The Autonomous Flywheel: Scaling Revenue Through AI-Native Systems

Julia Nimchinski:
Welcome to AI Practice Sessions. Today, you’ll hear from executives building autonomous revenue agents, and we’ll focus on what they are actually deploying, what’s working, what they’re learning, and where they think technology is heading. And welcome, Brian.
To kick things off, we’re excited to welcome Brian Peterson, co-founder and CTO of Dialpad. Brian will be showcasing the Autonomous flywheel. And can’t wait to get into it. How are you doing, Brian?
Brian Peterson:
Great. Doing great. Can’t wait.
Julia Nimchinski:
One second… we can’t hear you.
Brian Peterson:
Oh. Let’s see… Can you hear me at all? Okay. Hold on.
Julia Nimchinski:
Can you hear me? Yep.
Brian Peterson:
Okay, how about now? Nope.
Julia Nimchinski:
This is how you know that Brian is real. Can you hear me? I’m an agent. –
Brian Peterson:
Okay. How about now? No? It’s, it’s highlighted.
Julia Nimchinski:
About now?
Brian Peterson:
Can you hear me?
Julia Nimchinski:
Yeah.
Brian Peterson:
Okay. Awesome.
Julia Nimchinski:
Welcome to the show, and yeah, let’s get into it.
Brian Peterson:
Let me… Okay, can you… can you hear me now again?
Julia Nimchinski:
Perfectly, yeah.
Brian Peterson:
Okay. I couldn’t tell if it was me or you. Okay. All right. All set?
Julia Nimchinski:
Yep.
Brian Peterson:
Okay. I will share my screen if this… Second… Yep. Okay, good.
Julia Nimchinski:
Amazing, yeah.
Brian Peterson:
Thank you. Okay. Hi, everyone. I’m ready to start, right? We’re good? Okay. So, I’m co-founder, CTO of Dialpad, we do, a customer, AI customer experience platform for both sales and support conversations. So, we’ve been doing… AI on this for the last 8 years, and I’ve been doing communications for over 20.
So, been in the world of AI and agents, all in, so, I want to just give you an update on what this… what is it like today? Where are we at? How’s it evolving? What’s the future of, of sort of this Agentic sales look like? So… I’m gonna bet, and this is what we’re seeing already, is that a lot of people don’t know who their customers are.
They have CRM data, which is very basic. Maybe they have some Zoom info data, but maybe you have some call recordings, but you don’t… you don’t really know what’s going on with those.
But I think in this future of Agentic and these AI systems is… you… the data is where a lot of this value becomes, actually real, and really adjusts to how you sell and what you know about your customers. So, so, like, what does this actually mean?
So, What we’re dealing… like, the problem, like, right now is… the inherited architecture was never designed to learn, and I think that’s the biggest thing, like, to take away from this conversation, is that We have… a lot of our systems and data was built around just a human looking at it, and what’s gonna start happening is you’re gonna want to have the latest version of these Agentic AI, where it gets really, really good, is when it starts to learn, like, reinforcement learn on itself.
So… right now, like, everything is totally separate, and I think the future is gonna be these connected systems, where a lot of it comes down to context and shared context. So… Let me show you. So… This is… one of the things that’s happening right now is that everything is completely fragmented.
It is… you have data, but it… about your customers, about your prospects, but it’s all over the place. It’s… it is in completely different systems, they’re not connected, Some of it’s in the CRM, if it’s synced well, but most of the time it’s just basic customer information.
But, like, each one of these things is super important to understand what’s going on with your customers and prospects. So, the… This is something that you’re gonna start to see becomes a bigger deal, that you need to be able to connect all these dots together to actually understand and take it to the next level. And… This is where we’re at again.
This is… Every single customer conversation teaches us something. But we’re not looking at it. We’re just looking at basic stuff. So, it could be… you know, a product issue, it could be a buying signal, it could be frustration that they have, maybe it’s a coaching opportunity for your reps, even.
Every single… interaction, and all those different systems I just showed creates intelligence that you are not probably looking at. And these AI systems are not getting smarter over this data. They’re just repeating the same thing. They are using, basically, whatever data ChatGPT has, or any of the models you have from the internet.
It’s not really learning from these interactions that are very custom to your customers, to your business. So, on only, as you can see, only, like, a fraction of these conversations are actually, reviewed. So, And all this data is everywhere. It’s… it is… everywhere, but we’re not actually having our systems learn from it.
