#315 Jarrod Johnson: How Agentic AI Is Impacting Modern Customer Service

In this episode of Eye on AI, Craig Smith sits down with Jarrod Johnson, Chief Customer Officer at TaskUs, to unpack how agentic AI is changing customer service from conversations to real action. They explore what agentic AI actually is, why chatbots were only the first step, and how enterprises are deploying AI systems that resolve issues, execute tasks, and work alongside human teams at scale. The conversation covers real-world use cases, the economics of AI-driven support, why many enterprise AI pilots fail, and how human roles evolve when AI takes on routine work. A grounded look at where customer experience, enterprise AI, and the future of support are heading. Stay Updated: Craig Smith on X: https://x.com/craigssEye on A. I. on X: https://x.com/EyeOn_AI (00:00) Jarrod Johnson and the Evolution of TaskUs (03:58) Why AI Became Core to Customer Service (06:07) Humans, AI, and the New Support Model (07:16) What Agentic AI Actually Is (11:38) TaskUs as an AI Systems Integrator (14:59) How Agentic AI Resolves Customer Issues (19:52) Workforce Impact and the Human Role (23:26) Why Most Enterprise AI Pilots Fail (30:32) Real Client Case Study: Healthcare Impact (36:34) Why Customer Service Still Feels Broken (38:49) The End of IVR Menus and Legacy Systems (42:25) AI Safety, Compliance, and Governance (49:38) Training Humans for AI and RLHF Work (54:34) The Future of Agentic AI in Enterprise

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  After all the work in Fortune 500, it was really, really fun. So I've been here since 2016. I run the go-to-market organization. For those people who don't know Taskus, we're a digital services company supporting some of the most innovative businesses in the world. So what does digital services mean? A little over 50% of our work is in digital customer experience. So think about chat, email, in-app support for food delivery, chatting with your airline app to try to get your seat upgraded, all kinds of things like that. And then we do voice support, but it's 100% cloud-based.

• 0:34Craig S. Smith?

  Can you introduce yourself to listeners and give your background so far as it's relevant? And then I'll start asking some questions about Taskus.

• 0:44Speaker 00

  Yeah, for sure. And thanks, Craig, so much for having me. I'm excited to be here. Jared Johnson, Chief Customer Officer of Taskus. In terms of background, I spent my first 15-plus years in the technology sector. I was with IBM for 10 years as an application developer and consultant. And then I moved into account management and large services, contract management for Fortune 500 companies. So a lot of technology services in my background there, which was a great foundation for me in the early 2000s. Then I went to another IT outsourcing company. We sold that business to Xerox Services. And after that, I just said, I don't like these big companies. They're too slow. I need to do something a little bit more exciting and fast. And all my buddies were working in these private equity companies.

• 1:28Speaker 00

  And so I went to a SaaS business that was private equity-backed, which was super fun. Spent a couple of years there, kind of learning how to run go-to-market organizations in smaller and nimbler companies and building them from scratch. Then in 2016, I stumbled on Taskus. We were not very big. We were 4,000 employees globally, which for a services firm was not that big. The largest players were 300, 400,000 employees. But I really liked the services. It was a heavy mix of digital services. I'll talk more about that in a second. And then it was the coolest customer list I had ever seen. It was all these Silicon Valley tech startups. They had a contract with one of the biggest tech firms. Doing like forty million dollars a year and this client was just on a, you know,

• 2:12Speaker 00

  rocket ship. And I was like, I don't know how these kids figured out how to land these kind of clients, but I can learn something from them. After all the work in Fortune 500, it was really, really fun. So I've been here since 2016. I run the go to market organization. For those people who don't know Task Us, we're a digital services company supporting some of the most innovative businesses in the world. So what does digital services mean? A little over 50% of our work is in digital customer experience. So think about chat, email, in-app support for food delivery, chatting with your airline app to try to get your seat upgraded, all kinds of things like that. And then we do voice support, but it's 100% cloud-based. So there are these critical moments in financial services,

• 2:55Speaker 00

  fintech businesses and such, healthcare and health tech businesses where you really want to get on the phone with someone. We do that support as well. So that's the bulk of our business. We've got a trust and safety practice, which we like to think of ourselves as the police force of the social web. We protect users. We take down inappropriate advertiser and user content. We do dating apps, social media sites, streaming businesses. And we also have a financial crime and compliance business that's in there too for our fintech clients where we do KYC work, know your customer, any money laundering, things like that. That's about 25% of our business. And then the remainder is our AI services practice. That's everything from managing autonomous vehicles in the field, data tagging, annotation,

• 3:40Speaker 00

  building AI models, and emerging thing we call AI safety, which is helping make sure that those models are actually safe when they get deployed, deployed out to end consumers. It's a super fun, exciting business. Historically, one of the fastest growing in the services sector in my time here. So glad to talk more about it.

• 3:58Craig S. Smith?

  Yeah, I mean, the safety thing fascinates me. How, how do you, well, let's, before we get into that, why has AI become central to, to your vision of next generation consumer experience? And, and I'll say that I've talked to a lot of customer service companies that are implementing, you know, AI-generated voice and text and, and those sorts, sorts of things. It seems like a fairly crowded sector right now, but, but tell us about how AI is central to Taskus.

• 4:34Speaker 00

  Yeah, it's a great question. In the, in the customer experience space, it's all anybody's talking about. Like you said, it's a very fragmented marketplace right now. But everybody's trying to figure out, is there a better way to serve my customer and transform what I do through the use of AI and AI technologies, agentic AI, et cetera. Many of those clients are doing chatbot work today, but it's relatively primitive. And from a Taskus perspective, if that's what the market wants, and that's where the market's going, we're going to try to go get there. We're still not the largest player in the space. We have an opportunity to, if we can lead in helping customers get these kinds of solutions right, we think we can continue to grow and take share in the marketplace from kind of the bigger, slower competitors. There's another side of this though, Craig, that is like

• 5:20Speaker 00

  an existential threat, right? So if 50% of my business is humans answering customer contacts, I need to be relevant in the future. That may mean less humans and more humans in the loop combined with AI technology. We think if we can transform our business to be more that way, give clients what they really want and help them achieve their outcomes, we'll win in the long term.

