This 3x founder hit $1M ARR in 5 months. Here's his playbook. | Roy Moussa, Founder of GetVocal

Roy Moussa (00:00:00) :
We went with one salesperson from zero to one in five months and two weeks or something like that. Over Christmas, a big term got coined, which is context graphs. We put a deterministic AI based on the most intelligent and efficient way to do it. Which is graph based and go, and have a look. The world just woke up to that. We built this three years ago and we've been capitalizing on it. For businesses, AI that just chats is useless. That's the way I see it. The two worlds of process automation and conversation automation, or agents for CS. They have fully merged. We had deployed an agent for Courier in Europe. We deployed for one use case, showed a lot of value, a lot of ROI. Then all of a sudden, it became very contagious in the organization and so it spread without us doing anything about it. Another team wanted on, and they took it on, and they started creating, you know, agents. Not just for CS, but for onboarding new restaurants and then another team took it and it was for commercial. And, that team took it and that, and all of a sudden within eight weeks or so. There were eighty agents live within that organization and that was another aha moment for PMF.

Previous Guests (00:01:17) :
That's product market fit. Product market fit. Product market fit. I called it the product market fit question. Product market fit. Product market fit. Product market fit. Product market fit. I mean, the name of the show is Product Market Fit.

Pablo Srugo (00:01:29) :
Do you think the product market fit show, has product market fit? Because if you do, then there's something you just have to do. You have to take out your phone. You have to leave the show five stars. It lets us reach more founders, and it lets us get better guests, thank you. Roy, welcome to the show, man.

Roy Moussa (00:01:45) :
Thanks a lot, Pablo. Good to be here.

Pablo Srugo (00:01:47) :
I'm excited to have you on. You're based in Paris, right?

Roy Moussa (00:01:49) :
I am actually based in the southwest of France, but the company is based between Paris and Lisbon. We have two big hubs there.

Pablo Srugo (00:01:56) :
I say that because, you're obviously growing really fast. Last year you raised a $3 million seed round at the beginning of the year. By the end of the year, you raised a $26 million Series A and most of the time when I see that, it's companies based in the Bay Area. Sometimes in New York but there's more and more coming out of Europe. So it's great to see that. Let's start a little bit maybe just with your background. I know, I mean, you're a multi-time founder. GetVocal is not the first company that you started and I think, obviously one thing builds on the other. So tell me a bit about yourself and kind of the other companies that you built before.

Roy Moussa (00:02:29) :
I come from an engineering background, actually. I'm an engineer, and I've built a lot in my life. Before I started building companies, I built, you name it. I'm an aerospace engineer by education, so I've built a plane, plane models, even electric cars before they actually became a thing, just, you know. Built all sorts, and it was only natural that I started building companies afterwards. I'm an engineer turned business, so commercial. It's always a good plus and basically, thirteen years ago. I started with my co-founder and CTO, whom, still thirteen years later, we are launching companies together and we kicked off all in AI even before AI became a household name. Back when it was time series prediction, machine learning platforms, et cetera, and we started our first company in 2013. Two successful companies, this is our third, and we kind of put all our sweat, blood, and tears. And all our learnings from our past companies into this one. And I definitely confirm what you said earlier, you know, the more you do, the easier it becomes.

Pablo Srugo (00:03:37) :
What drives you to? I mean, there's a lot of engineers turned founders but, you know, as a discipline, it's very analytical, very precise. Oftentimes, many risk averse types end up in engineering, not always. Obviously but in your case, what drove you to leave that predictable world, let's say and go into startup land?

Roy Moussa (00:03:54) :
Good question. Engineering was my way of understanding the world or trying to really dig, to peel a few layers and understand how it's built, how it works, how it happens, et cetera. But at heart, I'm as well very much an outspoken people's person. I've lived on four different continents, traveled and lived in various countries all my life. So that kind of fed that other side of me, and both came together to make a decent combo for founding companies. And all the skills that that takes, whether it's leading a team and hiring the best talent around you, and inspiring them or raising funds. And having the confidence of your investors, and future investors in your startup or challenging technically, you know, what is the product vision without me being the builder or the one that's coding. Not to mention that coding is less and less a thing. But being able to challenge that vision technically and from a go to market or business perspective is two very good skills to have.

Pablo Srugo (00:05:00) :
Walk me through, what were your other two startups. What did they do and maybe just how big did they get just to build some context.