So, this raises, like, the obvious question of, like, why? So… This is what we’re seeing consistently. There is a bunch of fragmented stuff. There’s your CRM, as I said. There’s maybe your contact center for outbound, inbound, whatever it is, that are completely separate.
they try to sync the basic stuff, but they don’t really sync, you know, obviously full conversations. And then you have an AI. So most of the… most of the companies, including, I’m assuming, a lot of businesses out there internally, yourselves maybe, are taking AI and trying to see what you can do with it.
But it’s really hard to extract at scale, what’s coming out of your CRM, what’s coming out of your actual conversations, which is a lot of data, and how do you actually do something with that? And then even if you do figure out the right prompts. and you want to actually be able to get some information out of it, it never gets smarter.
It’s still just basically doing pattern matching. So. Based on what you give it. So, and then what happens is, like, the customer risk is recognized too late. So you don’t… you don’t get this data, maybe you parse some of it, you don’t really understand what’s going on with your customer until it’s too late.
And… This is why a lot of… you’ll hear this in the news everywhere, but this is why a lot of AI pilots fail. There’s a lot of people trying to figure out how to take this AI and make it useful. But it’s… it’s been bolted onto disconnected systems. Insights don’t connect to the actual execution. There’s no feedback loop to improve outcomes.
So pilots don’t scale, and, like, the ROI is… always unclear, almost always unclear. And this is a source from MIT, this isn’t just something I came up with, but this… This is a real problem that people are experiencing. I think everyone feels this right now.

So, What is the new frontier, which… we call the flywheel at Dialpad is… what does a real customer experience look like using the power of AI? Not just, can I parse things and summarize things and do it at small scale?
And what it looks like, and you’re gonna see this not just… not just with something like Dialpad, but anything you’re building in the future of Agentic AI is it’s going to be some form of this flywheel if it’s done right. And it’s really, really hard, but this is where the value gets unlocked.
These customer conversations Is where a lot of the, like, the richest operational data exists. But it is scattered, like I said, around, you know, across 5, sometimes 10-plus systems. And then it just doesn’t get smart.
It just continues to be sort of dumb AI with hallucinations and everything, and feeding in information that’s mixed with the entire world wide web of data and fake news. It’s just not… it’s just not designed for that. And so.
turning this into a system where you can take all of these interactions, every single interaction with your customer, and learn from it, the future is not going to be these giant. models that can do these things at once. It’s gonna be a bunch of smaller models, and it’s even gonna be custom memory and AI for each customer, and vertical.
So, imagine you can… someone calls in, and it knows them personally. It has to have the ability to contain, and knows what you talked about last time, and it adjusts based on what you talked about last time. This is where you’re gonna see things going. It is really difficult, but it’s this constant learn from humans.
AI gets better because of even human conversations, or your customers’ information, or conversations, and then the AI even will be training the humans, because I’ve been around this a lot. I’m a big believer that we’re not going to automate all conversations. I think we’re still less than 2% of all conversations are even automated right now.
maybe we’ll get to 80% someday, but there’s still going to be 20% or more of humans that need to have these conversations, but what is the AI doing to help them have better conversations? It’s going to be able to live coach those humans based on the data it knows about the customer as well, making the human conversations even better.
So, the overall opportunity, though, is to connect conversations directly to these this decision-making and actions, and even real time. This stuff can’t wait, you know, a week, months to process this amount of data. It should be able to react.
Like, when the call gets even transferred, to, you know, the AI to a human, which is a normal scenario I would expect to happen a lot still. Does the human have the context? Does the human know who this person already is? Can they have a… make it a better experience? So… Agentic systems, like.
need to have all this stuff that people, I think, don’t realize, because you’ve heard a lot of stories of… different businesses go out there and automate on their own their customer service, or do a bunch of just automation with something Agentic, something with ChatGPT, basically.
And then, a lot of times what you hear and what you don’t see, though, is that they don’t really work. And they don’t work because if you look at using something like just an LLM, like a ChatGPT, That’s only about 10%, I would estimate, of making something work reliably.
They have absolutely made it so that we can do this amazing stuff with automated workflows, but the LLM, the ChatGPT, is only about 10% of that. It is… it’s because of things like what you see here. The LLM is great for understanding humans, for summarizing things, for parsing out what they want to do. It is really bad at execution.
And you’ve seen this, if you’ve ever used any of these, which I’m sure you have, you can ask it the same question twice, and it’ll give you a different answer. That doesn’t work, like, that’s okay as your therapist, maybe? But it’s not… it’s not okay for a business that needs the same thing executed every single time, exactly.
And so, what you’re seeing is a lot of these things, and you’ll hear this term called harnesses, or harness, It’s becoming more popular, but it’s basically the system around these things that makes it reliable and actually work. So.