• 5:43Craig S. Smith?

  Yeah, and, and so the customer are, are enterprises, not BPOs, right? I mean, you're not plugging this in at the, at the call center. You're giving it to customers and then if they have a call center, they, they can, it can integrate with this, with Taskus's technology. Is that right?

• 6:07Speaker 00

  Yeah, we're using third-party technologies, but the concept you just said is correct. So the clients will select and implement that, select and choose to implement the technologies. We'll advise them, we'll help with the architecture, we'll talk about the workflows. And that'll be the first line of defense to try to triage and solve these end customer issues in the most efficient way possible and in the way that the customer wants to be served. And then the humans in the loop are tasked us behind that first line of defense. So when there's something that's critical, maybe it's fraud-related, maybe it's financially related, maybe it's a critical healthcare issue, or just an issue that the client feels like they want a human to help solve, either due to complexity or the emotional nature

• 6:52Speaker 00

  of it, that the human is still there to handle those exception cases.

• 6:58Craig S. Smith?

  Yeah. And then you did a deal recently with two companies, Regal and Decagon. Am I pronouncing those right?

• 7:07Speaker 00

  Yeah, that's right.

• 7:08Craig S. Smith?

  Can you tell how those partner fits? I'm sorry, can you tell us how those partnerships fit into your AI strategy?

• 7:17Speaker 00

  Yeah, you know, we wanted to have an agentic AI platform to help serve our clients. And so the difference between just, you know, simple chatbots and agentic AI, the agentic AI means this agent can take action for you. The example we use is in the past, we used to build a chatbot to support customer interactions for a neo-bank, a digital card services company. And we could tell them how to lock their card, but we couldn't actually lock their card with the technology. The technology didn't have the ability to go take an action. Now, we can implement that same solution and we can, through the chat interface, can say, would you like me to lock your card for you? And they can say yes, and we can just take that action. A human used to have to do that. So that's the big shift.

• 7:59Speaker 00

  These two players, we found, were the leaders in their spaces in

• 8:05Speaker 00

  Agentic AI for customer service. We looked at about 25 different players. It's a very crowded marketplace right now and nobody has a real dominant position. We looked at technical capability, willingness to partner, cultural fit, to be quite honest. Like, we wanted somebody who's gonna move as fast as we did, work with the kind of customers that we work with. And these two stood out. They're different. Regal is a legacy voice technology provider. They built a voice-first AI platform. So it's voice to voice on a voice-based model, which is unique. They were pretty early in the lifecycle of voice-based modeling. Decagon is more of an omni-channel platform, so you can do voice, you can do chat, email, different work types through that platform.

• 8:50Speaker 00

  And it's a full platform with quality management, analytics, recommendation engines, some other things. But it is, in the case of voice, it's voice to text, to text to voice. That's how the technology works, because it's a text-based LLM that's serving up the answers. We felt like we needed both. And they're both, you know, they've both been fantastic partners, a lot of good client traction.

• 9:11Craig S. Smith?

  And, but with, just, again, so I understand, Tescus is kind of like a systems integrator. You're not building these solutions. You're putting together a system with these solutions in them. Is that right?

• 9:29Speaker 00

  Yeah, that's exactly right. It feels a lot to me like the, you know, how we used to build the first generations of websites at IBM, or, you know, we used to do ERP or SAP or Oracle implementation. Our job is to be the integrator who has expertise and knowledge on how best to implement the platform. We're gonna configure it to meet the needs of the unique business case of the client. And then because of our history in the human operations of this work, we have the ability to write operating instructions and train the tech to do what we've done so well in training humans to do over time. Deliver high customer satisfaction, quality results, while delivering efficiency in the interaction for the consumer as well.

• 10:11Craig S. Smith?

  Yeah, and and so you you don't have, for example, a

• 10:17Craig S. Smith?

  customer service workforce yourself. It's it's you're you're you're helping enterprises find workforce and and find the tech to work with the workforce and all of that. Or or generally, does the enterprise already have all of those pieces? And you're just helping upgrade it with agentic possibilities?

• 10:44Speaker 00

  Yeah, so our agentic AI consulting practice does the the the implementation, business process management, outcome definition and configuration of the platform to meet the needs. But luckily, because we're a legacy, you know, digital services and outsourcing provider, I have the humans also behind that, so that when the technology either should not or does not solve the problem, we then kick that to a live agent who's in one of these taskless operations centers around the world to make sure that the end customer is taken care of. So the agentic AI consulting practice is the tip of the spear to help get clients where they need to be. And then we bring the human talent behind it when it's required.

• 11:24Speaker 00

  And we envision a world where sixty five, seventy percent of interactions may be handled by technology. And then we would just focus on the complex, critical interactions that customers need to talk to a live human.

• 11:38Craig S. Smith?

  Yeah, I see. And yeah, so you have both ends. You have the the tech, the agentic end, and you have the human end. And as part of your training of the human workforce, how to use these tools or how to interact with them, because that's one of the big issues with company kind of bolting on, you know, voice on the front of a system that's relied heavily on human call centers that the handoff can be very, very unsatisfactory.

• 12:13Speaker 00

  Yeah, that's right. I mean, it's certainly the the handoff from the technology to the human is still a challenge. Nobody likes to be transferred, right? Oh, I'm sorry. Let me transfer you to the to the airfare desk, right? Nobody likes that. But we, because we've handled the human interactions from the start to finish for so many years for so many of our clients, we have almost a better understanding of the process flows than our clients do. And we also stand understand some of the shortcomings in How they will need to evolve what they do in order to make the technology work so we can help fill those gaps.

• 12:45Craig S. Smith?

