Farewell, Chatbots: AI Agents Are Taking Over Customer Service | Mike Murchison, CEO, Ada

Why is customer service a perfect use case for AI?

Matt Turck [1:47] Hey, Mike, welcome.

Mike Murchison [1:49] Thanks, Matt. Great to talk to you.

Matt Turck [2:17] Thanks for doing this. So I'm excited to talk about AI for customer service because it's one of the historical spaces for applied AI. And in a world where there's a tremendous amount of demos on Twitter/X of AI agents and all the things that look cool but may or may not be deployed in production at scale, customer service is actually one of those rare areas where there's a history of deploying AI, and the space looks super vibrant. If nothing else, there seem to be more companies getting created, entering the space, in addition to companies that have been around for a bit.

Matt Turck [2:35] So maybe to start from the top, why do you think that is? Why does customer service seem to be such a ripe area for AI?

Mike Murchison [3:03] Well, I think it's simply because the quality of customer service that's delivered around the world today continues to be abysmally poor. Like, the average American will spend 40 days of their life waiting on hold. We all have a very visceral experience, probably in the last seven days, of a negative customer experience we've had with a business. And businesses expend an incredible amount of money delivering what is fundamentally a poor customer service experience. Globally, every year, $500 billion is invested in human-led customer service operations around the world.

Mike Murchison [3:29] And so it's a bad experience that's incredibly capital inefficient. And finally, we now have AI that is capable of powering not only as good an experience for less money, but a far better experience for less money. So I think those three factors make it a logical entry point for generative AI in the enterprise.

Why didn’t foundation models replace AI “thin wrappers” out of the box?

Matt Turck [3:56] And one particularly interesting part of the discussion around AI for customer service in the last year or two is that there was this moment in time when it seemed that foundation models were going to completely kill that area. And there was a lot of chatter, again, on Twitter/X about why would you need any kind of AI chatbot company in a world where you can just use Claude or GPT? And then that seemed to be corroborated by Klarna. At some point, there was this very well-publicized moment when they were able to decrease very substantially the number of people that they used in customer service using a foundation model.

Matt Turck [4:18] Sort of fast-forward to today, it seems that that discussion is largely gone. And in fact, again, the space seems more vibrant than ever. What do you make of that?

Mike Murchison [4:35] Yeah, I think definitely 12 months ago, there were a lot of discussions about, do you have proprietary models? Is your business a thin wrapper? How far down the stack do you go? And I think in the last 12 months, most of the market has realized that the value is beginning to accrue at the application layer. And that's because it turns out that it's very expensive and technically quite complex to control your AI, to measure its performance, and to steadily improve it over time.

Mike Murchison [5:09] And I think it's those three factors that are starting to really help folks realize that what matters most is the value that we're providing to businesses, not so much how we provide it. It's turning out that there's a lot of product to be built, a lot of software to be built to deliver that level of control, observability, and improvement. And I think Ada is a good example of that. But there are other examples in the market across industries, not just customer service, where we're seeing that play out too.

What is Ada?

Mike Murchison [5:27] Software development is the other one, and we're seeing lots of interesting application companies built to automate software code. And I think those same three principles apply.

Matt Turck [5:56] So, as a level set, let's talk about Ada. And you provide an AI product that covers chat, but also voice and email. So you've reached some level of scale and customers. But maybe walk us quickly through the history, because you've been doing this for a little while. In particular, you had to navigate that transition from the early days pre-ChatGPT to today.

Mike Murchison [6:28] Yeah, so our mission has always been to make customer service extraordinary for everyone. Our view is that businesses today—this continues to be the case, though it's changing—when we work with a business, they fundamentally have to make this trade-off as they scale their communication with their customers. They really have to choose between quality and efficiency. Perhaps all the entrepreneurs, all the builders listening can appreciate that when you have your first handful of customers, you know all those customers' names. You obsess about their feedback.

Mike Murchison [7:06] You will do anything. You'll hop on a plane to meet with them to ensure they love your product, your service, whatever you offer. But as you grow, if you've had the privilege of experiencing true scale, it turns out it becomes unfeasible to continue to offer that quality of service. And something sort of bizarre happens. You go from treating your customers as real people to treating them as these anonymous numbers that you're trying to keep at bay. Your customer service becomes a cost center.

Mike Murchison [7:34] Our view about what it means to make customer service extraordinary for everyone is to fundamentally eliminate that trade-off, that as you scale your customer contact, you no longer have to choose between quality and scale. In fact, our aspiration is to enable you to increase your quality as you scale. So, literally not only eliminate the trade-off, but actually enable your customer experience to improve the more you talk to your customers. Klarna's CEO, a couple months ago in their earnings report, shared that they're automating 50% of their customer contact, but also that they're talking to their customers more than they ever have before.

Mike Murchison [8:07] And I think it's here that we're starting to get a window into the true promise of an AI-native customer experience operation. It's one in which you don't treat your customer service as a cost. You're not trying to deflect, you're not trying to contain, you're actually trying to learn from your customers. You're trying to engage with them far more than you ever have before. And so I think that's starting to beget a whole other way of relating to your customers at scale, one in which you're learning from your customers more, you're increasing your customer lifetime value because you're selling and cross-selling to your customers in an empathetic and quality way.

Mike Murchison [8:52] You just relate to them in a different manner. But I'd say, to bring this home to your initial question, we've been on this journey for quite some time, and we started with very different AI models. We started pre-LLM with language classifiers, where the way we were focused on improving your customer service experience really continued to be on trying to deliver the highest-quality experience with the least amount of effort. But we did that with a different underlying set of technology. By virtue of our application being fundamentally built around AI models from day one, we were very early in recognizing the tipping point of the large language model, first with the BERT-based models and then the true GPTs.

Mike Murchison [9:29] And we've been really excited with the introduction of Ada 2.0 to see just the results skyrocket. We're now seeing experiences that our customers say are not only as good as what a human customer service rep can deliver, but actually increasingly better. And it's only January 2025. So, if we have this conversation again in March or June or December, I mean, I think we're going to start to see experiences that really blow people away.

Matt Turck [9:35] Maybe in February or March you'll be running on top of DeepSeek.