A lot of this stuff is… You have to have… you know, the ability to test it ahead of time, because you don’t want to throw these things into production, have them start talking to customers. Is it going to work? Are you going to be able to simulate that these things work against real human conversations?
That’s something that we put a lot of effort into. You need to make sure that all that is tested before it goes live. You need also the ability to govern it, so… you’ve all experienced the hallucinations. These systems can go outside the boundaries. When it goes out there, you can’t have it say something that you as a business don’t want it to say.
So you have to put all these guardrails in place to make sure it never says the thing it’s not supposed to say, and that it’s exactly acting like you want to act as a business, not like… ChatGPT responds with the same writing every time. Like, it needs to be custom to your thing, so… These are becoming very, very complicated systems.
They’re really hard to build. This is something I do not recommend, whether it’s for customer experience, sales, but even if for anything you want to automate in your business, I recommend buying someone who’s been living this, leveraging common frameworks, do not try to boil the ocean and do it yourself, because it is super difficult.
And this is… this is just, like, a little preview of the things that we built, to accomplish this system that works, that actually works at scale. And you will see things like this in other… other forms of Agentic, not just in customer experience and sales, but, like.
moving from this, like, experimental phase to a production phase is the biggest barrier for companies. They can do it in a little trial, and it works, and, you know, it seems like it solves, like, a few test cases, but then once they get it out into the open, and it has a bunch of different things that can happen to it, that’s where it falls apart.
So… Our system’s designed for full lifecycle. There’s all these pieces that… have nothing to do with just prompting, you know, an LLM, necessarily, that you need to have in place if you’re going to do anything that’s an automated workflow.
If you’re doing an outbound AI SDR, inbound SDR, inbound customer experience, anything you want to do, though, with Agentic inside your business, you’re going to need some form of these things. So, And the whole flywheel, we call it, of reinforcement learning is a big deal.
So, one of the advantages we’ve had is, in the first one you see, skill mining, is that we have… all these conversations that have been transcribed, that we have for your business. So our customers say, of course, they’re like, well, you have all this AI, you have this Agentic stuff, I don’t know what to automate. I don’t even know what do I do?
And so the first thing we do is we have a proprietary model that goes and analyzes all their human conversations, and then says, oh, we noticed that you do these things repetitively all the time. You do this, then this, then this, then this. And you do it on 10% of your calls, this one on 30% of your calls.
And these are all things that humans probably shouldn’t be doing repetitively if you can automate, but… finding out what you can do with AI is one of the hardest things to do, in the business. Internal, external, whatever it is you’re trying to do with AI is understanding, like, well, what do I do with it? Like, what can I do with it?
So, SkillsMind is one example. The ability to, like, visualize, so we have an Agent Studio, but it’s not just about the typical drag-and-drop. You know, workflows. that was the old way. The new way is… these workflows are much more simple, but at the same time, to make them work correctly, they’re more complex. So.
humans aren’t going to be good at building, necessarily, these workflows, so having AI automatically guide you to what the best way to set up the workflow is, based on the data it knows about you, is a huge thing.
The… Proving Ground, so one of the things we call Proving Ground is back to what I said, testing is one of the hardest things in this world right now with AI, and you can’t… if you’re gonna obviously put something out there that’s representing your business, you need to know if it works.
With Proving Ground, the whole concept is we take… you change something with your agent. your AI agent, you have no idea if it’s gonna work. You want to just throw it out there and see if it fails. That’s not good, obviously, for business.
So… whenever we’re doing Agentic, you need some way to do evaluations, and so you need some way to say, hey, when I make this change, is it gonna work? Without it being really in production.
So we built something that, you know, analyzes all conversations against your new agent, to see if it would work, and we use human conversations to even test against that. So, and then governance is a huge thing. You’ll hear in all of AI right now.
We have our own product called Guardian that’s constantly looking at live conversations to know if it’s going outside the bounds. Is it leaking, you know, private? identifiable information? Is it saying something that your business doesn’t want it to say?
I think you’ve all heard of maybe a story, or some of you, where an AI agent sold someone a truck for $1, and they had to give it to them for a dollar because It was a business chat, so this is something that’s real, and it’s something that is… a piece that most people aren’t thinking about as well. So… what does this look like?
Well… when you can actually connect everything, like, when you can get the data, the insight, and take action from it, and they’re all connected in one operational system, the outcomes, like, change dramatically. So, the cost to serve decreases. More issues are solved on your first interaction.
