Alex Levin, Co-Founder & CEO at Regal AI. Regal is the AI Agent Platform.

Generative AI Agents are transforming customer expectations and the types of customer experiences business can build. The biggest opportunity is in support, sales and operations calls at consumer businesses.

Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.

In this episode of the AI Minds Podcast, Alex Levin, Co-founder and CEO of Regal AI, unpacks how AI is reshaping the future of customer engagement in contact centers.

Alex shares his journey from marketing and product roles to building Regal AI, and how voice AI is transforming contact centers by blending automation with personalization.

He explores how AI agents now rival human reps in handling complex interactions, while seamlessly integrating with existing systems to drive higher conversion and lower costs.

The conversation dives into the art of conversational design, highlighting how Regal AI is pushing the boundaries of natural, intelligent voice interactions at scale.

Listeners will get a front-row seat to the evolving role of voice AI in enterprise CX—and how companies can deploy bots that truly sound and feel human.

Show Notes:

00:00 Rethinking Contact Center Strategy

04:14 AI Revolution in Contact Centers

10:11 Custom AI Agents for Enterprises

12:40 AI Agent Flexibility vs. Cost

14:53 Customer Notification Guidelines Uncertain

18:44 Enhancing Human-Like Conversation Models

20:59 Exploring Text vs. Voice LLM Challenges

More Quotes from Alex:

Demetrios:

Welcome back to the AI Minds Podcast. This is a podcast where we explore the companies of tomorrow being built AI First. I am your host Demetrios. And this episode, like every episode, is brought to you by Deepgram. The number one speech to text and text to speech API on the Internet today, Trusted by the world's top conversational AI leaders, startups and enterprises like Spotify, Twilio, NASA and Citibank. In this episode we're joined by Alex, the co founder and CEO of Regal AI. How you doing today?

Alex Levin:

I'm great, thank you for having me.

Demetrios:

So I want to dive right into the inspiration behind Regal AI. I know that support is one of these use cases that has come to to be at the forefront of AI and I'm wondering what made you get into this.

Alex Levin:

So my background is actually more as a marketer and product manager, but the last company I was somehow put in charge of the contact center and helped grow it from 20 seats to 3,000 seats. So I've seen sort of every scale of a contact center. I've been on every contact center telephony platform that's out there and really seeing all these different tools and you know, beyond just the obvious that the contact center has been underinvested in, I think the biggest shock to me was that everybody was telling me, don't let customers call you. That was sort of unilaterally the lesson. Hide the phone number, deflect to other channels, don't do voice, largely because it's the most expensive channel, but also just it's quite operationally complex to make sure you can do voice well. And as I got into it, sort of taking more of a analytical approach to it, I'd see that if customers were well handled on voice for support, they had a better lifetime value. And I'd see if we wanted to reach a customer, you send them an email, 2% would click on it, you call them, 30% would answer.

Alex Levin:

And I'd go, wait a second guys, why are you telling me that voice and calls are bad? It looks like this is our best channel. And so really sort of struggled with that for a long time and struggled with trying to get our current providers. So we were, probably a 30, $40 million a year account for our contact center software. I won't say who. And I'd go talk with their execs and say, let like help us. let's make this a first class experience on Voice and let's do for the contact center what let's say a braze or iterable have done for email marketing and make it more personalized and you know something you can a B test and they'd tell me, no Alex, you're crazy. Voice is dying. Like you're being silly.

Alex Levin:

I left and started building software for contact centers. And this was about five years ago. Honestly, AI agents just were not good enough at that point. So a lot of what we built at the beginning was the infrastructure around how do you use the right data in the call? What should you be saying? How do you do AB testing, how do you know who calls should be routed, to both for inbound and for outbound. And started with human, started working with our customers, with humans. And it wasn't until about two years ago that really we were able to build AI agents. Where my co founder and I went that like that is now at a place where we don't, the weakest link is the human agent, honestly. And everything that we're trying to do, we try to do really complicated things to be super smart about what we're doing to the customer.

Alex Levin:

And that human agent gets on, just doesn't do any of it right. So we said this is like very interesting. And so not only was interesting for us, but with customers, I think what changes? Instead of having to sell a two year sales cycle into an enterprise customer, we were integrating with Nice or Genesis or 5:9 or whoever. This was open, green space and they go, this is interesting, every year I lose 40% of my agents. So why not try this instead of, hiring human agents. And we don't need to replace a current software. It's easy to integrate with my current contact center software. This is really something I want to try.