Mike Murchison [10:01] Maybe. Yeah. I mean, we do. And that's another whole other topic, but we definitely spend a ton on what we've learned. One of the competencies we've developed inside Ada is model testing. And this relates to the thin-wrapper question, too. I mean, it's an unbelievably complex decision to make if you're a business who's selling a different service or product to figure out what type of intelligence to tap into to automate a given task. And it turns out that if you're going to automate a conversation in customer service, you actually are going to have to rely on many different models that work together.

Mike Murchison [10:19] If you're going to deliver that quality of experience across the channels you mentioned, across voice, across SMS, across social, across what have you.

Reasoning engine, model orchestration, instruction following, routing

Matt Turck [10:54] For anybody listening to this in a week or two, I'm making the joke because we're recording this on a Monday after the weekend when everybody freaked out about DeepSeek and called it the Sputnik moment for U.S. AI, but very much to the point of which foundation models to use. So what does Ada currently run on? Do you run on multiple models? And if so, how does that part of the stack work in terms of switching to one or the other?

Mike Murchison [11:09] Yeah, so we do. We run on multiple models. The way to think about Ada is what we call our reasoning engine: we orchestrate on your behalf anywhere between 7 and 9 models, depending on the inquiry. And those models work together to understand your customer's inquiry, to create a plan to resolve that inquiry, to call upon relevant tools and knowledge and company policies to resolve that inquiry, and then to finally follow up and make sure we actually resolved it.

Mike Murchison [11:43] And those models are always changing. In fact, part of the service that we implicitly provide to our customers is we stay at the frontier of the intelligence landscape on their behalf. And that's because our North Star is to resolve the most.

Matt Turck [11:43] Mm-hmm.

Mike Murchison [12:11] We're very aligned with our customers in that, and we put our money where our mouth is. Our mission is to be their number one customer service employee. And that means that we can resolve as well, if not better, than a human. So we're always seeking to resolve even more. And that requires developing quite a robust testing infrastructure behind the scenes to ensure that we're always leveraging the right combination of models to do that.

Matt Turck [12:17] And to the extent you can talk about it, can you mention specific models? Some GPTs, some Claude, some—what else?

Mike Murchison [12:53] Yes. So we work today—right now, our ensemble is mainly a combination of OpenAI and Anthropic models, although there are some open-source models that we're also leveraging. We test those models according to the module or specific use case that we're applying them to. We always have at least a capability that is leveraging the deepest reasoning, and that is connected usually to our planning module. So the intelligence that's determining what creative plan to propose to resolve your inquiry. That's typically a reasoning-based model.

Mike Murchison [13:28] I think we're very quickly about to reach a tipping point whereby the most performant reasoning model available in the world may be overkill for the average customer service use case. Sort of like, you don't need a PhD to chart a path to resolve your order refund request, even if it is adhering to a bunch of unique company policies. Although we'll see. I think that's TBD. So that's sort of an example of what we might test against. One thing that is definitely true that we measure quite extensively is a model's ability to adhere to instructions.

Mike Murchison [14:08] So we are quite maniacal about instruction following and testing a model's ability to execute on what you asked it to. And that really relates to part of the core value that we provide to our customers that I like so much about Ada, is the control that we offer our customers. We really do ground it in an employee paradigm. If you ask a member of your team as a manager to do something, you expect that they'll follow through. And the same is true of this non-human workforce that is evolving around the world.

Mike Murchison [14:33] And which Ada is a part of, you need to be able to follow instructions. And that turns out to be quite a technically complex problem when you start to have thousands of instructions. So we test regularly models' ability to adhere to those complexities.

Matt Turck [14:57] Fantastic. To unpack some of it: on the model front, it sounds like you do dynamic routing. So, meaning that for one customer, actually, each inquiry may go to a different model. Is that depending on the query? Do you think about cost as well? Do you think about latency? What are the factors in routing?

Mike Murchison [15:27] Yes. So it depends on the inquiry type, the number and specific types of models that will be leveraged. For example, if an action is required, we will leverage a different model than if only a retrieval, knowledge retrieval, is required. We optimize for performance. We are attuned to the cost. We're not very cost-conscious right now. We are focused really on maximizing resolution. And as you know well, we're seeing model costs plummet. They've plummeted 90% in the last 12 months.

Hybrid systems, finetuning, customization

Mike Murchison [15:45] I think there's lots of performance gains on the efficiency side to tap into ahead if we wanted to realize those. But we're most focused on performance, and that includes the quality and latency of our models.

Matt Turck [15:53] Do you still use some of your own models that you built pre-LLM?

Mike Murchison [16:33] We do. There are parts of our product where we have our classifier models that we trained ourselves. We also leverage a number of open-source models that we fine-tune. So we are finding that some specialization is beneficial. I think long term, there's a question about how specialized this ensemble of models we have actually is. It's definitely the case that as the instruction set, we think of it, of each of our customers grows—meaning as they go from having coached their AI agent to perform differently in 50 different ways, to 500 different ways, to 5,000 different ways—that there'll be opportunities to fine-tune and/or train new models in a customer-specific capacity.

Matt Turck [17:06] You customize and you fine-tune on sort of, like, an industry basis, meaning those are the things that you need to do to be a good AI customer agent, but not yet on a customer-per-customer basis.

Mike Murchison [17:38] Right. And we align each model to, or we instruct each model to adhere to, the specificities, the instructions, of each customer. But we don't train models from the ground up based on individual customer data. And because we haven't seen that that yields a performance improvement yet, I think it remains to be seen if that's actually ever going to be the case. I think that's an open question for us. I think there probably are cost advantages to doing that, potentially. But given the accelerated pace of the intelligence landscape and how quickly it's changing, we don't think there's—we don't think now's the time to do that.

Matt Turck [18:22] What it is, by the way, based on the way you described it, it feels like Ada truly deserves the moniker of AI agent. So AI agent is one of those terms that everybody uses, but it means different things. But you have different components that interact with one another, and there's a decision-making process, which is different from a copilot kind of scenario where you don't have a human necessarily telling you what to do. So just for definition purposes, again, in a world where everybody uses agent for different things.

Prompt engineering, observability, self-improvement

Matt Turck [18:52] So we talked about the models, we talked about how you route the query to the models, the prompts themselves. So we talked about fine-tuning. Prompt engineering is a part of what you all do, right, in terms of creating a set of instructions for the industry, but also per customer?