And all this, like, manual work that people shouldn’t have to do starts to disappear. We have customers who are doing hundreds of thousands of password reset phone conversations. Those shouldn’t be done, by… by humans anymore. That’s a waste of their time, and they should be on more important conversations.
But every single piece of customer interaction should become a source of intelligence that improves the system. Over time. That’s where things are starting to shift. It’s like.
ChatGBT, then Agentic for really hard-coded workflows, and then now the next phase is, which we’ve been building towards, and you’re gonna see this with other companies and other AI systems, is can it learn from itself, and can it be custom to your business, not some giant model.
So, this is… this is, like, where AI moves from just isolated productivity gains into very specific things into actual enterprise-wide, operational leverage. And… So, the companies that win won’t have the most AI. The customer is actually your competitive advantage as a business.
And the first thing people come and they ask me is, like, back to where do I get started with AI, what’s my advantage, as a business? Your biggest advantage? is not which model you can use, and how cheap you can get it, and can you scale it. It’s… Actually, your data. Your data, and a lot of people tell you, your data is your moat.
when you… You have all this data as a business. You have it everywhere. Taking that and turning it into something that’s a custom experience for your business, that’s bespoke, you know. adjusted for your customers, your business, is going to be the biggest advantage you have. And it’s… and AI right now is not good at taking a lot of data.
Like, if you were just gonna go and use any model out there, it has a context limit, it doesn’t have really good memory built in, it doesn’t have the ability to parse all this stuff and then figure out what to do about it, and then it can’t then get better.
So… this… this is starting to become a world where… where, you know, obviously AI is available to everyone, but, like, the customers that understand, like, their… like, the businesses that understand their customers will be able to build better products, will be able to deliver better experiences, like, make decisions.
That are better, and just outperform everyone else. So… Thank you, I don’t know if there’s questions, potentially, but, I know it’s a lot in a short amount of time. But thank you.

Julia Nimchinski:
Phenomenal presentation, thank you so much, Brian. Yeah, we have a lot of questions, and the first one is… People are asking if you’ll be able to share the deck after the session.
Brian Peterson:
I think so? But yeah, we can, we can follow up with that, I think so.
Julia Nimchinski:
Cool. And, you mentioned that, you know, the biggest advantage now is data. I’m curious your thoughts, commentary on, you know, something that Alex Karp has been very loudly articulating lately, when it comes to, you know, enterprising feeding their alpha back to the LOMs, to the interior labs.
What are your thoughts and Does Dialpad address this directly?
Brian Peterson:
Sorry, that was… question was about… sorry, the… what about the…
Julia Nimchinski:
About the enterprise concern, the…
Brian Peterson:
Oh, Trade, like, them training on your data?
Julia Nimchinski:
Yeah.
Brian Peterson:
Yeah, if you… a lot of the enterprise plans now. you know, have, you know, conditions where they will not be allowed to train on your data. I think that’s what most big businesses are doing now, is their enterprise plans, they do not train on your data. I think you can trust that. I wouldn’t say that that’s a big… deal, necessarily.
It used to be a big deal. In the early days, they were taking that data, and there was no rules around it. I think it’s safe to trust those models, as long as the terms and conditions of your contract say they cannot train on your data.
I wouldn’t say that’s actually the biggest problem, but the biggest problem to me is that it’s… you don’t really have the ability to customize any of it.
So, that’s kind of what my speech is a little bit about, is you’re going to start to see various custom models tailored to your verticals and your businesses that We’ll be trained off of different companies, not necessarily the big foundational model providers.
Julia Nimchinski:
Makes sense, and… One question from the audience. Is there a minimum amount of customer interaction needed before the flywheel actually starts delivering materially better outcomes?
Brian Peterson:
Yeah, you only need… so it gets better, obviously, with more data, but you only need, honestly, about a week’s worth of data. to get… to get some real initial value out of it, because a lot of these interactions, if you have enough volume, they’re very repetitive, or if any customer conversation continues to learn on itself.
So even if you have one customer conversation, if the next one is adjusted based on the previous conversation, you’ve already gotten value. It does obviously get better when you, like, have more and more data on the same customer, but you get it.
You can get it instantly after even one conversation, and for your whole business to understand what your common… like, what’s working with your sales inbound, what’s not working with your sales inbound, you can get within, like.
I mean, a week, probably, worth of decent volume of data, because we do things like, we have an AI customer satisfaction model we use for sales and support that automatically analyzes what conversations are going well with your customers, so that it can automatically adjust, like, hey, don’t give that answer again, they don’t like it.
And you would never be able to know that at volume, that that works well, because you wouldn’t be able to listen to all of those recordings, right? So… Yeah.
