The Recursive CX Loop

Julia Nimchinski:

And next up, we welcome Brian Peterson, co-founder and CTO of Dialpad. Brian, welcome back. How have you been?

Brian Peterson:

Hey, good. Thanks, Julie. I’m all good. How are you?

Julia Nimchinski:

Awesome, super excited to dig into this. We just had you on the summit, speaking to Recursive Self-improving everything as the major trend, so just, you know, we’re obviously in the early innings, and Excited to get into this.

Brian Peterson:

Yeah, for sure. Let me get my… Okay, we ready.

Julia Nimchinski:

Yep.

Brian Peterson:

Okay. Alright. I’ll take you all through this journey, on where I think AI is going. Just to give… some quick background… I’m gonna get used to using Zoom. We’re good, we can see this.

Julia Nimchinski:

Yep.

Brian Peterson:

Okay. So Our background, so… We were at Dialpad, we were the former Google Voice team. So we built, Google Voice to 10,000,000 plus users, and we thought, hey. This makes a lot of sense for businesses, and, and we saw that, obviously, for anything customer communications, that AI was gonna be a big part of that when we started. So we’ve been doing AI now, with Dialpad for about eight years.

So we have a lot of experience with AI, a lot of patents and papers submitted to AI, AI, conferences, so… A lot of history with this, but what we are trying to build at Dialpad, and the dream is is one communication platform that can do everything. So both AI and human agents together, we feel like this right now is broken. There’s a lot of disjointed systems, which I’ll go into a little bit here.

But we are full omnichannel. We call ourselves the AI platform for customer experience. Trying to basically solve the problem that the end customers had forever, which is terrible experiences, and wait times and everything with either AI agents or human agents or both. So… But I’m going to cover a little bit about not just Dialpad, but where I think AI is going.

So this does apply to Dialpad, but this actually will apply to anything you’re using in AI. So, when we talk about Recursive AI, I think a lot of people think we’ve seen what AI can do, but we’re not fully there yet into what’s going to happen. So, this is a little bit of a prediction as well. So, what happened? So, first thing. ChatGPT. I feel I call it, like, the before and after ChatGPT era.

There was… sort of, like, whatever happened before ChatGPT with AI, and then what happened afterwards, and it just changed everyone’s world. These large language models really changed everything. And then it was great, it was a really good, you know, oh my god, magic, it can write really well, it can summarize things. It could be probably an okay personal therapist for you.

But it that was you know, it it had its moment, and it was great, and it enabled all these other things. But that was kinda like, okay. Well, what’s next? Well, then all of a sudden and things are moving so fast. What was next? Well, everyone started talking about Agentic. Oh, everything’s Agentic AI. It’s… and Agentic AI is basically, can it take action?

Can it automate real workflows in your business, and everyone starts talking about Agentic, and Agentic’s the thing, and everyone’s talking about it, everyone has it on their website, just like they have AI on their website after… after ChatGPT comes out. But I’m gonna basically argue for, while a lot of people think this is the end state, that we are just beginning, and there’s something missing with all of this AI.

And Agentic sounds magical, but it’s really, really difficult. The LLMs and ChatGPTs enabled this. But it didn’t actually… make it… workable in real life for your business, especially not consistently. So… Are we done? We are not done.

And this, again, I’m going to talk about Dialpad a little bit, but this is also just, if you’re doing anything in AI. this is going to apply to anything, not just using, you know, our customer experience platform, but if you’re using AI for internal automation, or you’re using AI to do anything with your business, this is going to apply to you, too.

So… Agentic AI is not the end state, and I think this is something I talked about, at the AI4 conference, and it got a lot of attention because everyone was talking about Agentic, Agentic, Agentic, and I was like, wait, Agentic doesn’t really work yet. I was like, why is everyone forcing this on us? It doesn’t… it’s super fragile. When you put it into production, it doesn’t work consistently.

It’s really hard to maintain. And this is why most AI initiatives stall out. Like, this is… this is something that, again, if you just look at, like, what they do today, yeah, they can execute some things, but they do make the same mistakes over and over and over again. And they only improve when someone goes in and configures it and changes it, or re-prompts it. they’re not… for being AI, it’s not very smart.

Like, it’s really good when you automate the one thing, and it works consistently, maybe for a little bit, until your process changes, or the way your customers interact with you change. But it… it’s not smart, which is… kind of ironic that this is, like, the latest AI, and it doesn’t get better on its own. So… and then how did we get here? Like, why… why isn’t it getting better on its own?