Mike Murchison [19:21] Well, it depends on how you define prompt engineering. One way to think about Ada is: Ada allows you to hire an AI agent, control its performance, and observe how it's performing—the level of transparency we provide into what your AI agent is actually thinking, why it did what it did. That's where we expend a ton of effort. And then fundamentally, where we really are—our bread and butter, what we think most about—is the speed of improvement. Once we give you that level of control, once we give you that level of transparency, we allow you to improve over time.

Mike Murchison [20:00] Our agents are now autonomously resolving—we have a new high watermark as of last week, week before—up to 85% of all conversations that they're seeing. That's without human involvement. To your point, this is not a copilot experience. This is pure customer-facing AI. Underneath the hood, to actually deliver a personalized experience for each one of your customers, the level of prompt engineering that's actually required to do that starts to become really untenable. This is why you quickly start to realize, and teams who start to build their own agents inside their companies quickly realize this, that really what you need to do is assemble a custom prompt for every individual customer, right?

Mike Murchison [20:40] You need to know that this is Matt, that Matt is part of this VIP customer segment, that Matt is based in New York City, that Matt's on these product SKUs, and that your company objective is to upsell Matt, to ensure that he gets his immediate help he's asked for, and to make sure that you adhere to his prior preferences, which is to always follow up over email. Being able to adhere to that level of requirement, as an example of prompt engineering, that's an example of some of the complexity we remove from our customers.

Mike Murchison [21:08] So effectively, Ada, as this AI CX platform, as we think of it, really behind the scenes is creating a series of unique prompts for every one of your customers. And we're assembling those for you based on the instructions that you've given your customer service employee.

Matt Turck [21:15] So there is a whole layer of prompt engineering behind the scenes in what you pass to the actual model.

Mike Murchison [21:39] That's right. And that's a way to think about it: behind the scenes, when your customer service agent is talking with one of your customers, Ada has assembled and created a unique prompt for that customer that adheres to your policies, to your specific company objectives, whatever strategy you have for this customer based on who they are. And it turns out that's almost exactly what a human customer service rep does. They look up who you are, they try to figure out what the objective is, they learn who you are.

RAG (Retrieval-Augmented Generation) and AI as a judge

Mike Murchison [22:07] And it turns out that humans actually aren't that good at that. Like, it's actually a really hard thing to do when you start to add on all the different objectives that a business has. So, yeah, I never really thought of it that way or described it that way, but that is, in many ways, technically what it is doing. We're assembling a unique prompt for every one of your customers as you interface with them.

Matt Turck [22:14] And what about RAG and pulling information from customers' databases? How does that work?

Mike Murchison [22:52] That's one of the modules in our reasoning engine. So one of the steps that we will follow, again, depending on the inquiry, is we will perform retrieval-augmented generation to surface the relevant knowledge from your knowledge base or series of company policies that you've integrated into Ada. We are very focused on the accuracy of our generations. So we use language models as judges to ensure that your generations are grounded in the knowledge and policies that you've connected to Ada. So RAG is relevant to that as well.

Guardrails and security

Mike Murchison [23:06] But yeah, we think of knowledge search essentially as one of the functions that your agent might perform, depending on what task it has at hand.

Matt Turck [23:30] How do you think about guardrails and then security? Those being two problems. So, guardrails: making sure that your agent behaves ethically and there's no terrible language that's used, on the one hand, and then security, meaning that malicious actors cannot extract from your agent information that they're not supposed to provide. What do you do there?

Mike Murchison [24:02] Yeah, so we're maniacal about that. We think that adherence, accuracy, and safety—those three things are our job. You should not have to worry about Ada adhering to your instructions, the accuracy with which Ada speaks, or the safety or security with which Ada operates. And we expend a lot of our effort in ensuring those three things. Increasingly, we make those assurances in runtime. So it's actually not possible for Ada to deliver an inaccurate response. We will soon do the same in terms of adherence.

Should we expect perfection from AI?

Mike Murchison [24:35] It soon will not actually be possible for Ada to generate an answer or say something to you on the phone that doesn't adhere to one of your instructions, some of the coaching you've provided us. I think that will be something the market will demand and will just become table stakes. It will start to be weird before too long that your agent doesn't follow through on what it says. I think that there's also an important reflection, though, as we think about customer service broadly.

Mike Murchison [25:08] And I spend a lot of time talking to customer experience leaders. And one of the things that I find striking is that there continues to be this expectation of perfection from AI. And I think the most forward-thinking customer experience leaders understand that this isn't realistic. But some of the market does not. And it's a bit ironic because humans are far from perfect. And so there's a tendency when AI is deployed—we see this with autonomous vehicles—that it doesn't matter what the driving track record of humans is, the second that you get the first Tesla self-driving crash, it becomes sensationalist and front-page news.

Mike Murchison [25:55] I think the same thing is true with AI being deployed in production environments across use cases, but in customer service too. So I think there's an important risk calculation that leaders need to perform, where I think it's valuable to think about the performance and consistency of their customer-facing teams today, and then benchmark that against AI's performance. And I think, if you make that comparison today, the comparison is quite favorable, and it's only getting better.

Mike Murchison [26:11] But it's just interesting to me that a lot of the market doesn't default to thinking that way.

Measuring “resolution”

Matt Turck [26:47] How does that work? So I'm fascinated by the concept of resolution in the customer service industry, because resolution can mean that a conversation ended, or it can mean that the customer hung up in frustration, or it can mean that the problem was satisfactorily resolved. And you could make the argument that when you're on the phone with someone, the customer agent can sense how the conversation ended. But maybe AI does or doesn't, I don't know. But how do you think about the concept of resolution?

Mike Murchison [27:20] We think about it a lot. At the end of the day, we think of Ada as a resolution company. So our North Star as a company is to resolve 100% of conversations. And resolution for us means that the customer got the help they were looking for. We actually satisfied the initial intent of the customer while conforming to the policies and guardrails of the business. This resolution, in many ways, I think, is an example of how AI is transforming business-to-consumer, the economics of businesses and consumers interfacing.

Mike Murchison [27:44] Prior to language models, it was prohibitively expensive to understand the quality of your customer service. Everyone's heard that saying, "This call may or may not be recorded for quality purposes."