  Yeah, and well, in the past, for example, closing a card, that would be done by the call center or they would send a message to someone else in operations to do that. And now that you have that agentic capability, it happens right away at the front end.

• 13:05Speaker 00

  Yeah, that's right. So I would say, you know, most clients turn over to task us or other outsourcing partners, ninety percent of the contact volume. There might be ten percent of things that are like escalated refund or, you know, some dollar amount that they want to handle in-house, but ninety percent of the work would come to a, you know, a human capital talent partner like us. Now, you know, we're seeing, we're hoping that we can take fifty, sixty, seventy percent of those front-end interactions and then pick up the bulk of the remainder and execute on those for clients.

• 13:38Craig S. Smith?

  Yeah, but in the past, something like, you know, did you say, what's the verb, closing the card? Or, you know, if you have a credit card that's been lost or stolen. In the past, your call center representatives would do that? Or does it involve more than that?

• 13:59Speaker 00

  No, so in, so in the past, in kind of the legacy world, only a, only a human agent who had logged into a system had the access rights in order to turn off a card for, for a customer, right? A customer may or may not be able to do that. Some of the self-service apps can do it. You know, you can do some things yourself, but if a customer called in, a human would have to go take the action to log into the system, validate the user and say, yep, I'm gonna, I'm gonna close this card down. And they might even try to save them, offer them a lower interest rate, you know, do, do something. Now, we can train the technology. The old chatbots could not take that action. They could send you to a help article, or they could try to walk you through the steps to do it yourself if it was available on self-service.

• 14:43Speaker 00

  But normally, the chatbot would have to get you to a human agent to make that action. Now, with agentic AI, you empower the chatbot to actually go into that system, authenticate and take the action on behalf of the customer without a human having to be in the loop.

• 14:59Craig S. Smith?

  Yeah, so from your perspective, how do you define a Gentek AI? Because there's a lot of automation out there that isn't truly a Gentek. And why is it so pivotal in the way companies like yourselves are deploying technology?

• 15:18Speaker 00

  Yeah, the primary difference in a Gentek AI is the ability for the virtual agent to go take an action or have agency, right? That's where that kind of term comes from. These systems on the back ends that our our human teammates use are very complex, right? You might have might have to log into a CRM system to validate somebody. Then you might have to toggle over and get into a financial system in order to lock and close a card. And we have to train teammates for two, three weeks before they can get on the floor and be proficient. Now with a Gentek AI, I can train technology to overcome these previous systems limitations and actually take action on behalf of the end user or the customer or the member

• 16:01Speaker 00

  to deliver the result they need without so much human intervention and frankly, error. You obviously have to train the technology to take the right actions. And there will, there will be things you have to tune over time. But that's the primary difference from, you know, the old websites you go to and like you're going to a help center and you're looking up a knowledge article to do something yourself. Or the chat bot who might feed you information but can't take an action on your behalf. So the the big leap is the ability to actually create the integration. The Decagon or Regal system can actually go talk to the CRM and validate. You can actually go talk to the financial system and lock your card and then tell you, hey, you're you're taken care of. So it's a big leap in functionality. And I imagine this sort of, this will continue to evolve and they'll be able to take even more actions on behalf of consumers.

• 16:45Craig S. Smith?

  Yeah, and how does Regal and and Decagon work together? It doesn't depend on, yeah, tell me how that works.

• 16:56Speaker 00

  Yeah, so they, in theory, they could. We have clients who are using multiple agentic AI platforms for different use cases, where sometimes they want to test both and see which one is performing better than the other. But in, in most of our cases, our clients are, and maybe it's because it's early days, but our clients are picking one that best meets their needs. And again, in the case of Regal, if you, if you have a use case, like a voice interaction that is costing you a lot of money that you think can be handled by technology, you know, they're, they're a great fit. Decagon might be a great fit for multi-channel interactions where you're looking for a different set of solutions. So we, we generally don't present both at the same time. We try to recommend one that we think works best for the client.

• 17:36Craig S. Smith?

  Right, but you do recommend one or the other. I mean, as, as opposed to just making that decision on your own based on the, the case.

• 17:47Speaker 00

  Yeah, I mean, we normally try to look at each customer use case and fit and make a recommendation for, you know, where are they trying to get to? And then based on our set of partners. And frankly, there's others we're not as close to, but there's others we've evaluated that we understand that maybe, you know, maybe there's even a third option. So we're trying to act as that consultative partner to help get them to success in their, in their project implementation.

• 18:11Craig S. Smith?

  Yeah, if sixty does, did you say sixty, seventy percent of call interactions are resolved

• 18:20Craig S. Smith?

  with technology, with either agentic or, or some voice technology? Is, is that what you were saying?

• 18:29Speaker 00

  Yeah, so the example I was giving is in, in the old chatbot worlds, about forty to fifty percent of interactions are handled by the chatbot and resolved, which is kind of surprising, right? We have one client. You know, maybe a million interactions a month were handled by the chatbot, which is great. And that didn't exist five years ago, really, right? The difference with the Gentic AI is we're able to push that, what they call deflection, or the ability to resolve inside the tech, we're able to push that up to sixty-five, seventy percent. And I think over time, we'll get even better because now we can take that action for the for the end customer that that before the chatbots just couldn't do. So seventy percent is kind of our benchmark. That's where we think we can get clients

• 19:13Speaker 00

  to if they give us all the volume they have coming in. We think we can get them not only to the old forty fifty number, but get them up to seventy. And then that incremental, you know, we think is twenty twenty-five percent incremental resolution with the technology, which is significantly cheaper cost per interaction than the humans are, even if you're doing it in, you know, different parts of the world, low cost parts of the world.

• 19:36Craig S. Smith?

  Yeah, absolutely. So is your workforce then shrinking, your human workforce? Or as you grow, you you just distribute the twenty-five percent that require human interaction across that workforce?

• 19:52Speaker 00

  Yeah, that's a phenomenal.

• 19:53Craig S. Smith?