Alex Levin:

So I think part of why we're lucky is that we come from a background running big contact centers, really understand the space a little bit better than some in this industry. And then we spent years building a lot of infrastructure beyond just a voice AI agent. So a lot of the companies in this space have a voice AI agent which is not that valuable. You need to actually in production have the ability to orchestrate outbound, orchestrate inbound, do many different channels, take actions, obviously, do post call analysis, things like that. So I think we had a leg up because of the history and now the shift has moved really fast. So obviously Maybe in the last 12 months, what's changed is AI agents or we are able to get our AI agents performing as well or better than human agents. And that's just opened the floodgates where every company now wants to try and see exactly how this is going to work for them. And I think it's kind of fait accompli that whether you think it's two years from now or five years from now, 90% of these voice interactions will be handled by AI agents as long as it's unlicensed and, there's not a legal requirement that it's a human.

Demetrios:

Did you just drop some French on me right there?

Alex Levin:

I'm French, actually, I'm not American.

Demetrios:

That accent sounded nice. I was you're pretty good at French. All right, so can you double click on this idea of how the voice agent is a small piece of the greater system?

Alex Levin:

I think people came at this problem and here's the string. There's speech to text, there's an LLM, there's text to speech. Maybe there's some rag and some latency things, but that's the whole problem. Actually that's the easiest part of problem. Like if you spend time in the contact center, a lot of the challenges are beyond just the AI agent itself. It's how do you, let's say I go back to the days, pre contact centers, pre Internet.

Alex Levin:

There's a small town insurance salesman who's really great at selling life insurance. What do they do? Well, they probably go to the local Rotary Club meetings. They get to know you, they come to your house, they see you have a kid. So they're going to come over and say, life insurance is really important. This is one you should have. I see your house. Based on that, I think this is what you need, so on and so forth. And it's a highly personalized, maybe inefficient experience, but highly personalized.

Alex Levin:

And any local business that wasn't highly personalized went out of business. And the interesting thing is as everything went online and scale went up, everyone assumed at huge scale we have to treat everybody the same. So if you call any online life insurance company, they read the same script to you that they read to a million other people. And they don't even care that you're a different human being and need something different because it's not in their mental model. Their mental Model is, high scale, very efficient, not personalized. And I think the really fascinating thing about AI is it brings back the possibility that you could be at huge scale, but very personalized because that AI agent can look at everything that we've done with our customer and say, I see you spoke with Bob and you're in the city where the Kansas City Chiefs are really loved and the weather is great there and actually based on the thing that's happening today, I'm going to fix another problem for you you didn't even know about. But I see it in our system and human agents really struggle to aggregate all of that. But AI agents do very well.

Alex Levin:

But that's only possible if you have a much wider platform than just this thing that, like I said, takes speech to text, puts it in LM and puts it back to speech.

Demetrios:

Well, context is key here.

Alex Levin:

Right.

Demetrios:

And you want to get all that data and information on the customer. I see your market brain shining through with that idea of let's segment out each customer so that when they call us, they do have that personalized and customized experience. How do you balance this ability to acquire all of this data versus now we need to plug into every single system that the company has.

Alex Levin:

So it's not like day one we go and do all this work. Like when I was running contact centers, I'd always be going to the IT lead and saying, build me a button. That was the ask. And it's like humans build me a button that does X things so that my agents can do it instead of having to send it somewhere else or manually do it in the back end. Now it's a different ask. It's please it build me an API, externalized API endpoint so the AI agent can take an action. Because, let's be honest, the RPA is not good enough to be actually doing something.

Alex Levin:

You really do need an app. And so what we're typically doing with customers is working through what are the best examples that you want to start with, where, it's very high volume, so it's important for the business. It's relatively simple, so the agent will handle it. Well, there's not that many integrations, the API endpoints. So either they exist already or you're not going to have to make that many. And so that's a great place to start and we don't have to start 100% of that use case, let's start at 1%. And it's something that we do very well, is setting up an A B test to have even multiple versions of agents and then different versions of their human. And they can actually see, let's say across the main metric.

Alex Levin:

So average time on call, csat, the outcomes and perhaps containment, what how do you compare across those different variations? So you can keep deciding, what is the right mix for your company between AI agents and human agents. And then, as it works, you scale up the AI agents and then you go to the next use case. But it is very much use case by use case. I think there's many companies in the space that either are sort of mid market vertical specific where they say this is Bob, Bob is the insurance lead generation agent. You all use Bob. It's exactly the same. And then there's some that basically go in and say, it's not industry specific, but this is an agent that does a specific task. We take a different approach.