Matt Turck [27:46] Your call is important to us.

Mike Murchison [28:13] Yeah, your call is important to you. It just is these brutal statements that erode your confidence in the business. But those, up until very recently, most of those transcripts were obviously not transcribed. And those that were transcribed, very little was actually done with that information. We now live in a time when 100% of interactions are automatically transcribed, and their quality is automatically assessed. There's no more human annotators who have to painstakingly, at $15 an hour or whatever, review the quality of a conversation.

Mike Murchison [28:42] An AI model can now do that as well or better than a human being. We've spent a lot of time doing this at Ada. We've reviewed so many conversations as a company, and we regularly test the performance of our language models on an annotation-quality basis compared to human annotators. And we can say with confidence right now that, yes, that is exactly true. We can measure the quality of a conversation transcript as well as a human annotator, and we can do it with 100% coverage.

Mike Murchison [29:12] And so when you can do that, you can actually raise the bar on your definition of resolution. You can actually know: did the customer get what they were expecting? Was this conversation safe? Was it relevant? Was it accurate? And you can measure those things with confidence. You can also learn a ton from those conversations. So I think things are changing in two dimensions. One, there's greater assurance than ever that the quality of conversations you're powering are, in fact, high quality.

Mike Murchison [29:28] And two, there's a greater opportunity to actually learn from these conversations at scale because every single conversation is annotated.

What actions can Ada AI Agents take?

Matt Turck [29:56] And what can Ada do to help solve some of those problems? Right. And the history of chatbots back in the day was that it was largely FAQ, right? You could just answer whatever question was in the knowledge library. But today's AI agents are infinitely more powerful. What kind of tools, I guess, do Ada agents have at their disposal, and what can they do?

Mike Murchison [30:23] They can do quite a bit. They can take actions, so they can perform, execute complex workflows that would formerly require a human agent clicking around in a back-office system, making changes to an account manually. They can do that proactively, so they can recognize that you're trying to accomplish something within a product and proactively reach out and offer to do that for you. They can very easily personalize the experience, so they can know who you are and what your relationship has been with the business historically.

Mike Murchison [30:52] We like to think of Ada as having perfect memory. One of the things that underpins such negative customer experiences is when you don't get credit for the relationship that you have with the business. Like, you've been a customer for 20 years.

Matt Turck [30:52] Yes.

Mike Murchison [30:53] But the person on the line doesn't know that.

Matt Turck [30:58] After you give them your name, your date of birth, and they have no idea who you are.

Mike Murchison [31:29] If you only knew how many people I have brought your way and how valuable I am to you, you would treat me differently. That experience will disappear. So Ada can bring you perfect memory. What I'm very excited about is, up until relatively recently, the actions that we can perform, the processes that we can automate, have been mostly limited to API availability. And as we've started to see, about a year ago, we gave a preview of what we call Web Actions. You've probably played with Operator, which came out a couple of days ago, which is just a mind-blowing example of a language model's ability to navigate user interfaces autonomously.

Authentication and personalization

Mike Murchison [32:12] I think we're very soon going to be in a position where there's really no action that Ada can't perform on behalf of a customer. And the limiting factor will be whether or not the business wants that action to be automated. And so we're very excited about that. There's a whole set of use cases that have been limited because of a lack of API availability, where we've worked closely with our customers to create those APIs that perhaps may not be as necessary ahead.

Matt Turck [32:38] And a little bit to the Operator discussion, how does authentication work? And I'm saying this in the context of Operator because there were a lot of discussions about needing another browser and the fact that you were never kind of logged in and therefore you had to re-log in for Operator, which—not to say bad things about Operator, because it's a brand-new product and it will evolve tremendously. But how does that work in the context of a conversation with Ada?

Mike Murchison [33:01] So you can deploy Ada in an unauthenticated manner or authenticated manner, or both. Many of our customers deploy us using our SDKs, and they put us behind a login. So Ada's automatically authenticated by virtue of you signing into your application yourself. We also support the experience for you to actually log in through Ada. If you ask a question, Ada detects that you're a customer, or you state that you're a customer and you're not logged in, Ada can log you in so that you have access to your account.

Mike Murchison [33:38] And it is a personalized experience. And I think increasingly ahead, we're very much trying to get to an experience where Ada always knows your name regardless of where you're coming from. And we make auth as seamless and frictionless as possible. Things start to get complicated when you start to think about omnichannel and multimodal experiences. But by virtue of Ada being a platform where we build once, deploy everywhere, you manage your single AI customer service agent, and that persists across email, across voice, across all text-based channels, messaging channels.

Mike Murchison [34:17] The ability to traverse channels and maintain an awareness of who you are becomes something that is beginning to unlock experiences that didn't exist or were much harder to find before. An example of that is a lot of our voice AI agents will often send you instructions over text. So you'll be speaking with them. It's so annoying to talk to a voice agent or someone on the phone and have them give you a password or give you a set of instructions or something that you have to write down yourself.

Mike Murchison [35:08] Ada will just automatically text you whatever instructions you needed or text you a link to follow up after, what have you. But the ability to sort of unify this customer experience across modalities is something that we're pretty excited about. And the ability for your identity to persist across those channels is something that we're trying to make sure is the case for every one of our customer experiences, that you have a single record and that traverses all channels.

Handoff vs human delegation

Matt Turck [35:31] And as of now, not just for Ada but for the industry in general, there is still a concept of handoff to a human at some point. So, a couple of questions. One, how does that work, and how do you know, other than when the customer asks you, when to do that? And two, is that something that you think disappears over time?

Mike Murchison [36:05] So, yes. This is interesting and very topical for us. Ada is an AI-native application that has no human customer service agents in it. So whenever a human customer service agent is needed, we'll route, we'll hand off to a human customer service agent who lives in an incumbent system, like a Service Cloud, and we will pass over all context. We'll make sure that customer service agent can seamlessly pick up the conversation where it was left off. As we've started to achieve higher and higher resolution rates, we're starting to challenge our idea that a handoff is a failure, because for most of our company's history, it's been a failure.

Mike Murchison [36:36] If we have to involve a human, it's on us. That's a mistake. Increasingly, particularly as we look at this remaining 15% of conversations, those conversations tend to be incredibly complex, and many of them actually—

Matt Turck [36:36] Are important.

