Talkdesk Calls Out the AI Hype Machine – And Offers a Way Out

Pedro Andrade discusses why most AI deployments fail, why your KPIs are lying to you, and how the Talkdesk CXA Operations Center changes the game 

Rhys Fisher, Associate Editor at CX Today, sits down with Pedro Andrade, VP of AI at Talkdesk, to dig into one of the most pressing challenges facing contact center leaders today: deploying AI at scale without flying blind. 

With AI agents now handling real customer interactions, the stakes have never been higher – and as Pedro makes clear, the gap between expectation and reality is wider than most vendors want to admit. 

AI is everywhere in the contact center, but visibility, governance, and accountability are lagging dangerously behind. Pedro Andrade pulls no punches in this conversation, breaking down exactly why so many AI projects underdeliver and what it takes to fix that. 

🔴 The black box problem: Pedro explains why deploying AI without real-time observability is like "driving at high speed, blindfolded" – and how Talkdesk's

Transcript

SPEAKER_01 0:08

Hello and welcome to CX Today. I'm Reese Fisher, Associate Editor, and today I'm delighted to be joined by Pedro Andrade, the VP of AI at Talkdesk. Pedro, thanks for joining me. How are you doing today?

SPEAKER_00 0:20

Hey, thank you very much. I do appreciate the invitation.

SPEAKER_01 0:23

Thank you so much for no, no, pleasures absolutely all ours. I'm uh really looking forward to the chat. You know, today we're gonna be we're gonna be talking about the the Talkdesk CXA operations center and guess and the wider implications for the CX space around that. So maybe just to start things off, Pedro, could you just explain to us what the solution is and maybe why Talkdesk has decided to introduce it?

SPEAKER_00 0:46

Right. So before explaining what the solution is, is kind of uh framing what the problem is that we're trying to solve, right? So, Rhys, when you are launching, when you are bringing AI into the automation, into the automation space, especially when you are looking at the CX space. Um, one thing that it's clear, especially recently with uh Gentic automation, is that you one of the things that stops companies from adopting AI at scale, and I mean at scale is making it available to all their um customers, patients, um, is the fact that they have low visibility on what's going on inside AI. Most of the times is governance issues comes into place, and this is one of the key blockers of adopting AI at scale. So to solve this, companies are starting emerging with a new role, which is the CX operation management management or manager, right? And so this new role um has a responsibility to oversee the work that is being done by AI. And this is because we need to frame that AI now represents labor. And this represents labor, labor the same way that you have supervisors to manage human labor, you have the operations manager, the CX operations manager to oversee this hybrid workforce between human and non-human labor. And that place is the CX operations center, is uh observability and validation and governance platform that will allows you to understand what who is doing what and what is the quality that is coming out of that, and what is the right balance between uh human labor and AI labor in the CX space. That's pretty much it. Um, that kind of represents and frames the new role and the solution for this new role.

SPEAKER_01 2:58

Yeah, it's interesting. You mentioned governance there sometimes being maybe a bit of a stumbling block, and it's obviously it's it's become uh a very serious CX conversation because of the proliferation of AI. I guess, in your opinion, how does giving leaders these real-time observability into AI agent behavior? How does that actually change how confident perhaps they feel about actually implementing AI?

SPEAKER_00 3:24

Well, it makes, as I used to say, the world difference in the world. Um, this is the difference between you launching AI, not understanding what is going to be, what are going to be the implications, what's the quality that you are doing. So basically it's like driving a car and high speed blinded, right? When you what you basically need to do when you are launching AI, one of the key concerns that we are getting is understanding what's inside, understanding uh observability, understanding how an AI agent able to understand, decode, reason, and make actions when he's having uh an interaction uh in any me, any any channel with with the customer. So when you look at that in that perspective, um observability means and deeply understanding when there was a response, deeply understanding what were the several steps and the reasoning behind those steps. That is going to give you the power not just to understand, but then also to do course correction. And that is a fundamental capability. That is what removes the blindfold when you are driving this type of uh machinery, right? And um it's a fundamental step. I would say that it's totally impossible or irresponsible not to understand and just trust a black box machine when there is so much at stake. There is trust from your customers, there is revenue at uh at stake. There is especially um the confidence that helps you then to increase the automation that is so needed here um in this space. So, in sum, this is what it's all all about: understanding, deep understanding, give you actionable data to your course-correct AI, remove the black box um concept from out of AI and give you the tools that you need to confidently deploy AI at scale.

SPEAKER_01 5:45

Yeah, yeah, thanks. I really enjoyed uh I enjoyed the driving metaphor, I think it's uh it's a really good way of kind of picturing things. Um I guess I suppose a big part of this solution is it's and a big part of the CX space in general right now, actually, is is that relationship between human and AI. You know, a lot of organizations are already deploying AI agents or almost everyone in in some capacity, but I guess part of what you're saying is that quite often these are designed for humans and not designed for AI. What what's actually going wrong, do you think, when that gap isn't addressed appropriately?