It’s because a lot of this stuff is built on fragmented systems. and you all experience this in your businesses, things are all over the place. There is, you know. disconnected context, there’s… it’s… it’s coming from different systems, you maybe have a CRM, you have something else, you have internal databases, data lakes, it’s… it’s just kind of all over the place, it’s really hard to keep track of.

It doesn’t work even well with your… your databases and your CRMs. So… This is… sort of the reality of today, and why… why it’s really hard to make Agentic AI work, but also why it doesn’t get any better when you even make it… some of it work, is because we’re still dealing with things being all over the place, and AI is really good at connecting to these things, don’t get me wrong.

Like, it is really good at finally connecting to these things. But they don’t… it doesn’t do a good job of taking all this stuff and being able to really process it and get better over time. It’s really good at the, like, point A to point B of, I need this, then I’m gonna do this, then I’m gonna read, you know, the basic information about the customer and do something. But it doesn’t actually get smarter over time.

The context window is really just the current basic stuff you know about the thing you’re trying to solve, the workflow, the customer, whatever it is. But it’s it’s all over the place. They can’t get smarter because it doesn’t know enough, and it’s it’s not controlling all of your, all of your interactions. So… Every customer conversation, though, this is where we need to go.

This is where we put a lot of investment in Dialpad, and this was our vision from day one, is why does… like, I’ll just talk… we deal with sales and customer support heavily. I’ll talk mostly about the customer support side of things. you… we all still… we have all this amazing AI in Agentic, and customer support still sucks. Like, really sucks.

It’s actually even more frustrating now, because there’s so much obsession over deflection, deflection, deflection, because the Agentic AI customers… companies just want to deflect to get their money and avoid talking to a human, but that makes the customer experience way worse, and no one’s really looking into what is the frustrations with the customer?

If you look at a lot of the deflection stuff out there, or the fancier term they’ll call it is, like, containment, which still sucks, is… they will… say, oh, we handled this for you. But then when you actually look at the conversation, and you understand what the person’s intent was, most of the time it didn’t actually solve their problem. It just, like, pissed them off and got them to give up.

And that’s not good, obviously, for our businesses. That’s not good for us who need our customers to love us. And so none of these systems are trying even to go and analyze those conversations much. They’re starting to try to figure it out, but it’s obviously really difficult. even traditionally, right, if you look at a contact center, you’re only reviewing, yeah, like, about 2% to 5% of any interaction.

Like, you can’t get real data from that, because it was very manual. And even the new Agentic systems are not putting a lot of effort into having the full customer journey. They don’t even control the full customer journey. And so they’re not they’re not able to put this all together, and they’re kind of just ignoring it and throwing most of it away. So… This fragmentation breaks down all of these decisions.

So it’s… even if you had this amazing AI and Agentic AI, and there’s some really good Agentic AI-only companies out there, they don’t have access or the ability to learn from the full loop.

And so, in our case, the full loop is everything about your customer, it’s every interaction. we believe that we’re one of the only ones in the world, or maybe the only ones in the world, who can own the whole customer journey, and if you can’t control the whole customer journey, you can’t make it a better experience, and… For example, everyone has experienced the same problem, and it’s still happening today, and you all know about this.

If you go to some company’s website. They got a chatbot, you start talking with the chatbot, it’s, like, pretty smart, you know, you tell it who you are, you maybe authenticate a little bit. And then… You know, it can’t solve your problem after 20, 30 minutes of back and forth. it’s… it needs someone to help with. It says, hey, you should talk to a human representative.

It then has you call in to the… to the number, or maybe even takes you to a human representative in the chat. The second you get to talk to a human, what do they do? Who are you? Let me verify you again. What are you calling in about? It’s like, I just had a 20-minute conversation with your bot. how do you not know my problem already? How come I have to repeat myself?

And this is kind of the problem where you… if you don’t control this whole end-to-end loop, you can’t have these amazing, seamless experiences. So, and that’s… and this is why… this is according to not just us, but this is a source from MIT. 90% of AI pilots fail right now. This is, again, not just customer success stuff, not just… customer experience, not just customer support and sales.

This is, like, anything you do in your business with AI is not really working. There’s a huge promise of, like, this is magical, and then when you actually try to put it in your business. it’s not good at doing things automatically on its own yet.

It’s not consistent enough, and that’s probably one of the biggest things I’ll call out, is ChatGPT seems magical when you’re interacting with it, but in a business use case where you need to use ChatGPT to do step 1, 2, 3, 4, 5, It’s… So all five steps have to happen in order exactly every time to be useful in your business. And with LLMs, they’re just not consistent enough. Right? They’ll hallucinate.