Mike Murchison [36:47] Require some level of human approval just based on the processes that exist inside our customers. And so increasingly, we're starting to think less in terms of handoff and more in terms of human delegation. And we're very excited about that because what it means is that instead of just simply handing off a conversation to a human agent and having them figure it out, increasingly, we're going to be essentially the quarterback for this issue inside the company.

Mike Murchison [37:19] We're going to tell the customer, "Don't worry, Matt, leave it with me. I'm going to need some time to resolve this inquiry." We're going to go inside the company, and we're going to involve everyone who needs to be involved in order to resolve this inquiry. And then we're going to get back to the customer.

Matt Turck [37:20] Mm-hmm.

Mike Murchison [37:37] And so, was that conversation handed off? Yes, sort of. But we stayed in it. And we ensured that the resolution was incredibly high quality and that the right people were involved.

Matt Turck [37:45] Interesting. And then presumably you'll be able to track a whole log, summarize the issue to date: this is what we tried, this is what we said.

Mike Murchison [37:53] Exactly. And there's lots that a business is going to be able to learn from Ada spearheading that conversation that they wouldn't have been able to learn otherwise.

ACX (AI Customer Experience) and the future of customer service professionals

Matt Turck [38:29] Yeah, which brings this really interesting part of the discussion around ACX that you alluded to earlier. It sort of feels like this whole world of customer service, from a job perspective, a profession perspective, is completely changing. So what is today, but also two, three, five years from now, if I'm a customer service expert at any of those companies, Monday or Square or ZoomInfo that we mentioned earlier, what will my day-to-day job be?

Mike Murchison [39:02] So I think broadly, the biggest shift that's happening is customer experience, customer service professionals are evolving from reacting to incoming customer service inquiries to proactively managing them by really serving as the manager of an AI agent or a group of AI agents. The role is quite different. It's the same in that they're focused on improving the customer service experience, but the way that they do that is much higher leverage. And by virtue of their agent being so performant, it starts to become a far more creative role.

Mike Murchison [39:37] They start to be able to be far more focused on not just driving the efficiency of their customer service operations, but generating new revenue and fundamentally increasing the customer lifetime value of their customer base. We spend a lot of time helping our customers build their ACX organizations and identify the right kind of talent and chart the career paths of folks.

Matt Turck [39:41] And ACX, I don't think we ever defined it. That's AI customer experience.

Mike Murchison [39:41] AI customer experience.

Matt Turck [39:43] AI customer experience.

Mike Murchison [40:14] Yeah. So this is the discipline of deploying customer-facing AI in your company and ensuring that you're operationalizing it most effectively. I think the skill set to develop there organizationally—our customers call it ACX, but think of it as AI customer service management—is the discipline of learning how to continually improve the performance of your AI, your customer-facing AI. And it turns out that the companies that are doing that most effectively have new roles. They have folks who are responsible for the performance of their AI agent, and those folks are given a level of accountability and access inside their companies where they're empowered to do that.

Mike Murchison [40:51] So that's been really interesting to see, and it's been very interesting to see the career paths that those folks go from working in the contact center to now running the ACX team, and just the scope and scale of the impact that they've had has been hugely different. It's one of the things I love most about the impact it's had so far.

Matt Turck [40:54] So they went from managing humans to managing AIs?

Mike Murchison [41:23] Yeah. Or in some cases, not managing humans at all, but being a highly productive, highly empathic customer service agent to now managing an AI, going from helping 20 customers a day to helping 2 million customers a day. And then I think one of the things that's most interesting about this transition, and it kind of goes to that point of when you deploy customer-facing AI effectively, you end up talking to your customers more. The ACX team has access now to all this conversation data that didn't exist before.

Mike Murchison [41:54] And so they are now increasingly informing company strategy. Folks are coming to them saying, "What do our customers want us to build next? What do our customers think of this initiative? Are we supporting our customers in Asia effectively? What do you think we should do? What are our customers like?" There's long been this promise of voice of customer. It's always been this little thing.

Matt Turck [41:58] Voice of customer is the best poll of all polls, right? Because you talk to customers all day.

Mike Murchison [42:12] Yeah. And now it's becoming real because AI is enabling customer-facing teams to have the actual data and credibility to back their instincts. What they've long known, they're now backing with real data. And that's changing the influence they have inside a company.

Leveraging analytics and customer support data

Matt Turck [42:30] So if I'm a CX employee or head of CX and I live in Ada all day, what do I see in terms of analytics? If I want to know what my customers are talking about, what is it like? I do have dashboards.

Mike Murchison [42:54] A wealth of information at your fingertips. So you have everything from abstract metrics like, what is my resolution rate? How well am I resolving? What's my customer satisfaction or NPS? To, hey, what are the topics that my customers are asking me about? And how well am I doing within those topics? And how have those topics changed in the last six days or seven days? To, hey, what are the opportunities I have to improve along these dimensions that I care about, either globally or within specific topics?

Mike Murchison [43:35] All of that is presented in a really easy-to-understand manner. And some of our customers, they'll drill down into the individual conversation level, where they'll actually see, for this individual customer, I want to treat this individual customer differently. And so it's really a level of transparency into how your customers perform or how your agent is performing that is teaching you new things about your customer experience. We had a customer recently. They're a large e-commerce business.

Mike Murchison [44:09] They discovered through our reporting that they should probably create a new product. In other words, in this case, it was a SKU of sunglasses that they didn't have. It was pretty clear that there were enough complaints around one of their existing products, and there were enough recommendations or feedback in general that it was suggesting, like, we should probably create a new SKU. And they did.

Matt Turck [44:10] Yeah.

Mike Murchison [44:29] Like, they did. And so it was such an amazing example of this team that, prior to deploying Ada, was this reactive customer service team, to now this team that was fundamentally driving innovation for the business because they've been empowered with customer data they didn't have before.

Matt Turck [44:46] And you mentioned revenue and increasing lifetime value. So I guess this is one example, but is there a concept, I assume, also packaging offers and trying to upsell? How does that all work in a way that doesn't annoy people?

Mike Murchison [44:47] Doesn't suck. Yeah.

Matt Turck [44:47] Yes.