Alex Levin:

We go and work with very large enterprises. So Fortune 500 and are going and trying to build custom AI agents for their business and their end user. Which means that we are not building Bob, for everybody, every time it is a different agent and it's not necessarily the same use case for every customer because for different, even within insurance, which is a big market for us, for each company they have different models and there are different use cases that are going to be the most important. And so for one customer it may be really important to start with support. For another, it may be important to start with lead generation. Another one an operations call or a claims call. So depending on the customer, we end up building very different agents first for them. Eventually again all of it will be AI agents.

Alex Levin:

It sort of starts to be the fact that, all these use cases will be done, but the very first one is different across customers.

Demetrios:

Can we get into the weeds a little bit on these metrics? Because I would love to know how you are counterbalancing false positives in a way. I imagine that when you're looking at time on call, the lower the better. But there might be people that get on a call, they realize it's an AI agent or it's not human and they hang up. So how are you weeding out those types of metrics and making sure that what you're getting is a realistic view of what's actually happening?

Alex Levin:

So it's interesting when you run human contact centers, average time on call and after call work. So totally that's called the average handle. Time is one of the key metrics you're constantly looking at. It's not as important with AI agents. And so for two reasons. One, when you first put an AI agent on to do exactly the same use case as a human, tends to be about 30% less time. It's missing some of the filler. It doesn't sort of go around as much for a variety of reasons.

Alex Levin:

Just on average it's about 30% less time. And at first people go, that's great. Now go, cool your jets. Like not necessarily, let's talk about the difference between your newest agents and your most experienced agents. And they go, what do you mean? I go, what do your newest agents do on the call? And they go, they read the script and they follow exactly the thing. And okay, what are your most experienced agents do? And they go, they don't read the script, they don't follow it, they know where they need to get to and they go around more. And I go, so which one would you prefer on the call? They go, my more experienced agents, the outcomes are better. Like, it's the same with AI agents.

Alex Levin:

If you limit the AI agent to operate like your newest agents, the performance is going to be like your newest agents. If you allow the agents to have more flexibility and have the conversation be wider raising, perhaps the average handle time will be longer, but the outcomes will be better. And because, AI agents are, let's say at 10 to 20 cents per minute versus, an offshore agent, if you really calculate per minute of talk time is maybe 30 cents per minute and a US based agent is a dollar a minute, the savings are so massive already and they're only going to become more as the cost of AI comes down. Doesn't matter if you're saving 10% on average handle time. So I really dislike the press around sort of AI agents being sold on outcomes. I think it's a terrible idea because all that happens is it incentivizes the company building AI agents to you to build the shortest possible conversation and for them to lower their cost and not pass any of the savings on to you. what we do is we do it on a permanent basis. And so what that means is we're completely aligned with the customer in that as we bring prices down, they get the savings we're not like keeping them for ourselves.

Alex Levin:

And if it's better for them to have a slightly longer call or slightly shorter call, that's fine. And often we're not optimizing for that. We're optimizing is the CSAT where we want it to be is the first call resolution on a support sort of ticket or where we want it to be is the qualification rate, the conversion rate on an outbound lead generation call where we want it to be. And then you were asking specifically about containment. We have all this post call analysis again because we've been in this world for a while, so very easy for us to tell on a call when somebody's asking for a human, and trying to get off so on. when I used to do this five years ago in the contact center world, the legacy providers used to basically say at the Most we'll do 20 or 30% containment. And they're still, in the market today saying 20 or 30% containment. Our AI agents at Regal do about 97% containment.

Alex Levin:

So about 3% of the time somebody's saying, give me a human, I want to talk to humans. So it happens, but it's quite rare actually. it's striking how well these agents perform already.

Demetrios:

Do you have the agents identify themselves?

Alex Levin:

About half of customers do, half don't. There were, there was a notice of proposed rulemaking last summer by the FCC where they suggested this was going to become law, that you'd have to. And under the current presidential administration, zero chance that becomes law. So it's a bit the wild wild west. The way I sort of explain this to people is I say, it's up to you now whether you need to, but just make sure you do it in the right way. Don't get on the phone and say, this is Alex, the I agent. Like don't say that, like get on the phone and say, this is Alex, how are you? When the person responds and say, I'm from Bank of America, here to help you find the local atm, and by the way, I'm an automated agent and that is perfectly fine. The same way today you might say this is a recorded line.

Alex Levin:

That's how you do it. You don't say this is a recorded line like before the person has a chance to speak. So that's one piece and the second piece is for outbound calls, even though it's not law. Yet we have all of our clients add to their TCPA consent the right to use artificial voices to make sure that they're getting that permission so that at some point in the future when it becomes law that you have to have TCPA consent for artificial voices that they already have that you can't retroactively go and get it.