They’ll give you a slightly different answer every time. What seems magical is is really, really hard. But yeah, a lot of this stuff, though, is just not based on, like. It’s not turning into any ROI. And a lot of people are focused on, like, oh, it’s not even… they’re just focused on cost savings. But, like, ROI isn’t just on cost savings.

ROI is supposed to be on business revenue improvements, if you want to go very specific. So, if I reject that customer. and they don’t go to a human, you’re like, oh, great, I saved the average, you know, human interaction cost between $10 to $15, on a call center. Oh, great, I saved $10 on this customer interaction.

But you just pissed off one of your most important customers, and maybe they would have bought more from you. Maybe they wouldn’t have churned if you actually gave them a really, really good experience. So… how do we get to that really amazing experience? Well, one, you need to own the entire customer journey, so when it goes from AI to human, the human immediately knows what the AI talked about.

It not only knows what the AI just talked about, it knows what your humans, or your AI, or anything about your customer that they interacted with them from even months ago. Because if I call in two months later, I don’t want to be, like… I want the human to… or the AI to immediately know, like, who I am. what I’ve called in about before so that we don’t repeat the same problem over and over again.

I’m frustrated with this, and I go through the same, did you try this? Did you try this? When you should already know I already tried this stuff. I already did this. I did it months ago. Why aren’t you smarter?

So One, you have to control the entire customer journey, but the next phase of AI, and this… this chart is not just a chart for Dialpad and what we provide, but it’s a chart for anything you’re going to do in AI, not Dialpad related to. This is where AI is going. some pieces are far away from this.

I think our customer experience side is much closer, and we have a lot of these pieces already, but every single thing with this Agentic AI needs to first understand what is going on. So it needs to actually analyze the conversation. Is it working? Is the customer happy?

Then, you have to take that orchestration, make it happen, you have to resolve their problem, and then the biggest step that people don’t… like, so resolve their problem, not actually just deflect, contain, like, actually resolve their problem. And then you need to take all of that understanding, orchestration, and what worked and didn’t work, and you need to optimize.

And it needs to be this continual self… learning loop, which… that’s where AI is going across all businesses. We’ve seen this coming, we’ve been building this, is that when this stuff happens. it should automatically get better over time. You shouldn’t have to adjust your Agentic workflows and mess with, you know, some sort of complicated workflow system and drag-and-drop thing just because Something broke.

You should already know if it’s breaking. The system should tell you. The system should be automatically fixing it, or at least suggesting what you should do differently to improve not just deflection, but actual customer satisfaction. So… It needs to start with, okay, Discover. What… where can AI create impact without introducing risk? How can we maximize capability without losing control?

So… building it. what do you build? How do you know what to build? So Discover is… and I’ll tell you a little bit of how we’re doing this, but, like, the biggest problem with AI is you go to your AI bot. And, you know, the typical, I think, AI bot is like, what would you like to do today? It’s like, I don’t know, you need to tell me. Like, why… I don’t… and this happens on anyone, so for our existing customers.

They’ll do things like, okay, I have millions of calls coming through my system in a year, or millions of customer interactions over text, website, whatever it is. I don’t even know what to automate. So, like, this is great, this sounds great. Oh, Agentic AI, like, awesome. What do I automate? I don’t know… what… where to begin.

I don’t know what we do repetitively, because they don’t have access to the millions and millions of minutes or, you know, or text responses and conversations to know what even to build. So, they need to be able to discover what they can automate. Then, when they build it, they need to be able to build it easily, not something that’s going to take them weeks, years, and require a bunch of fancy, crazy interface stuff.

After they’ve built it, the next hardest problem is you have all these Agentic AI agents, and they’re fine for your personal life because it’s not a lot of risk there. But when you put them into your business. they… you can’t just throw this Agentic AI over the wall to your business, your customers, and just cross your fingers. You can’t do that. Like, it could do some serious damage.

It could write the wrong bill, it could, you know, cancel the wrong subscription, it could do so much damage. And even if it’s not a lot of damage, maybe it just isn’t a great experience. You don’t want to just test that out with your real, important customers. Gil Preuss, By just, let me see if this works.

Gil Preuss, And a lot of this is happening where you’re doing trial by fire on this stuff, so we have to build in amazing, what is called, you know, evaluations.

You need to build this into the system, and the system has to be able to simulate all of these conversations, right, and know that, like, hey, I think, you know, we’ve guessed or estimated that this change is going to have a, you know, much better improvement because we’ve analyzed all previous conversations, we’ve tested this with all these simulations, it’s going to work, we feel comfortable about it.