Mike Murchison [45:10] I mean, everyone's experienced this. This is an interesting topic. So, for folks like you and I who have been tormented over the years online with really negative cross-sell and upsell experiences from businesses, I think we have a bad taste in our mouth from those. I'm quite focused on making sure that as Ada sells, it's not a negative experience. It is the case, though, that I'm convinced that there are a number of problems that a customer has where actually the right way to solve it is to offer either more of your service or an adjacent service that you offer.

Mike Murchison [45:51] And those are what we think of as high-empathy, serve-to-sell experiences. In the retail or hospitality world, that's as simple as, like, “Would you like fries with that?” because you've ordered a burger and a lot of people like fries, to a more technical problem that you're facing within your current marketing SaaS product that you're paying for that turns out you have an adjacent product that solves exactly that problem. There's this idea that sales and support and marketing are like these three different categories.

AI agents for cross-selling and upselling

Mike Murchison [46:25] And I think that's because folks tend to think too department-first as they think about the customer experience. But if you really just think about how do we resolve the most with the highest quality, how do we just power the best experience? The customer doesn't care about the departments that you have internally, right? They just want the help they're looking for. And so that's how we think about cross-sell and upsell. We think about it as helping the customer. And so long as it's a high-CSAT experience, there really is a win-win-win to unlock a better experience for the customer, a better experience for the business, and a better experience for the person who's managing it.

Matt Turck [47:00] I guess the reason why there is sales and marketing and other departments is also because we as humans can only learn so much, and it takes many, many years to be a good salesperson, many, many years to be a good marketing person. But if you have a form of intelligence that knows everything about everything, then yeah, you don't need the specialization, right? You can have one central brain.

Mike Murchison [47:31] Yeah, I think actually the specialization emerges because of this trade-off between quality and efficiency. Like, I think if you look at small businesses, think about a restaurant. The best waiter is serving and selling you all the time, right? They are interrupting your conversation in a way that's not rude, at the right time. They're very attuned to when to come and speak to you. They are deeply knowledgeable about the menu when you ask questions.

Matt Turck [47:37] They know that somehow the best recommendation, the best thing that they recommend, is always the most expensive one.

Mike Murchison [48:04] Maybe, but if they do that too overtly, then that's a negative experience, right? So they toe that line really, really well. They will upsell you on the bottle of wine that tastes better. But the point is that at small scales, serving and selling are the same thing. They are unified. It's only once we've traded efficiency for quality that they become specialized. And AI's impact on customer-facing experiences is going to be the unification of those things again.

Mike Murchison [48:23] There will just be a single customer-facing AI that most businesses deploy, and that AI will not only serve as well or better than a human, but it will sell as well or better than a human too.

Traditional AI chatbots vs the new generation of AI Agents

Matt Turck [48:46] And you alluded to this a minute ago, but sort of the elephant in the room is indeed that chatbot experience that we've all had that basically sucks, where it's bad enough to talk to a human, but it's even worse to talk to a chatbot that just doesn't seem to understand and gets you into loops, all the things. Where do you think we are, for all the progress, in terms of that holy grail of people wanting to be online and actually just talking to a chatbot because the chatbot is so much better than any human they'll ever speak to?

Mike Murchison [49:32] I think from an industry perspective, things are changing very quickly, but most people have still had a mostly negative experience with the chatbot. And I think there's a couple of reasons for that. One is because the way this technology is being deployed is with the wrong measurement, in our view. Most of those chatbots that you talk about, those really annoying ones where you're just trapped, it's because the chatbot's job is to deflect you or to contain you. In other words, the way the business thinks about the value of that chatbot, they think about it exclusively in terms of keeping you away from their human team.

Mike Murchison [50:05] But if you focus on resolving the customer's issue, the quality of that experience improves dramatically. What we see inside Ada is when an issue is resolved, when we resolve an issue, it is almost the same thing as a five-star human-reported CSAT experience. They are so tightly correlated. So experience improves when you have the right measurement, is my first point. My second point is that while most people have still had a mostly negative experience with chatbots, there are increasingly people who have had some unbelievable experiences with AI.

Mike Murchison [50:37] If you zoom into that cohort, what I think we're increasingly going to see is that cohort is actually increasingly asking for the AI. They are increasingly hoping that they can speak to an AI on the phone, over SMS, over email. And that is the experience that's going to become the norm. That will switch. It'll start to be the case in the next few years, I think, where unless you have the right AI deployed in your contact center, for your brand, you'll actually be at a disadvantage as a company, that your customers will begin to expect and demand this from you.

Emotion, empathy, personality

Mike Murchison [51:24] Unfortunately, there will be some pain to get there because you're right, there are still so many negative chatbot experiences. And just as there are still so many mostly negative customer experiences that we all, for some reason, continue to tolerate, it is amazing what we're subjected to. We've been beaten down so much as consumers into thinking that it's okay for your bank to waste four hours of your day.

Matt Turck [51:51] Yeah. What's very interesting in this whole customer service space and the conversation we're having is that there's a clear emotional aspect to this, which may or may not be present in other parts of the business world. I think you mentioned the term empathy. How do you think about that? Sort of almost like emotional customization. Is that built in the foundation models? Is there something that you build on top to do a good job there?

Mike Murchison [52:23] Yeah, we do, out of the box, make Ada, as a sort of default experience, as empathic as we can. And the reason for that is, again, if we're focused on resolving a customer's issue, key to resolving a customer's issue is ensuring that you convey that you understand their issue. It's not sufficient just to do it unless you explain what you did and why. And so empathy, I think, is inherently connected to that. I will say that we do see our customers customize that a fair amount.

Mike Murchison [53:02] So the way that a business empathizes starts to really touch on the brand personality of a company. And different companies like to handle that differently. Some like to be quite verbose in how they empathize. Others like to sort of empathize after the fact. They just want to get to it, resolve the issue, and then explain what they did afterwards. I think it depends a lot on the business. I also think that over time we'll see this dimension of Ada really become personalized on an individual user basis.

Matt Turck [53:14] Like during the conversation.

Mike Murchison [53:21] During the conversation, yeah. Depending on who you are, we will adapt to your individual preferences.