  Yeah.

• 19:54Speaker 00

  Because this is the this is the big existential threat part. Like, are we just accelerating our demise? You know, am I the MP3 player? And I just created the iPhone or something like, you know, I don't know. But what what we've seen so far with clients is they're able to drive that ten to twenty percent incremental efficiency on the simple contacts, the the high volume, simple contacts, you you want to allow a technology solution or a bot to handle. But then what they're doing is they're taking that savings, and maybe they pocket some of it. But they're coming to us and they're saying, hey, you know what we really want to do now? We want to put your best agents into a tier two group that is just going to handle my most valuable members, my most valuable customers. So we're getting these premium queues or fraud queues, or maybe they deploy more people

• 20:37Speaker 00

  to saving customers who are cancelling or retain, you know, retention activities. We're seeing them redeploy the bulk of that cash. And so we're still growing in the most case. But I will say that the agent profile is getting more complex. So because the work that we're getting that's replacing the simple stuff is is getting more complex.

• 20:58Craig S. Smith?

  Yeah, I mean, I'm at a conference right now where there's a lot of research conference. So they're talking a lot about advances, advances and agentic capabilities. Where do you see it going in your industry?

• 21:16Speaker 00

  Yeah, I mean, I don't know that I have a crystal ball on this one. But when I, if I look at the most advanced users of the technology that we've deployed and that we've kind of studied in the market. I think that agentic AI will take these simple interactions that are already probably enabled for self-service, but the users don't know how to do it. I think it'll almost wipe out that simple contact volume. We think that represents maybe 20 to 30% of our interactions that we get today over the next three years. So I think that that, what I'll call level one work, that simpler work, will go away pretty fast. But I think on the horizon, what's really interesting is the translation technologies are becoming more advanced on top of these agentic solutions.

• 22:01Speaker 00

  So things you couldn't automate before because they were in, they were in Japanese and that, you know, the technology didn't, couldn't handle it. All of a sudden, this advanced translation work is going to mean that you can solve even more work globally than you could with your old chatbot solutions. And even 18 months ago with the LLM-based AI solutions, the translation wasn't quite there. Now it's there. The other big advancement has been voice. So I anticipate that translation and voice on top of agentic AI will continue to drive more and more adoption. And then I, I imagine that today we're taking sort of one-step actions for clients on simple things.

• 22:43Speaker 00

  I imagine you'll get to a day where the workflows will be, will allow the agentic AI technology to take multiple steps and actions for you and understand the context of your entire trip or your entire customer journey or experience that you're going through. And then actually be able to start making recommendations and interact with you. I can see a day where they'll have contextual history on you. So when you call in one time or chat in one time, when you come back, they know you're booked on the next flight to Chicago, they know where your seat assignment is, and they start to predict, you know, what it is you may be needing based on your historical interactions. So I think there's a lot of potential for this to continue to evolve, at least over the next, you know, five to 10 years.

• 23:26Craig S. Smith?

  Yeah, and, and you know, I was gonna ask about getting from pilot to production because so many enterprises are experimenting, but so many of those pilots are failing. And you guys must have gone through that, that arc. What, what challenges do companies face when moving into deployments of the Gentech AI? What challenges did you guys face? That's one question. And then are you, is task us moving toward? When you were talking about more complex agentic workflows, are, are you moving towards

• 24:11Craig S. Smith?

  maybe beyond customer service to helping companies implement multi-agent systems that, that do more than, than respond to customers?

• 24:24Speaker 00

  Yeah, so on your, on your last question, the, we absolutely see the, a day where there will be multi-agent activity, handling customer interactions. So a lot of the model developers today and the, and the Decagon and Regal examples, they're trying to say laser focused on, I am a customer service chatbot or AI solution. Because they know that market's plenty big, they wanna stay hyper focused, don't wanna spread their resources. But then we talked about other complex interactions like fraud and financial crimes. Someone's gonna go build an AI solution for just specifically that problem that will know how to handle your disputing a charge on your credit card. I need to investigate that dispute and understand from the merchant, were they really supposed to charge you or not? Somebody's gonna build that, right?

• 25:08Speaker 00

  So you can see a day where you've got multi-agent orchestration, handling the complexities of what really has to happen to solve and to solve a customer's problem and get to end of job. So we can, we see us going along that journey with our clients. And we can see it when we've our longest standing implementations, it starts with simple, starts with the baseline, goes to, goes to the next, then you tackle a new problem. In terms of challenges, I'd put them into three buckets that, that we've, that are pretty consistent. Number one.

• 25:43Speaker 00

  Everything you're training an AI solution to do is based on the information that you feed it. So these knowledge articles, which are your operating procedures for this is how you this is how you open an account, verify that it's you, check the account balance, like all those instructions have to be accurate. And most clients do not have enough up to date knowledge management systems and policies in order for an AI solution to ingest them and then execute correctly. So we find one of the first things we have to do is get the scope tightened, work on those knowledge articles and policies and procedures, document them correctly, and then come up with a way to consistently update and manage those over time. So

• 26:25Speaker 00

  challenge number one, the base knowledge management's got to be in place. Challenge number two, our clients are some of the most sophisticated technology companies in the world, and they're most of the apps on your phone, and they're beautiful. When you, when you log into them, the user experience, the how you select an item and do my tip and auto refund, all these things, it's beautiful facing the customer. They don't exactly put the same kind of engineering talent on the back office systems that a that a Taskus teammate might use to go execute work every day. I'm never more fascinated than when I watch a Taskus teammate and I sit by side by side in the centers and they're going between tabs and multiple systems and copy paste here, this, in order to make the, make the process work.

• 27:10Speaker 00

  So they generally have limitations in what they can ask an AI agent to do because they have their own system limitations on the backend, which maybe humans can work around, but a technology solution is a little hard. And sometimes they say, you know what, I'm not going to solve this right now with Agentic AI because I'm, I need to go fix my backend system first. Once I fix that, then I'll deploy the AI on it. So that's, that's bucket number two is backend system sophistication, I'll call it. And then the third is just an industry reality, which is the sophistication of the platforms and the partner technologies are still evolving. So Two years ago, when we started deploying agentic AI systems, the partners did not have the same level of capability that they had that they have today.