SPEAKER_00 6:22

Well, that's true that there is um statistic where AI projects don't go sometimes they don't go as well as um Pictet expected, and in my perspective, there are two problems. The first one is um inflect inf uh inflated expectation that is uh mostly injected by technology providers. So technology providers are part of the of the problem. They are trying to make it sound like AI is amazing, they are jumping into this wave of uh awesomeness and amazingness, and so basically I I see some competitors sometimes, they bringing up um fantastic stories or fantastic promises that AI magically is going to solve those problems. And that is part of the problem. So basically, you create an expectation, high expectation, and at the same time, CEOs and CIOs feel the pressure to increase efficiency by reducing costs, by bringing more automation into CX. So when you blend those two things, you basically are creating a bump in terms of uh expectation. Now, the second problem then is going to collide with this incredible amount of expectations, which is the technology itself. The technology itself, RIS, is far from something that you can magically touch and is going to transform everything into gold. That is not the scenario. Building AI agents, no matter if those AI agents are being deployed as customer-facing conversational AI, or if they are being used to support agents during conversation, like a peer having um assisting during a conversation, or if you are using those AI agents to assist in back office operations running as an engine that is running and moving data from place A to place B and taking actions on it. No matter what the situation that you are doing, the reality is that you need to proper set up those AI agents. And this means that you need to understand how they work, provide clear instructions, click provide clear prompts to so these systems can execute without giving having second interpretations on those instructions. That's when one of the issues when that happens when you are running those agents. And as well, very important, is the data readiness. You can't imagine how many situations I went through on customers that they believe that they have fantastic knowledge bases. And but the reality is that most of them they still need there is there is work to do. There is work to create, to help those customers to curate those knowledge bases, those knowledge articles, those governance procedures. So basically, AI can read clear clear documentation and with clear data and clean data, so you can actually then provide the promise. So you see here two worlds, right? A world of super high expectations, there is money involved, people want to do that, they have the budgets, there is promises being made, but then the delivery, delivery requires certain work and certain amount of work. Um, we're trying to bridge and talk desk, we're trying to resolve this problem. We we we have based on the experience that we have, we create reasonable expectations of what can be actually achieved after an audit on what is your current data readiness, what is your current capability. So we do an advice on which type of solution should the customer employ. So, and then create um a plan according to their maturity, maturity level, so they can adopt AI at scale, but always responsibly.

SPEAKER_01 10:55

Yeah, I really like that framing of kind of two worlds of AI. I think that's a really nice way of looking at it. I might uh I might actually steal that for myself, Pedro. Maybe put that in an article. But I wanted to pick you up on on the data point because I think that's an interesting one. Because we hear a lot, you know, from people like yourselves, other vendors, they go to these customers, the data isn't ready, it's not in an appropriate condition. Why do you think that is, first of all, because you mentioned like a lot of people, they believe that it is, you know, that it's uh they've got a strong data set, but when it comes to it, it's not it's not prepared for AI. So I guess two parts, why do you think that is, and what can organizations do to kind of to improve that?

SPEAKER_00 11:37

So the biggest the biggest problem uh that I'm seeing when it comes to that point about the quality of the data, the quality of the APIs, right? Um is that is that understanding or is it created by the problem of understanding how then the technology works and what it takes in terms of requirements? Uh let me give you an example. When you have um when you have when you see a demo on a voice AI, everything sounds absolutely amazing. Supernatural voices, um, very fluid conversations. But in practical in in real life, RIS, what happens is that the system is connecting is to connect to an API on their procurement system or their patient record system and a healthcare system, and that API may take two, three, or four seconds to retrieve the data. And that is when the problems start to appear. Because there was they saw a demo, that demo was amazing, it's super fluid, and all of a sudden there is a gap that needs to be filled. There are techniques to overcome that. But the first thing that you that it comes to mind is that those issues, this uh data access, those APIs, they're all going to work magically. And the reality is that you need to prepare. If you have an API that takes eight seconds, you cannot just feel, keep feeling, uh, let me take a look at that data for you, uh, hold on just one more second. You can't do it for eight seconds. That's not the expectation when you are bringing automation. So there's a need to optimize and to understand if that API is the right one and to prepare that access to that data. So you basically you deliver up to the promise of a fluid, natural-sounding conversational AI experience. And that is one example. Another example I would like to give you is when it comes to knowledge articles, for example, knowledge bases, right? Um, I uh in my life I saw um situations like this. You have in the same document instructions to provide to human agents, and they are in a given color. And below that text, there is other part of the text that is written in other color. And there's um there's a concept and there's an uh a knowledge across these human agents that whatever is written above in in one in the in one color can be fully disclosed to the customers, to the contacts. But the one that is on the other colour cannot, it's just for internal reference of the of the of the agent, cannot be disclosed to the agent. So so to the customer. Now, what happens? If you just expect that your AI agent connects to that, those knowledge articles that have this um, well, this uh definition of um data boundaries by inking them with a certain color, that doesn't work. See what I mean? That doesn't work, and that that that's when this is a full stop no-go. You cannot proceed with a knowledge that is based on those codes of colors of what can be disclosed, cannot be disclosed. An AI agent is going to read through all of that. It doesn't know that rule of whatever's in blue you can say, whatever's in red you cannot say. So this is yet another example of that readiness that I'm expecting. Um they they they basically can have um in across human agents, they give instructions, don't say whatever is in red, and that's okay. And eventually you have mistakes as well, provided by human agents as well. But at at in the era of AI, these mistakes, if they happen, they will happen at scale. So it's a high risk that you cannot just cannot afford to run.