Matt Turck [53:32] Is that part of your onboarding when you have a new customer? Do you ask them, okay, well, how do you want us to handle your customers? Are you long-winded, to the point?

Mike Murchison [53:46] We actually offer a bunch of different sort of out-of-the-box personalities that businesses can pick from. And then, because of what we call coaching—it's one of the most powerful capabilities Ada has right now—it's the ability for the ACX team to, in their natural language, coach their agent to improve. So they can offer specific instructions at a conversation level to behave this way differently next time, and then observe their agent perform that differently for the next customer, both in a real and simulated capacity.

Mike Murchison [54:33] And we see that empathy is something that is often coached on an individual basis. So, in this situation, make sure to be more empathic, because I observed here this is actually a big deal to us. When a package is late, we actually, as a brand, really want to make sure the customer knows how deeply apologetic we are. And we really want to make sure that we're playing back to the customer, like, "I know that this is upsetting to you."

Mike Murchison [54:55] "I can only imagine how upset I would be if I were in your situation. And here's what we're doing about it." That's an example of empathy, customers dialing it up and dialing it down on an individual basis differently.

Transparency and AI improvement

Matt Turck [55:27] And just to verbalize the thought, that's another example where the CX team gets to program AI in a no-code way, for lack of a better term. But I guess it's an important point to not leave unsaid, because there was this perception that AI was a black box. But actually, very much part of what you provide is this ability to configure it in a way that's fairly transparent.

Mike Murchison [55:49] Exactly right. So that level of transparency allows the business to improve. They can see that the reason that their agent wasn't super empathetic with the order refund conversation they were powering, well, it was because it was adhering to your guidelines of being really direct and to the point and actually not expressing empathy here. And because we give that level of visibility, then we make it incredibly easy for the CX team to coach their agent to improve and then adhere to those instructions over time.

Mike Murchison [56:25] That feedback loop is just so fast and so instant that we just see the performance of these agents improve week over week over week over week. And that kind of touches upon what's different about this moment we're having in the world of technology broadly, but within AI customer service specifically. And that's that this is not like traditional software at all. Businesses should not be purchasing customer service software in 2025 and expecting the results to be consistent.

Mike Murchison [56:49] They're not buying a piece of software to solve a problem today. What they should be purchasing is an expectation that their new non-human employee is continually improving. And that's the same expectation we have of our human workforce. When you hire someone, you and I both know that we expect that person is going to take some time to ramp, but we hope with the right coaching, the right management, the right support, that they're actually going to be way more productive and have a much bigger impact six months from now, a year from now, than they have today.

Mike Murchison [57:37] And we'll be really proud of them when we get there if we know that, as their manager, we helped them, we played a small part in helping them improve. That's what we see at Ada. And that's, again, one of the things that correlates with our customers' career trajectories, is this level of pride they have for their agents. They remember how capable their agent was on day one, and they're so amazed and so proud of what it's doing at day 120.

Matt Turck [57:39] Look at them, AI of the month.

Mike Murchison [57:49] Yeah, literally. That's funny, but it's literally true. Many of our customers have pictures of their AI agent's performance on the wall.

Matt Turck [57:50] That's amazing.

Managing AI: the measure-coach-improve loop

Mike Murchison [57:58] And the monthly performance of them. It really is amazing, and it will become totally normal, I think, ahead.

Matt Turck [58:14] So you've been doing this for a while now across a bunch of different customers. What have you seen works and doesn't work in terms of accelerating that progress that you just described? In particular, the concept of onboarding AI, how do you accelerate that?

Mike Murchison [58:53] We observe effectively the companies that are maximizing, achieving the best results, they're going through a cycle of a couple different phases. They are, as you mentioned, onboarding a new agent, measuring its performance routinely, coaching it to improve, and then extending it into new places, whether that's in front of more of their customers within the channel where they first deployed it, or across all other channels, or building it natively into their application if they're a software business. So it's really that loop of measure, coach, improve after they've onboarded that we see. Really, the speed at which companies move through that is a key predictor of their overall results.

Mike Murchison [59:27] I think one thing I've learned over the last couple of years of helping businesses operationalize this has been on the onboarding dimension, we are now very focused on making sure that the first day of your new customer service employee, like on day one, that employee is amazing.

Matt Turck [59:28] Mm-hmm.

Mike Murchison [59:56] It used to be the case we provided far more configuration and control. It was almost like, we're going to give you a platform, you are going to create the agent of your dreams on day one, and then you will launch it. Now we've discovered that we can just accelerate results far faster by giving you an agent in a box that is great, and then just ensuring that you know how to coach it to improve over time. That's been a big unlock for our customers, and that's been something that language models have made far easier, because it's far easier.

Ada Voice and Email

Mike Murchison [1:00:15] It's so much easier to create that initial agent using Ada that's preconfigured for your business than it would have been pre-transformer.

Matt Turck [1:00:33] Let's talk a little bit about voice and email. So, the multimodal aspect of Ada. When did the voice product come out? What does it do? Any kind of surprises, lessons learned there?

Mike Murchison [1:01:06] Ada, for most of our history, has been messaging-focused. We launched Ada Email about six months ago, and then we launched Ada Voice around that same time, maybe a couple months before that. Results and adoption of both these products has been awesome. It's really compelling to see how, after you have a very performant, amazing messaging-based AI agent, how easy it is to just now deploy it in messaging or now deploy it in email and on the phone. In the sort of contact center world, the term for this is a hybrid agent.

Mike Murchison [1:01:43] It's a customer service rep who handles multiple channels. They both text, they both type, and they talk to you on the phone. A lot of companies specialize. It's very expensive to have hybrid agents. So the fact that it's so easy with AI to have a hybrid agent is a big deal. That being said, we've learned about the idiosyncrasies of each modality. For example, with email, we've learned that a lot of our customers find it weird that Ada responds so eloquently, instantly to email.

Mike Murchison [1:02:22] And depending on the business, we actually had to build a capability to allow our customers to delay Ada's emails. It's pretty interesting where user expectations are, right? So, in other words, customers weren't opening these emails because they didn't believe that the email could be helpful if it came so instantly, right? So there's like an eight-minute delay that a lot of our customers use.

Matt Turck [1:02:25] Please reply to me within 24 hours, but no sooner than 24 hours.