• 27:55Speaker 00

  Right. So they're constantly adding features, functionality that serve that serve the clients, which is great. But you're you're you're going to continue to have this, you know, sophistication curve on the platforms where they're continuing to build out more functionality. Decagon, I think a year ago, Decagon had barely released their voice, their voice modules. They didn't have voice to start, you know, translation now is at a point where it it can be plugged in real time, real time analytics, sentiment monitoring, like a bunch of these, you know, features and functionality have come in over time as the platforms get more mature. So that's the first, that's the third challenge for the clients is like to make sure you pick a partner and you're willing to ride that that sophistication curve. So you don't, you may pick the right partner today, but you need the right partner in two years also.

• 28:41Craig S. Smith?

  Yeah. And and for you guys working with the agentic solutions, I mean, did you go through that, that, you know, having to fill out your knowledge documents and, you know, getting getting prepared for using these agents? Or is that only really something that happens on the customer side?

• 29:09Speaker 00

  It's mostly on the customer's side, but we we did our first implementation with Decagon on ourselves. We had about 150 people who did HR service desk work, like our own employees calling in saying, I have a payroll issue, I have a benefits issue or something. And we said, let's, I mean, if we're going to go market this to customers, clients, let's go learn and use it on ourselves. And boy, did we learn, you know, the harsh realities of, you know, what it takes to make these systems really work well. So that's, that's one thing. And then I'll be quite frank. Thank goodness, we we had such strong partners in Decagon and Regal. They really held our hands on the first few so that we could learn, because these these technologies are so new.

• 29:51Speaker 00

  It's not like you can go out and find a certified, you know, Decagon consultant. We basically have forward deployed engineers who are custom engineering solutions for clients in real time on systems that are being updated and upgraded kind of every day, it feels like. So there's been a lot of handholding along the way.

• 30:13Craig S. Smith?

  Yeah, and the, can you give an example of a customer, you don't have to name the customer, but we're a deployment with either Regal or Decagon, together with Taskus, move the needle? And I mean, how do you measure success in this?

• 30:32Speaker 00

  Yeah, that's a great question. The one I'm most proud about right now is a, it's a health technology platform. So it's an app that connects customers to counselors or psychologists, mental health professionals. And you can imagine, this is a big matching problem for the, for the platform. They're trying to get the right person with the right counselor at the right time. And setting the appointment is relatively simple through the technology to set the initial appointment. The problem is the percentage of reschedules, cancellations, and other things is a huge revenue loss for them. So let's say they had an 80% conversion rate before for appointments set to something, an appointment that actually

• 31:13Speaker 00

  got held. Well, they were deploying, because of HIPAA regulations, they were deploying US-based call center workers to do outbound calls, to reschedule that appointment, confirm the cancellation, find a new time that works. And we were able to deploy the Regal solution in that case, to do these relatively simple calls. This is just a reschedule. And if the, if the patient actually just went back to the platform, it can be done very, very simply. But they have to get back to the platform. So we were able to deploy that solution to help significantly reduce costs. They can actually do more outbound phone calls at the same quality and accuracy as the human-based solution. And our goal is

• 31:58Speaker 00

  to actually exceed the human-based KPIs and metrics eventually. We're about equal right now. And then at the same time, it's a huge revenue lift. If you, if we can lift them from eighty percent set to held to ninety percent, I guess you could say conversion. That's a significant revenue and profit uplift for that client. So we're in what I would call phase one of that deployment. So it's not, not global, not every use type, use case. But it's, it's, it's significantly enhancing the client's ability to attract and retain those patients.

• 32:31Craig S. Smith?

  Yeah, you don't have any metrics on that yet, do you?

• 32:34Speaker 00

  I don't have, I don't have any specific ones that I can measure. I know, I know the previous conversion was eighty percent. I know the goal is ninety percent. We're fluctuating between eighty-two and eighty-five today.

• 32:45Craig S. Smith?

  Yeah, okay. And, and what, what's the feedback you're getting from your human agents? Do they like working alongside Agentic AI?

• 32:57Speaker 00

  You know, there's, I don't think we have enough scale yet. You know, we've got over sixty thousand employees globally. You know, we're only in a dozen, half dozen to a dozen pilots and production programs out today. But I will tell you the sentiment of the human frontline. So one, there's fear, right? They're like, is my job going to be wiped out by AI? Maybe I shouldn't embrace this, right? But there's another population that sees the opportunity for them to not be in a simple job that pays, let's call it, you know, like, for example, let's say, pays twenty dollars an hour in the US today.

• 33:35Speaker 00

  Well, if I can move up into more clinical work, more medical coding work, more complex interactions, more fraud and dispute work. Well, maybe that's a twenty-five to thirty dollar an hour job. So there's an opportunity that if, if, if they can, if the AI can take away the simple work and allow more of the complex work to come their way, there's an opportunity for career growth. The other thing we're trying to do with our entire employee population. Is encourage them to become AI literate. So whether it's internal tools that we deploy to them to help make them do their job far faster, smarter, more efficiently, that's better for clients. That's better for the end customer, right? So we're asking them to embrace those tools, adopt those tools.

• 34:19Speaker 00

  Because we believe, and this is not just a call center thing. I think this is a, you know, employee thing around the world in multiple industries. In order to be effective in the future, you're going to have to be AI literate and know how to integrate AI tools into how you do your work to be as effective as the next person coming out of college who's already native, AI native, right? Been been working on this stuff for, you know, four, six, eight years or something. So we think this is a, we're trying to educate the population to say this is an evolutionary trend, just like, just like using the internet to look up information to be more efficient in the past, or or using, you know, a new system that a client deploys to do the job better. This is just the next iteration of that technology.