Demetrios:

Now you bring up something fascinating there on how to tactfully weave into conversations different threads or themes. This conversational design is something that I've been harping on quite a bit because it is now so open ended and we can do so much with agents. Have you seen specific ways to increase the agents actions? Actions might not be the best word the and it's not accuracy the word that I'm looking for. Just the making the agents better and getting that final outcome that you're looking for through certain tricks with conversational design. And I'll give you an example. We had Elliot on here a few weeks ago and he was talking about how making the agent a little bit quote unquote dumber has humans speak to it slower and enunciate their words more which gives it a better chance of success.

Alex Levin:

So we honestly we don't do that. We actually it drives my co founder crazy. I run the go to market org and she runs the engineering Org but I say to our customers like you need to treat this as if it's a human and we've and it's on us to perform as if you're talking to a human. Like we don't want to create a situation in which somebody's talking to this like 1, 2, 3, 4. you don't want that. You want them to be speaking naturally this right. So I'm on the other side of that specific one. But I say the biggest thing that people don't realize is how widely you can push the personalities of these agents.

Alex Levin:

So if you go to regal AI slash dogs, we did something fun where we created these vastly divergent personalities for different dog breeds and you can go and talk to them and and people are shocked and they go oh my goodness. Like that one was like cynical and whatever. And this one was like funny and cracking jokes and this one like love bicycle, like whatever. So it's actually I think one of the more valuable tools that you can imbue these agents with true, with their own context and personality and history that they bring to the conversation. And it's quite valuable. Of course the voices and the accents you use are part of that as well. And the Ability to have sort of words like and like and things like that more in the voices is becoming something that's very interesting. And the other one that people are trying to do is create mirroring in the tonality.

Alex Levin:

So right now most of these models the way they work, you can set the tonality, but it doesn't mirror what the human is doing. And I think we're getting close to a place where you can mirror what the human is doing. So all those pieces make it feel, I think a little bit more lifelike. So same with like background noise, you add a little bit of background noise and it does feel much more lifelike. I think like in terms of the actual conversation design, it's one of the things that I think we bring to this. Like when we work with enterprise clients, it's not in a state where we do have a self serve product, but the industry is not in a state where we just give them a product and we say bye. I take about, half of the service area a customer gets from us is we give them a four deployed engineer who's helping them do this and is working through this stuff and saying, well guys, let's maybe push in this direction and it'll help you get to the better outcome. Because we've done this enough times that we know how to handle objections, we know how to handle awkward pauses, we know how to handle the openings to make sure it's going to result in a slightly better outcome.

Alex Levin:

And look, In a year, two years maybe, it'll be instead of 50, it'll be 10% professional services and 90% self serve. But we're just not, it's not where the industry is today.

Demetrios:

You get to recommend if it is a golden retriever versus a pit bull. Basically.

Alex Levin:

That's why we made those breeds to show off the wide ranging personalities.

Demetrios:

Now before we go, I would love to know how you deal with the differences in cadence of the end user because I find that to be incredibly difficult. I, for example, take long pauses while I'm talking in the middle of my thoughts. Other people speak really fast and it's very hard for an AI agent, a voice agent, to pick up on each individual pattern and speaking flavors. So how do you go about that, especially when you deploy across these gigantic call centers and it's a lot of scale and you're trying to keep that customization like you were talking about in the beginning.

Alex Levin:

So it's your point and same point on mirroring today. It doesn't do it well, because today all that's happening is right, we use Deepgram was great, but like obviously we're translating the voice into text and the text is being fed into some LLM. The two strains of work that we're going through and a lot of people are going through are one, if you're going to use this sort of text based LLM, which is probably always going to be cheaper than a voice based LLM, a true voice LLM, you need to have a secondary sort of analysis of the audio stream happening at the same time for tonality, for pauses, for all these things that is then constantly being fed into the model to change via text, to change what the LLM is generating and how your agent is going to operate. And the other obviously direction is true voice alums where the audio stream is being put directly into the LLM and voice is being given back without any text being generated. Where those can do mirroring, can do things. They're not from a quality like they're not there yet and the cost is too high and they're not stable enough. And so we don't use the voice LLMs in production, but they're getting better for sure.

Alex Levin:

I think always the voice alums will be 5x the cost of the text based elements. So depending on the use case, different one, different customers will use the different strategy. But I think we'll get there. We're not far from a world where we'll be able to handle the mirroring, we'll be able to handle the pauses and incorporate that into how the agent is acting.