Mike Murchison [1:02:56] That's right, exactly. So that's obviously a temporary phenomenon; that'll disappear. But it does speak to wherever our customers' users are. It speaks to the importance of perception. On the voice side, we've learned a lot about latency. I mean, that's, I know, a big topic on this podcast. Definitely it's been hard to achieve human levels of latency. I think a human conversation is about 200, 250 milliseconds of latency.

Mike Murchison [1:03:27] We're not quite there yet. The world isn't quite there yet, but we're scratching at it, and we'll be there soon. But we definitely have learned about just the relationship between latency and experience quality. Five seconds or four seconds of latency, it's a frustrating experience. Doesn't matter how capable you are, if you're not fast enough over voice, it's painful.

Matt Turck [1:03:48] Which is a very psychological thing, right? Because, I don't know, at least I personally will gladly wait like three, four seconds seeing GPT-4o think, quote, end quote. But yeah, four seconds on the phone with something that sounds like a human, that's much more painful.

Mike Murchison [1:03:49] Yeah, really painful, right?

Matt Turck [1:03:50] It's strange.

Mike Murchison [1:04:17] We've also learned a lot about, we hear this from our customers all the time, the difference between real-world deployments of voice AI and internal or sandbox deployments. It's one thing to ship a voice AI agent that works over the web, leverages WebRTC, has next to zero latency, and is a human-quality-level voice. It's a totally other thing to deploy an AI voice agent in front of tens of millions or hundreds of millions of users in traditional telephony, where it's a low-bandwidth signal.

Mike Murchison [1:05:02] There's a ton of noise in the background. There's different languages. I mean, it's a much, much more challenging experience. That's where I think the opportunity really is, and that's where we really focus. We really focus on making sure that Ada Voice works for real customers in different languages over traditional telephony.

Matt Turck [1:05:19] Yeah, of course, the language aspect that we haven't explicitly talked about, but that's a huge part of this, right? You have now a customer service representative that will speak whatever language perfectly, which we take for granted. But just that is mind-blowing when you think about it.

Mike Murchison [1:05:53] Totally. And it's also the case that often, many of these customer experiences where the business has the greatest opportunity to really increase their customer loyalty, to really make your day, those are often the experiences that are actually hardest to resolve. Because it's someone who is calling you in their car on the highway in a panic, and it's really, really hard to hear what they're saying, but they actually need you most in that environment. And so we really are trying to make sure that we're solving for those real-world constraints and those real-world environments.

Future predictions for AI

Mike Murchison [1:06:25] And that's very, very hard. We are very excited about voice ahead. We're tripling down on native voice, growing our team with native voice, growing the team across the board. But we have a particular focus on native voice, and we have a lot of work to do there, I think, to deliver the kind of experience we know we're capable of.

Matt Turck [1:06:45] So maybe to close, it sounds like we're in an early inning of a bright future that some may consider scary, some may fully embrace. Where do you think we are in two to three years from now? Any kind of thought or prediction, whether for Ada, for the industry, things you're excited about?

Mike Murchison [1:07:16] I'm really excited for people, for the average person, to have a very positive experience with a customer-facing AI. I think that is going to come in the next couple of years, potentially much sooner. But I think the reason I'm excited about that is the quality of customer experience that we as a population expect from businesses. Like, we are going to raise the bar for ourselves, and AI is going to do that. And the consequence of that will be that if you are a business that offers a low-quality customer experience, you are in trouble.

Mike Murchison [1:07:52] And so I predict that customer service will become a far more strategic investment for the average business as AI demonstrates to the world that the quality of customer experience that lies ahead is way higher than we've come to expect historically. And that's going to mean a really great thing for those leaders who take the bold step of deploying this stuff now.

Multi-agent collaboration

Matt Turck [1:08:14] So as we get closer to the end of this conversation, so far in this entire conversation, we've talked about customer service a lot. As we move into this agentic future, how do you view your customer service agents maybe collaborating with other agents? What does that look like in the future?

Mike Murchison [1:08:46] I think it's likely the case that customer service becomes far more horizontal. Certainly, our focus at Ada is not only supporting your customers, but also selling to your customers, because we believe that the best customer service experience spans both sales and support. However, I think secondly, it's also the case that the best customer service teams collaborate very effectively with other departments inside the enterprise, like the product development department. And if you think about the root cause behind a given customer inquiry, often the root cause of the customer service problem is a product or service deficiency, and that is just such valuable feedback for the product development team, the R&D team, to build something new or fix something existing.

Mike Murchison [1:09:39] And so what I'm very excited about is Ada's ability to collaborate with other agents. I fully expect that, in collaboration with some of the code generation agents, for example, Ada will be able to go from customer inquiry to new product being shipped fully autonomously. And that is ultimately, I think, the holy grail for many customer experience leaders as they think about not only, how do I resolve this issue, but how do I eliminate this issue from ever appearing ever again? And the way you do that is you make the product and service better.

Mike Murchison [1:09:58] And so that will be enabled by the collaboration that we facilitate. And it's something that we've thought of very extensively since the very early days of Ada. I know we're in a very strong position to do that.

Matt Turck [1:10:19] I love this. That's such a wild concept. I guess, in a way, that's what we've all been talking about and investing in for many years, is kind of like the fully automated enterprise, which has inside the feedback loop and data network effects and all those things. But it's wild that it seems to be closer than ever now.

Mike Murchison [1:10:56] It is, totally. And you can see it. We have some software development agents deployed inside Ada that are shipping basic paper-cut features, we call them right now, or paper-cut improvements. But if you look at what the Jira tickets are associated with those features that are being autonomously deployed, they're really not very different from a summarized customer service issue. And so if you squint, you can kind of see this at play already a little bit. And so I agree, it's not that far away.

Matt Turck [1:11:05] That feels like a wonderful place to leave it. Mike, CEO of Ada, thank you so much for doing this. This was great.

Mike Murchison [1:11:06] Thanks, Matt. Thanks for having me.

Matt Turck [1:11:27] Hi, it's Matt Turck again. Thanks for listening to this episode of The MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing, if you haven't already, or leaving a positive review or comment on whichever platform you're watching or listening to this episode from. This really helps us build the podcast and get great guests. Thanks, and see you at the next episode.

Mike Murchison [1:11:27] Bye.