• 34:59Craig S. Smith?

  Yeah, and and hopefully they will be able to move up into more complex tasks. And and as the market grows, it doesn't mean shrinking that workforce. It just means moving them into higher level work. Is that is that the idea?

• 35:18Speaker 00

  Yeah, that's right. The, you know, I've been I've been in the technology industry now for over 20 years. And I can't tell you the number of times we saw, you know, four or five different technology things come in and we say, oh, we're not going to need humans anymore. And the reality is there's more jobs in the in in our industry and in the technology industry than there has ever been in the past. So it's just a matter of evolving and adapting, being curious, learning and moving to the next. And our responsibility as an employer is to create the training pathways so that our teammates today can make that leap and take that journey with us to where the clients will want us to be.

• 35:58Craig S. Smith?

  Yeah, yeah. And so you were saying you have half a dozen either pilots or deployments. It's, you know, I've been talking to a lot of companies in this space. Yet every time I end up, you know, calling the customer service line or or trying to interact on a chatbot or even by email, it's a disaster. I mean, I just, I don't understand why it's taking so long for adoption to spread. Do you have any insights into that?

• 36:34Speaker 00

  Well, you know, I would say this and not insult you directly, but what I, it's amazing how uneducated or unfamiliar the customers are with how best to manage their service from our clients.

• 36:55Speaker 00

  And it's, it's the same with humans and it's the same with technology. The difference is human empathy and understanding can guide you. So when you're saying, when you're saying something, you want to lock your card, but you don't know how to say the words lock my card. You're saying words like, I just want no one to be able to use my card. Or, you know, you're saying something else. A human just naturally knows how to direct you down that workflow path. And then they can do the translation to, oh, that means lock card. The technology is struggling with some of those semantics. You have to train it, like you have to feed it. And we've, we feed the technology solutions, you know, cases after cases after cases to give them the transcripts to say, okay, this was the right outcome. This is, this is how I need to handle this. And you give them the operating procedures. But boy, the, the uneducated user is a challenge.

• 37:39Speaker 00

  It's, and I think it's always going to be a challenge. The other, a frustrated and emotional user is always going to be a challenge. Teaching humans to have the right empathy, teaching AI agents to have the right empathy. These are going to be, you know, ongoing challenges and they're going to slow adoption. That's why we all yell agent, agent, agent, right? Sometimes. So it's, you know, it'll be interesting to see how it evolves. But I think, I think consumers will get more and more and more used to it. And, and used, used to how to get getting things done, especially with those companies they interact with a lot. So me and my airline, when I'm on the road, you know, 50 weeks a year, or your, your, your bank, when I'm trying to get my kids their allowance, you start to get, they start to train you as a consumer, right?

• 38:20Craig S. Smith?

  Yeah, yeah, although I'm not talking about the the consumers inability to adopt. I'm talking, there's still phone menus and, you know, press one for this, press three for that. I there's still those systems out there. How long do you think before the kinds of solutions that you're providing really become the norm?

• 38:49Speaker 00

  I don't think it'll be very long before the what we would call the traditional IVR, right? That press one for two, press one for help, press two for reservations, but whatever. I think that that is going to be replaced relatively quickly. And what what you're going to want to do is put an AI agent or an AI solution at the front end to triage all that based on natural language. This the L L Ms can absolutely do, right? You have to you have to train it to understand what the customer is asking for. But it'll be a much more natural interaction rather than you having to wait for your list of eight things. And then you you forget the number on the third one and then you gotta go repeat the list to me. You know, you're going to actually have a natural language interaction like you're talking to your Alexa device, like you're talking to Siri.

• 39:34Speaker 00

  It's just got to be programmed and, you know, that migration has to happen. There's cost implications to it. There are some cases where you may still have a phone tree because they they can't can't put the agent in front of all those call types. They can only put in front of two, but I think that's only a matter of time. I would say in 24 months, we should see pretty significant

• 39:55Craig S. Smith?

  Really?

• 39:56Speaker 00

  information there. Yeah, I think so.

• 39:58Craig S. Smith?

  Yeah, yeah. And then, yeah, from from you you were talking at one point about not only improving customer service, but also opening up new revenue streams. How how do you see this leading to new revenue streams for enterprises?

• 40:19Speaker 00

  I would say it's not a hundred percent clear to me yet. But just the ability to take cost savings and redeploy them to revenue generation initiatives.

• 40:30Craig S. Smith?

  Yeah.

• 40:31Speaker 00

  It's and and revenue retention initiatives seems to be something most of our clients are more than willing to do. Yes, they want the cost savings and maybe they pocket a little bit. But in the end, they have a business model and if they can redeploy those savings into things that have ROI. They're going to do it. I think the easy ones to begin with are going to be new customer acquisition, lead generation, retention and premium support. I think they're gonna, I think they're gonna move into kind of those areas next. And then I suspect that as they have the capability, they may be able to service more products than they've been able to service in the past. So it may accelerate new product deployment, which then would also drive revenue growth potentially.

• 41:13Craig S. Smith?

  Yeah, but the agentic, the agent that's interacting with the customer, couldn't it also upsell or or work on retention or without handing off to a human agent? Or do you think those things really require human touch?

• 41:30Speaker 00

  I think it's gonna be a business decision. You know, the complexity will be one thing. The brand experience may be another, but I absolutely think almost every example, I think every example I just gave you. A gentec AI will move in to take over pieces of the human work. And there will be very strategic business decisions on, no, I want this to be a human interaction. And you can hear, you can hear on the radio or TV, people advertising, you know, live, live human support 24/7 and things. So it's actually a brand differentiator for a number of people. But that doesn't mean you don't want to take advantage of the technology where it does meet the customer need, right?

• 42:08Craig S. Smith?

  Yeah. How do you, I mean, you're dealing with partners who you're relying on their technology. But how do you ensure these systems maintain accuracy and reliability? And compliance with data privacy and security?