Julia Nimchinski:
Makes sense. And another question here, what’s the implementation process like, and how difficult is it to deploy in a typical enterprise?
Brian Peterson:
Yeah, the, the… the deployment, so we… we own the entire stack of, you know, your contact center, your inbound, outbound, dialers, all of that with… with our AI, so… Rolling out the, we were… we were the former Google Voice team, so we’ve spent a lot of effort making it super easy to deploy to all levels. We have a lot of enterprise customers.
they can get their entire communication stack moved over in as fast as they can set it up, usually. The Agentic side of it, the AI, is automatically built in, so they get all the benefits of of the transcription, summarization, the unified customer journey, pretty much immediately.
The Agentic side, we work with you with a engineer to understand exactly what your workflows are. Move. their data over, and then we train on it, like the flywheel, and then we give you your own skills mining, as we call it, where we say, okay, now we’ve analyzed what you do already, here’s what we recommend to automate.
And getting stuff up is a lot of what we’ve been working on, is to make it super simple and quick to do that with all those pieces I talked about, with being able to test it, so you can test it at full volume before it’s even live, you can You can roll it out slowly, just to a percentage of your customers. You can version it.
These are things that you just get a complete advantage if you have this whole system together. So, and it’s getting better every week. So, just being honest, I mean. it’s… you can get up and running almost immediately. For more complex scenarios, we work with you, and it takes a little bit longer, but it’s all getting better every single day.
And so, once it is set up, it’s something you can also maintain yourself.
Julia Nimchinski:
And Brian, one more question here, is comparing Dialpad with platforms like Genesis, Nice, Five9, what are the biggest differentiators?
Brian Peterson:
When we built Dialpad, there was… never really unified communications. You had… everything kind of was different.
You had contact centers, you had unified communications providers, but it was all felt like it should be the same thing, and you really lost all this data when you had separate systems, so we thought that this was broken when we started.
When we came out of Google and we said, no, this is… this is the way you, like, communication’s supposed to work, and then 8 years ago, a few years after we started, we knew AI was going to be the biggest part of communications, so we completely invested in AI starting in 2008 with an acquisition, and we’ve been embedding AI directly into our systems, so… we think, we think, we can control the entire customer journey better than anyone else, so that when it goes from your AI chatbot to a human, you’re not like, hey, who are you?
What did you just talk about? Who… like, can you verify yourself again? We think we’re the only company who can actually have this entire unique flywheel, because we own a lot of our AI ourselves, too, which I didn’t mention, but a lot of the stuff to make it super custom to your business has to be your… you have to own your own AI.
So we have our own models that are constantly trained based on what is best for the business, and then it’s also real-time. So, a lot of our competitors don’t have real time, we have real-time coaching, we have real-time customer satisfaction evaluation.
But basically, we thought all of that stuff wasn’t the best for the customer, and we thought we could do it better. And we think we have, and of course, I’m biased, so… but yeah, we think it’s just a complete experience for the customer, where they actually love the experience they have, because they have everything in one.
Analytics is completely unified. We do very large businesses, enterprise across you know, 50 different countries, so it scales instantly, and it’s just an overall better experience. We put a lot of effort on design, on our AI systems.
We have, you know, many PhDs and advanced, scientists inside of Dialpad, and linguists, so these are things that these systems haven’t really, Put a lot of effort until recently, so…
Julia Nimchinski:
Thank you so much, and one last question. Folks are asking, how do you keep bad habits from becoming part of Flywheel? Say, mediocre sales conversations. And, yeah.
Brian Peterson:
Well, yeah, you have to keep… you have to monitor… like, there has to be observability, there has to be evaluations, you have to continue to… the system itself should do this, that’s why we knew that this is going to be a thing. Your conversations over time change. So, like, it has to continue to learn.
Like, the way you answer a question or pitch someone, your playbook will be different over time. It needs to constantly be looking at that. And then, our biggest thing that we’ve had, and it’s our own proprietary model, is we… we have stuff that’ll evaluate these real… these conversations, and tell you Is it working or not?
And is the customer… we actually do sentiment analysis, so we can tell you, does the customer like your AI agent? Does it like your human agent? Which human agents are the best, based on what they respond to?
Because, last thing, the… People just want their problems solved faster, and they want they don’t care if it’s a human or AI, but they… but you really need to make sure it’s a good experience. So, we do a lot of stuff to make sure you’re always monitoring that, and can adjust for that.
Julia Nimchinski:
Thank you so much. Phenomenal session.
Brian Peterson:
Thank you, thanks for listening, everyone. Thanks.