Roy Moussa (00:05:07) :
Yeah, for sure. My first startup was actually that. We learned a lot there because we started horizontal. As 2013, there was a wave of massive, what we call, massive machine learning platforms. So it's actually at the time Palantir had just started as well, and that's what we call the massive machine learning platform. So, data science for enterprise and how you build models and with the right data for various use cases. So we started very horizontal, Data IQ, as a European company that started at the same time as us, and quickly we verticalized. Because typically that's what you do as well, into actually one of the use cases that showed most traction. So already, you start quickly understanding focus, what it means for companies, and how critical and important it is. One of the verticals that was working a lot, and where we were having more, and more customers, and traction, was for retail and logistics. And, we were deploying computer vision systems for retail stores and warehouses to help them manage inventory instantaneously. And everything merchandising, and everything that comes with that. We scaled that over five years. We were a hundred and something people at the end. Raised a couple of rounds, and it was right before COVID that we exited, and sold the company. I stayed for a short stint at the Choir. That was a great, very eye opening experience as well. Taking an SVP role there before I had to go back to what I like to do, even just for the heck of it. Building from scratch, there's a lot of adrenaline that comes with just building something. Adrenaline, you know, creation is, I think, one of the really incredible passions or an incredible thing.

Pablo Srugo (00:07:02) :
Something innately human about zero to one. I mean, that's why I love zero to one. It's just making something from nothing.

Roy Moussa (00:07:08) :
Yeah, right. Whether you're a painter and you get a white canvas, and the inspiration comes in, and you paint, or whether you're an entrepreneur and you build a company from thin air. There is something extremely gratifying. So there's a lot of growth, personal growth in it, and so I had to get back to it. And who better to do it with than my good time friend and co-founder. And so, once you find a good working marriage, so to speak, you know, you stick to it. And so, we kicked off again, this time we changed from computer vision to conversational, 2022, 2021, 2022. That's when obviously conversational has become the modality. So it was for us, OK, systems thinking, how we built systems before that did computer vision AI for reliable real-time actions, driving efficiency within organizations. How do we take all our learnings from that system building for that modality, image and video, and convert, and transfer it to conversational?

Pablo Srugo (00:08:11) :
And this is before the ChatGPT moment.

Roy Moussa (00:08:14) :
No, that was, I forgot, what version we were on but that was early days. Yeah and what we did, actually, we wanted to do something personal for a personal cause. That happened to affect both of us, my co-founder and I. So we launched conversational agents for Alzheimer's, actually. For Alzheimer's patients and dementia patients. And that kicked off, fundraised, et cetera, and.

Pablo Srugo (00:08:36) :
When was this? Is this 2023 or '22?

Roy Moussa (00:08:39) :
Yeah, by that time it was 2023.

Pablo Srugo (00:08:40) :
OK.

Roy Moussa (00:08:41) :
And then quickly there were clinical trials, long cycles, et cetera. So that's when we thought, OK, great, doing a personal, you know, thing for a personal ambition and having an impact in that sense. But this is going to be a long ass ride in terms of the equivalent of sales cycles and your buyers in that market, and all that. And so we shifted from there, and we spun off what we built there, and that's when we met again with our VP of sales from our first company. Who was managing a team of sixty BDRs, AEs, customer success, et cetera. Having a lot of all the issues we face within organizations that big in terms of efficiency, fluctuation of performance, cost, et cetera. On customer service, on support, success, even on the commercial side and that's when we kind of came together. We're like, all right, well, let's go with this and we had already the foundations built, and thinking about it in a systems way. And so we kicked off with GetVocal in 2023 there.

Pablo Srugo (00:09:49) :
Did you create a new company or just kind of part of the same company?

Roy Moussa (00:09:51) :
No, no, no, we created a new company. That was kind of a handover of that company. We handed it over to a more medical leadership team that's still ongoing and scaling at its base. And then we kicked off GetVocal.

Pablo Srugo (00:10:04) :
And so this first idea that you're doing late '23, let's say early '24 is what exactly? Within this new kind of with the VP sales or whatever.