• 42:25Speaker 00

  Yeah, this is a major concern that we have. We don't think that clients understand how much care and feeding that these systems will need. We talked a little bit earlier about knowledge management and policy updates. That's, that's just one in-house, you know, thing that they have to manage and control. They should be doing that today, or maybe they are with a human process, but they're going to have to be able to do it with a technical process in the future. And then keeping up with regulatory, security, privacy frameworks is, is going to continue to evolve as well. The good news is the partners are coming with a good baseline. Like, they know the systems have to be PCI compliant for credit card usage.

• 43:09Speaker 00

  They don't want to really be holding any private individual information. So they're, they're going to come in with a good baseline. But what will happen is each industry will have a couple of, and will have unique needs. So banking and financial services will be different than e-commerce or other things. And then you're going to have company-specific things as well that will have to be built on top. One of the things that we do is we'll do that upfront, systems integration work, and get you to go, you know, to live for production. But we can actually stay on, do that knowledge management work. We can help implement the policies, right-to-age and operating procedures to manage new regulations as they come down. That'll be, you know, there's probably a chunk of work upfront that's kind of large, large scope.

• 43:54Speaker 00

  And then we think there's this tail of work that clients most likely will, will learn that it's the care and feeding and maintenance that is not strategic enough for them to necessarily want, want or need to do themselves. They just want to direct it and tell them what the business requirements are, and then get the technology partner and somebody like Taskus to help make sure the system's maintained. Even little things like managing Bias in the LLM models, other offensive content that users might yell into the technologies. Like you have to know how to manage these things. So there and there's always going to be new cultural nuances that you have to adapt to. And the the models underneath will have to adapt to. So there there's this is not a set it and forget it environment.

• 44:38Speaker 00

  It's going to be ever changing, I think.

• 44:40Craig S. Smith?

  And is that something that Taskus provides? Because the customer may not have its eye on that ball all the time.

• 44:49Speaker 00

  Yeah, we do. We actually have an AI safety practice that helps for the largest model development, foundation model development firms, train their models to avoid bias. We we do what we call red teaming, prompt engineering, prompt evaluation, essentially making sure that the model is is executing the way it was intended by the developers and that it has guardrails on behavior. You know, we've heard stories in the news about these chatbot solutions, getting into emotional conversations with teenagers, for example, that have resulted in harm. Well, now you have to go retrain that model to not not engage in such ways.

• 45:31Speaker 00

  And so we're helping the foundation model developers, which will help the the, you know, agentic AI companies, the companies who are building products on top of AI. And then we'll help the end clients who then are using these technologies and have unique needs. So what we're trying to work in our AI safety practice to do that sort of analysis and training at every level of the model development.

• 45:56Craig S. Smith?

  So if if enterprise comes to you for a solution, who re-trains or fine tunes the underlying model? So that the agent has the expertise required to do what it needs to do? Is that something you guys do or something DuckGong would do or?

• 46:22Speaker 00

  Yeah, it's two, it's two, two phases. So the, the underlying model, whether it's Gemini or OpenAI or Anthropic or something, these, these product companies have one underneath. That's a

• 46:32Craig S. Smith?

  Yep.

• 46:33Speaker 00

  standardized model. So that's already been trained on basics. But someone has to go train it on, you know, client number one's knowledge. And so we, along with the partners from Decagon and Regal, are feeding that knowledge in and training the model in the specific instance that is developed for that client because it's closed, right? It's not that none of the data that we send in there goes outside that, that instance's walls for privacy and data breach reasons. So, but it has to be trained on the knowledge articles and the operating procedures. So those have to be written and we're, we're doing that along with our technology partners who are doing most of the detailed engineering and, and maybe the more complicated integrations to

• 47:18Speaker 00

  CRM systems, backend systems. But the business process, logic, policies, knowledge, that's where we really have expertise because we've been doing this to train teammates for, you know, 15 years.

• 47:30Craig S. Smith?

  Yeah. How, how large is your human agent workforce? Or, or does it grow and contract depending on business?

• 47:41Speaker 00

  We've been mostly, you know, pretty consistent growth mode in terms of total number of employees. We just crossed 62,000, I believe it was, we, we released in our Q3 earnings. And that's in, last I checked, like 31 sites across 13 countries, I believe.

• 47:58Craig S. Smith?

  And, and that's, are they doing, are they acting as customer service agents or are is part of that workforce also helping train or fine tune foundation models, you know, when there's safety and, and the safety aspect of what you guys are doing?

• 48:18Speaker 00

  Yeah. So 50% of those teammates that I just mentioned, so call it 30,000, 60,000, are doing customer experience related work. Maybe some of it's sales and lead gen, but it's, it's basically, it's what you would consider customer care-ish.

• 48:30Craig S. Smith?

  Right.

• 48:31Speaker 00

  But that business is only growing, let's call it seven to 11%. Trust and the teammates on the front lines doing trust and safety work, you know, there's probably, let's call it a, a

• 48:45Speaker 00

  Fifteen thousand teammates doing that work. That is growing at twenty plus percent. So more and more of that work coming up. And the remainder are AI services practice. That work is growing at fifty plus percent. Because there's just so much demand for implementation and management of AI systems. So there's, you know, a bunch of different dimensions of that. But it's so, you know, yes, we still do a lot of the customer support work, but the other pieces of our business are growing much faster.

• 49:14Craig S. Smith?

  That's interesting. And and then how how much training is required in, for example, in in the of the aid of the human agents that are working on fine tuning or reinforcement learning with human feedback or or those things? Or are the tasks fairly straightforward? I've always wondered that.