And then, you probably still don’t want to throw it out there 100%. You probably want to throw it out there, let’s start small, 5%, 10%, 15%, and be able to get insight on is it working, so that you’re comfortable to go full 100%. And then once it’s out there. This is still ChatGPT’s version. I mean, not everyone’s using… it’s basically the LLM stuff, right? It hallucinates.

It will do things, it’ll answer the same question differently. one time out of 10. You need to have something that’s gonna have guardrails around it, so that it can know when it’s going outside of the bounds of what it’s supposed to do, and shut it down. Or alert your team to say, hey, this is not doing the way… what it should be doing.

We need to shut it down, or change it, or automatically… we’ll automatically change it for you, but we need to let you know that this did this thing. And if anyone tells you that theirs is foolproof, they’re… they’re definitely lying, because this is so, so difficult, and if they can do it.

They’re only automating, like, 1%, because that’s the only percentage they can do that is very deterministic, very direct, where they can for sure make sure it does exactly this. But if they’re trying to automate. 80% of your business, or whatever your complicated workflows, you’re gonna have to worry about the guardrail side of the government’s governance side.

So these are the things, some of the examples of what we do, so… and I will tell you this a little bit. This is, again, this is a Dialpad thing, but I will tell you, you’re going to see some version of these features in a lot of different AI Agentic systems out there. We’ve spent a lot of effort into building these specifically for customer experience, and… and fine-tuning them and everything, but… Find it.

I told you the biggest problem is discovery. Our customers don’t know what to automate. Well, we built something called Skill Mining, which automatically analyzes all of their human conversations to find out exactly what they do repetitively.

So, like, you do this, this, this, and this on 10% of your calls, on 5%, you do this, this, this, and this, and it breaks down exactly what they can automate, and what they should automate, because it’s super repetitive, and you save the most important conversations for the humans, and not do the repetitive things with humans.

Then you need to build it, and it can’t be some super complicated way to build it, because We’re not all, like, data scientists and AI engineers, like, it needs to be something that’s super easy. Well, we take the skills mining and let you automatically create. skills. for those things. So you just press the button, and it’ll automagically go and create the workflows you need, and then do the evaluations on them.

So, it will test it. We have something that’s very proprietary that we’ve built called Proving Ground. How do you know this is gonna work? And you know what most people do to figure out if it’s gonna work? They have humans test it. And what do you think the human’s gonna do? The human’s not gonna do, yeah, my appointment’s for the fourteenth. No. I think it’s actually no. I need the fourth.

They’re going to say it in a very… And it’s obviously going to be limited in volume, too, so they’re going to test it, say, okay, yeah, it looks good, and then throw it over the wall, and it’s not going to work at volume with real customer conversations. So we knew we had to build something.

This evaluates your changes before they go live on every single interaction In past… As well as, like, in real simulations for what’s gonna happen if it goes live. So… replays it on real human conversations, replays it on your previous AI. Agentic conversations, your AI agent conversations, and gives you a real, score on, like, we think this is gonna work or not gonna work, and here’s why.

And then, of course, once it’s out there, we have our own proprietary thing called Guardian. Which monitors in real time all of your AI agents to make sure they’re saying the right things, they’re doing the right things. And with the ability to automatically adjust if they’re not, and to alert the team and the stakeholders that this is an issue. So, when you’ve done all this.

What this looks like, like, when data, insight, and action are connected in this one operational system, the outcomes change dramatically. The cost obviously decreases to serve this stuff, but there’s more issues resolved on the, like, on the first interaction, rather than back-and-forth conversations over multiple channels and AI and humans.

Yeah, and then every single customer interaction becomes this source of intelligence that improves future interactions, which is how AI… that was, like, the dream of AI when it came out, was like, it’s supposed to be really smart, why isn’t it automatically getting better?

And there is some part of the AI that is, but then nothing was getting better on this side of things, so… This is just a general, like. the benefit you have with these things. So, In summary… And this is something that, again, will apply to you across anyone who’s talking about AI. Because everyone’s saying AI, and everyone has AI on their website, and Agentic on their website now.

It’s hard to know what’s… what’s real, what’s not. And a lot of them will just keep saying AI, AI, AI, AI. It’s like… it doesn’t really mean anything anymore, because everyone said they were cloud, cloud, cloud. Everyone should be AI, but the next frontier isn’t just who has the most AI, it’s going to be the companies that learn faster, and then turn that learning into real money.

And I urge you to not just think of, this is a cost-saving thing. Like, we need to grow businesses, like, we need to grow revenue. Growing revenue is way more valuable than saving a few dollars on your customer interactions, because if you have this amazing learning experience where the customer feels like. I can go from AI to human, back to AI, and it just knows me, and knows what to do, and it’s solving my problems.