SPEAKER_01 15:56

Yeah, yeah, kind of it leads back to what we were talking about earlier, a little bit. I guess it's these systems that are still built for humans, not AI, so inevitably you're gonna have those those frictions like you outlined there. I think there were some good examples.

SPEAKER_00 16:09

Yeah, exactly.

SPEAKER_01 16:10

That um cool. I want to Pedro, I want to talk to you about um kind of metrics, I guess, in in the age of AI. Because I know TalkTest has spoken about this before, this idea that some of your traditional contact-centered KPIs, things like average handle time, they they perhaps don't tell the full story or the correct story, you know, in in an AI hybrid world. How do you think CX leaders should maybe rethink how they measure performance?

SPEAKER_00 16:38

Well, that's a great question. So, Rhys, um let me give an example. And this is a true story, right? Um traditional metrics on the contact center, they are based on a few very well-known KPIs. I'm going to give an example of the Havergendal time. It measures the time between start and end of an interaction. And in most times, most of the times, contact centers pressure is to reduce this KPI. The shorter time possible to resolve an issue, maintaining the quality, but as much as possible, reducing the average time. So when you start introducing AI, what is going to happen? Machines start taking on the low-hanging fruit of those interactions. Where's my order? Opening hours, locations, can I park there? And this type of low-hanging fruit usually is what takes the faster, quite the faster type of interactions. So when you start removing those interactions to get into a human where the avargental time is calculated, what happens is that the humans start getting higher avargental times. And if you don't look at this holistically, then you are in trouble because you will need to explain to someone why are you investing that amount of money in AI, but your KPIs are getting worse. So you need that's why the hybrid workforce and the management of a hybrid workforce is managed in in different terms. You you cannot just rely on the old metrics. You need to understand that what is automated deflection. And that when you start presenting that you have a certain amount of deflection, or uh if you are having um, if your case is not for deflection, what is the impact in other things like first call resolution? You need to start looking at more realistically view in terms of other metrics so you can explain the value that AI is bringing. And the my final note on this point in terms of KPIs is that majority of the times providers, technology providers, they they they miss this important part is explaining what is going to be the impact on uh human effort and human labor after you start deploying AI. So let's talk about the elephant in the room. People expect that a manager in the contact center that they basically invest on automation for one of three things. Number one, they want to save money, they want to make more money, or they want to reduce risk. Period. Right? So you here you if they want to, in the case if you are deploying AI to save money, you need to show where the savings are. And you just say that, oh, I'm deflecting 60% of the calls, but you don't see that number reducing into the human labor on the FTEs that you have hired, you need to explain that. So you need a tool, and that's what we are presenting here as part as well of this hybrid workforce management. You need a tool that the more you deploy AI beyond the metrics of controlling AI quality itself, then you can also see the impact on the forecast of human labor that you need. That difference between your current forecast versus the for the new forecast after deploying AI, those are the savings. There are material savings counted in FTEs, not in other type of uh metric that may not tell the world truth.

SPEAKER_01 20:40

Perfect. I think uh yeah, I think that's a great place to enter in Pedro. I really enjoyed that. Thank you so much for your time. I think yeah, it's always great to get time to speak to people like yourselves who are really kind of involved with the technology and really, yeah, getting it from the horse's mouth, I guess. So thank you very much for your time.

SPEAKER_00 20:56

No, I do appreciate that. Thank you so much, Ruth.

SPEAKER_01 20:59

Great. I would also like to just quickly thank our audience as well for tuning in. If you enjoyed this, please do remember to like and subscribe to the channel, and also you can head on over to cxtoday.com for more stories like this. Until next time. Thanks for watching.