• 49:38Speaker 00

  Yeah, it's a good question. So the the typical

• 49:42Speaker 00

  customer client of Taskus, a teammate agent doing the work, it's a it's a typical two to four week training. And you might have a mixture of classroom training and then you're doing some what we call nesting, which is you're taking some production cases, but you are getting a lot of off time for coaching and understanding what you did right and what you did wrong. The most complex work we do can be up to a twelve week training. So something that has to know how to work with airline reservation systems is a much longer training cycle. But on average, let's call it three to four weeks, something like that. Interestingly, the RLHF, red teaming, prompt gen, those AI technology things, it's a shorter training duration,

• 50:26Speaker 00

  but it's a more complex skill set. So they have to have a higher language, English language proficiency, and they generally have to have a little bit more technical acumen and critical reasoning skill. Those projects on our LHF are often mini sprints. So a client will call and say, hey, I need 50 people to do this. And I need the outcome by three weeks from now. And we sprint it. So the training durations are shorter upfront. The work comes in these project sprints sometimes. So it's just very different, you know, nature of some of the different parts of the work we do.

• 51:05Craig S. Smith?

  Yeah. And the customer service is kind of the proven ground for agentic AI. Do you think that's because it's it's such low hanging fruit? The requirements are not that complex for the EAFI agents. So it's just natural that they would be the leaders in implementation? Or do you think there's there's something else going on there? That maybe it's the variety of of problems that need to be solved? Sort of test the agentic capabilities or something?

• 51:47Speaker 00

  Sorry, give me one second.

• 51:52Speaker 00

  So the the reality is the venture capital firms all know that the customer support business is just in North America is probably $100 billion in services companies like me. And another $50 billion in in-house call centers, let's say. So it's a massive market. I think that's why so many of these AI companies have targeted the space to go to go develop solutions. It's a it's a it's a total addressable market equation. But if you look in our trust and safety business, and we do content moderation for a lot of the big social media companies, there's a lot of automation that's been happening in there for years.

• 52:32Speaker 00

  Machine learning systems detecting nudity or offensive language and automatically taking it down. That that's going to pick up with the use of AI as well. So there's right now, the one everybody's talking about is customer support. I think because of its size and it's going to impact each of us as consumers. But the it's happening in, you know, many of, I believe it will happen more in all the service areas that Taskus touches and probably many others.

• 53:04Craig S. Smith?

  Yeah, and and are there lessons from from your work with Regal and Decagon? That apply to other industries, healthcare? Or retail? For their own AI strategies? I mean, beyond customer service.

• 53:24Speaker 00

  Yeah, I mean, I think I think the best advice that I can give to clients is that they should be doing something or they will be left behind by their competition, I suspect. They need to they need to get out and be testing and piloting. And they need to think about not just getting to solve some edge use case. But think about one real problem that has some ROI. And get partners like Taskus, Decagon, Regal in the room to stare at that problem. It's got to be a big enough problem that it's financially motivating for your partners. And that there's real ROI for your business. So that's that's my advice, you know, in a couple of sentences is get in and get started.

• 54:06Speaker 00

  And try to find one thing that matters to your business and get see if you can have some success. And then, once you have initial success, bolting on additional workflows, additional problem sets becomes much easier because the framework and foundation is set.

• 54:24Craig S. Smith?

  Yeah, yeah, and where do you see the partnerships with Regal and Decagon? In three to five years?

• 54:34Speaker 00

  That's an excellent question. I hadn't really thought about that. I mean, you know, we we bet on these two because we think they're going to win. And knowing how technology companies that are winning work, I anticipate that their platforms will get broader and broader and broader. And I could see a day where CRM systems as we know them today may end up being replaced with something that's more AI-based and an easier architecture to get work done for customers and clients and have a more modern set of data privacy, user privacy standards behind them. That is the advantage of being the late mover or the second, third generation of technologies.

• 55:18Speaker 00

  So, right now, Decagon and Real are sitting on top of a lot of underlying technologies that are basically, I don't want to call them legacy because they're all SaaS-based, cloud-based, you know, they're modern-ish, but they're not AI. I would anticipate that you'll see, you know, the the foundation LLM models, the OpenAI's and Anthropics move up to provide more functionality and you'll see the AI products move down and squeeze the current technology providers out, which is why I think all the the Salesforce's, the Zendesk's, the these kind of players, ServiceNow's, ServiceNow just today announced a billion-dollar acquisition of an AI agent company. Salesforce has announced AgentForce.

• 56:02Speaker 00

  These are, these are huge initiatives because I, because they see the license revenue risk and they see the opportunity. So, that middle ground is going to be super interesting to watch.

• 56:13Craig S. Smith?

  Yeah, yeah, it's exciting times.

• 56:16Speaker 00

  Oh, yeah.

• 56:18Craig S. Smith?

  Yeah.

• 56:18Speaker 00

  I'll add one more thing and you kind of touched on it, but I just want to reinforce it. I also think you'll start to see evolving industry-specific solutions from the providers. So, right now, it's like, I'm a customer, customer service agentic AI platform. I think you'll see a day where it's like, I, this is how we do it for pharmaceuticals. This is how we do it for financial services. And you'll, you'll see very tailored, you know, business solutions for specific industries, I suspect.

• 56:47Craig S. Smith?

  Yeah, companies like Taskus but that are in effect integrators but for other specific use cases.

• 56:55Speaker 00

  Yeah.

• 56:56Craig S. Smith?

  Is that what you mean? Yeah, yeah, yeah, yeah. Okay, well, this is really interesting. Is there anything I I didn't talk about that you'd want me to talk about? I'll put David up here.

• 57:08Speaker 00

  It obviously depends on where is the work being done by a human today? If that work's being done by a human in the US for $22 an hour, the potential cost savings with an Agentic gas solution is significant. You could be talking 70%. On average, when we're talking to customers, I'm willing to sign up for almost a guarantee of 20% to 30% savings.

• 57:31Craig S. Smith?

  Wow.

• 57:32Speaker 00

  Even in cases where they've already taken advantage of labor arbitrage, much of the work has already been offshored for cost savings, we still think by optimizing the processes and using the technology, we can deliver 20-plus percent savings.

• 57:46Craig S. Smith?

  Wow, that's remarkable. Yeah, okay, great.