I’m not on hold, I’m getting my problem solved, then it actually becomes the reality we started Dialpad to build, which is to make the end customer interaction better, so that they want to buy more from you, they want… they want to say how amazing you are. So… That is it. Thank you. Appreciate the very quick time, okay.

Julia Nimchinski:

Thank you so much, Brian. You can’t see the chat, but our folks just, you know, love your grounded, realistic approach to AI. Lots of questions here. The first question is, love your definition of problem resolution. And it obviously dovetails to the context problem. And so from your CTO lens, how much context is actually enough? In order for a problem to be solved in a high-quality way.

Brian Peterson:

I think… I think that the answer that’s gonna make me sound smarter would be, oh, it’s never enough. But the actual reality is it’s more than we have now, but it’s not like you need… every single everything about that customer for the history of time, either. There is a… there is a middle ground where you get enough that is going to be a 10 times better experience for the customer, or for whatever you’re automating.

So, for example, right now, just context is mostly in the present. It’s like, what do you know from, like, the last hour, and maybe you know something from a couple months ago, there is a little bit of, you don’t want to feed it too much context, so it has to be smart context. Like, I should have information about you from 2 years ago, but I don’t need to have everything about you from the last 2 years.

I need to have just the right amount so that it doesn’t go off the rails either. So it is a little bit of a mix, but the problem right now is no one really has historical context from multiple sources. I’d say, like, that’s kind of the biggest problem right now. And even, obviously, like I said in my talk, a lot of people are even just struggling for Context between agents.

Like, that’s gonna be another big frontier because it goes from one agent to another agent, and they don’t know anything about what just happened. You’re gonna see that, too. There’s a lot of people working on, like, agent to agent. context passing as well. But basically, it’s not there yet because of those disjointed systems. That’s probably the simplest way to…

Julia Nimchinski:

Makes sense. And, to your second point, what to automate. We just followed our community, and one of the biggest challenges, exactly as you mentioned, is the lack of strategy, and actually knowing, you know, the art of possible. What would be your, you know, like, obviously Dialpad solves it in a product form, but what would be your recommendation in terms of, you know, vision, strategy, management of AI?

Brian Peterson:

I think a lot of people are trying to boil the ocean and throw it and say everyone just start go using ChatGPT or whatever. It always fails when they do that. I think the most important thing to do is start small with a pilot team.

Work on the business problem, and how you’re gonna automate it, and start small, because even in your giant business, it’s going to be really hard to understand what you can and can’t automate, so you kind of have to just try things out. Try, like, try something you know of. that is actually something you think you can automate.

Start a small program, apply AI to it, see if you can automate it. report back with the findings, have them learn, you know, see what they’ve learned, and then start applying it to more and more places.

I think everyone thinks they have to, like, throw AI at everything, and it’s like, it just ends up being a disaster, in my experience, but there’s so much pressure on all these CTOs and CIOs to… and CEOs to, like, go, we’re all, like, AI forward, go use AI everywhere. It’s like, it won’t work that way. You can’t just throw You have to be calculated, so I would just say, please.

Like, you need to use AI, it’s super important, it will help your business, but don’t try to be an expert day one, like, try something, small pieces, one at a time.

Julia Nimchinski:

And last question, a visionary one. You showcase your loop and central methodology to all things Recursive AI and self-improvement. I’m curious. just your from your lens. Do you think that customer contacts or any siloed functional contact sales marketing you name it. put it, essentially. Re-establish the product, the value prop of the product, and even how the product looks, or its functionality.

Brian Peterson:

If you have all this context, is that what you said? Context?

Julia Nimchinski:

Yeah. If, say, you have just an Agentic conversation, your agents gets an intel that, say, your customer would like, you know, your product to reshape around a certain workflow, would that happen automatically?

Brian Peterson:

Yeah. And not only would it happen automatically and tell you what you can automate, like I said, it absolutely will be coaching the humans. I think a lot of people think that AI is gonna replace humans. I’m not a big believer in that. I think it’s a it’s a mutual partnership, and a lot of people are trying to just go all AI or go, we’re human and we don’t understand AI.

And I think the future is the AI is teaching the humans, and the humans are teaching the AI. And that’s like, coaching everything. So I… that is the auto-magical experience that you’re gonna see, is it’s never gonna be 100% automation. It’s gonna be, like, 50-50, 80-20, where they… they are constantly learning from each other, and that’s the biggest thing. And again, on all forms of AI, it’s just gonna be like this.

It’s not just a customer experience thing. That’s my hot take, you know.

Julia Nimchinski:

Thank you so much.

Brian Peterson:

Okay, thanks, Julia. Appreciate it.