australia or the U.K. The same thing happens in Chile, comparing to Peru, Colombia, or Mexico. So, we have kind of a dictionary depending on the geography you’re in, and you can find those differences. These are the four layers that we have for training: the general model, then the industry layer, then the specifics, and finally, the country-based layer.

Host: 

So, for verification, let’s go backwards from the most superficial layer. If we talk about the geolocated layer, how granular can you get with that? You could even argue, though it might be a bit far-fetched, that you could have a layer for each person, in the way they’re taught. What is the radius of that particular model? Is it based on the country or the size of the town? Does that make sense?

Victor Garayar: 

Makes total sense, and it depends on the company and what we’re going to do. For the experience base, we usually go for the country level because sometimes the cities are too small. At the end of the day, AI is a statistical model, so... I'm sorry, I might have oversimplified that. No, no, just to put it in simple words, it’s not like that. I’m sorry. Iceberg, but at the end of the day it's a sample, so if we want to be very effective, let's use the sample, and the sample is going to be more country-based rather than city-based. So that's my last explanation behind it.

Host: 

Okay, got it. And is that because it only produces valuable results at that scale, or because actually, it takes that much data to get what you're looking for? And you do this per organization, so do you... So what happens if an organization doesn't have enough data? If it takes that much data countrywide to guess a value will set, then what happens to the organization? I would assume that there are certain sizes of organizations you cannot deal with because they just don't have the amount of data you need to provide the kind of...

Victor Garayar: 

We start with something. We start with the general, the industry, and the specific itself. So let's start with the methodology. Let's start small and quick, and then the customer is going to tell you whatever they're going to tell. And then the AI has to improve. The key thing is how fast you can improve whatever the customer says. So that's the most powerful part that we have in Amelia. At the beginning, we have what we call the "hypercare." So we put them in production, and then we start analyzing all the time what the customer says, what kind of words, etc., etc. So at the beginning, the first two to three months, we are able to say, "Hey, Amelia is good enough to have a very good volume, to have a very good impact." It doesn't mean that that's all folks; there is no more work to do. No, obviously, as any AI solution, there is continuous improvement. There is customer change, organization change, and why Amelia shouldn't change, you know? So that's more or less the limit. When we analyze the interactions, we gather valuable insights from AI and customer data to understand customer needs better. By leveraging this AI and customer data, we are able to make more informed decisions and refine our models, ensuring that we continuously improve and adapt to the customer's evolving needs.

Host: 

So, that was the first— the most... that's the layer that you see or you hit initially. And the second... I forgot what the second layer was, or the layer below that. What was that?

Victor Garayar: 

So we have the general, then we have the industry-based layer, and then... Yeah, got it.

Host: 

So for the industry-based, I guess my question there is, how do you go about collecting that? Is it a mango... a bunch of companies that you've interacted with, and they say, "Okay, we're going to stuff this in here because it matches," or did you go out and collect it in general from the web or from a different source?

Victor Garayar: 

Good question. There is a combination of... First, it's based on its own experience with previous customers. So based on what we have trained, I already have six years in the company. So I have six years of experience with different customers, different industries, and so on. It's the same way we have different kinds of NLU architects, what we call them. These guys are the ones who really put the architecture of intents and data, and so on, and then train them on it, and then create the model first. And based on that, they started improving the model. So that's basically what they do. But also, what we have are people like me, who have done many, many works before Amelia, in different kinds of industries, not necessarily in the AI world, but they know the processes. What kind of processes do we have in banking, telco, or in hospitality, etc.? So based on that knowledge, we also create a base. So we have, "Okay, how many intents or how much data do we have in this model?" Okay, we have X amount. Okay, I think we need to increase this by 100 or whatever the amount is. So that's more or less what we do. This is a combination of previous experiences in terms of customers, but also previous experiences in terms of what processes we've seen before. We also gather insights from AI and customer data to better understand user needs, adapting our approach accordingly. As we continually enhance our models, analyzing AI and customer data helps us refine our strategies and deliver more effective solutions.

Host: 

And also, to keep all this in context, are you consistently piping in, and or what's your cycle of improvement? Is it just that the model will continue to improve based on having more and more data? And then, obviously, you know, there are mechanisms to make sure that the data you're getting is net improving the accuracy of the model. Like, for example, you talked at the top of the hour about deduplication, right? So how do you... First of all, are you... is it a live process where you just plug it in and it has its way of making sure that it improves? And then the second question is whether...

Victor Garayar:

This is pilot, what you say, so this is a general, at the end of the day, it's going to be a general model and this is something that we as a company we understand. This is an accelerator; this is not going to be just a plug-and-play, that's not possible because every company has particularities. So what we say is like we put the general model, the last day industry model, and then what we do is like we talk with the customer saying, "What is the comparison of what we have versus what they have? What are the gaps?" And based on that, we train the model based on the differences that we have.

Host: 

So you separate it out and you, okay, so you separate that and you give it to them, right? And then it's theirs, and then sort of you can replace the base, but the base is managed by Amelia, and then we, you know, put it back, but it's not like directly integrated. It doesn't work that way.

Victor Garayar: 

Yeah, exactly, exactly, exactly. At the end of the day, what we do is like a specific implementation of whatever, because what we use is accelerators or assets that we already have.

Host: 

Yeah, well Victor, thank you very much for coming. Thank you very much for answering all the questions.

Victor Garayar: 

Thank you very much for having me.

Host: 

I enjoyed it. I hope so. I tried to make sure it was interesting for you, and I'll try to simplify it as much as possible because I mean, I feel like this is one of those things that you could talk about for hours and hours, but—and you barely understand—but I tried to make it easier to digest and thank you for playing along, so I appreciate it.

Victor Garayar: 

No, no, no, no, no. I appreciate the invitation. As I said, I really enjoyed it. I enjoy talking about these topics because you have so many stories behind them, and it's very interesting. It's very interesting because you learn a lot from companies, so for people like me, I really enjoyed this.

Host: 

Where can I find you?

Victor Garayar: 

Well, if you have any questions, please go to my LinkedIn profile, Victor Garayar. Sometimes, depending on where I am, I'm more so I faster than other days, yeah. So yes, reach out to me, and I think the best way is going to avoid traveling being, and then based on that, we can talk about any topic. So if you have any questions about your current initiatives or how AI or virtually use can leverage from the company or how it can impact what they can make and so on, so happy to talk. As you could see, I don’t need a Peter Pixel; I can talk many, many hours.

Host: 

Oh Victor, thank you very much. I appreciate it.

Victor Garayar: 

Thank you so much, Chris. See you soon.

‍