Roy Moussa (00:10:13) :
Yeah, the first version was GetVocal. Actually, it hasn't changed much. We were focusing on omnichannel conversational agents for, despite what the name of the company. We outgrew the name now clearly, and there will be a rebrand, and a rename very soon. As often happens with companies, you know, whether it's a Monday.com or you name it. Most company names have been changed somewhere on the way once you outgrow them, and you almost always do. But it's omnichannel conversational agents for the customer journey. The entire customer journey as we started. So by actually having a deep insight on a customer journey end to end, from prospection to activation, or qualification to activation, to support, service, NPS, et cetera. The whole shebang, you truly can create a full fleet of governed, interconnected agents for that. We started there, and then same thing. You hone in and you find your focus, and I'm a big fan of focus in that sense but it's always a cycle we find out of exploration and focus, exploration and focus. And so we narrowed in, obviously, on the use case that is most transformative. Which is customer service particularly, and that's where cost pressures are high, and that's where, you know, adoption or buying, at least is there. Not to say that our product is still, and our platform is still horizontal and we can talk a bit about that. You know, since the show is PMF, I can hover around PMF a bit. Basically, we started at customer service now, whereas before it was the full customer journey, et cetera. But now we started customer service and then expanded across the customer journey. And that was kind of the aha moment on PMF when, for a few of our customers. We entered through one function, customer service, and we built agents for that that performed and showed a lot of value. And all of a sudden, it became contagious within the organization. And because of the way our product is built, and how easy it is to grab it. We build a centralized enterprise. We're based on a graph. I don't want to get too technical, I guess, on the show but we can. But basically, it's a centralized spawning ground for agents for your entire enterprise.

Pablo Srugo (00:12:37) :
Just to get more specific, tell me a little bit about. Let's say, you meet this VP sales, he tells you a bunch of problems. Do you start working with them and solving. yeah, and solving their problem or where do you go from there?

Roy Moussa (00:12:49) :
Yes, indeed, we started with that. We started with our VP of sales. He's our CCO today, but he was scaling a Series C company and so we solved their problem first.

Pablo Srugo (00:13:00) :
What was that problem? Yeah, because I mean, high level efficiency, whatever. But, you know, what are the specific things that were that were going wrong? Let's say where he's like, yeah, I can really step in and make a difference here.

Roy Moussa (00:13:09) :
First, it was on the commercial side. So it was BDRs. He had tons of SDRs and BDRs, and performance was fluctuating. It wasn't there. There were so many chunks of their responsibility that could be automated and actually ensure better, and more consistent performance. And so we started on that, and we took all past call recordings, transcripts on how they're doing outbound. How they're doing inbound lead qualification, how they're maintaining records in the CRM, and how that information is flowing. The processes behind that were very human heavy and so, oftentimes, prone to not being done or done in a poor fashion. And so we automated that. We went in and put commercial agents, BDR agents, actually in the beginning. Doing a lot of the heavy lifting from those processes.

Pablo Srugo (00:14:04) :
So what was the initial part? Was more on the outbound side or was it more post call, take the transcript, enter into CRM?

Roy Moussa (00:14:10) :
No, it was both. There was outbound to reactivate cold leads, for example, that had been there. There was inbound qualification. There was one more use case, now I can't remember. But it was these three use cases, top of funnel.

Pablo Srugo (00:14:25) :
Did you do them serially? Were you trying to figure out where the need was most clear, or why did you solve different problems at once? What was the approach?

Roy Moussa (00:14:34) :
Yeah, that's exactly it. You want to get to the best bang for your buck, right? The highest ROI, highest urgency for that customer, and see if it's repeatable, if it's the same for the rest of the customers, whether it's in that vertical or across horizontally. So indeed, we did them sequentially, but very, very fast.

Pablo Srugo (00:14:55) :
And what did you find? I mean, ultimately, you ended up going in customer support. Was there not as much ROI in these cases? Or what was the outcome of this?

Roy Moussa (00:15:03) :
It was an intersection of many parameters. ROI, once you think about all the factors that need to come in. Customer ROI is one of them, urgency, how much investment, and, where you are on that cycle of buyers and whether they are ready to buy. Which ICP and how you define your ICP there makes a huge difference on what you're going after. And what kind of organization or what kind of go to market you want. And what you're good at, whether you're going after SMBs or enterprise. So all that, feed regulations, regulation risk, et cetera. All that feeds into it and, you quickly need to see and evaluate all those factors to actually understand, where to orient the company, where you have the highest chance of success in getting quick traction and fast growth.

Pablo Srugo (00:15:56) :
Where did you go next? I guess I'm curious on how you? What drives you to ultimately go from these sales efficiency, AI for sales efficiency to customer service agents?

Roy Moussa (00:16:06) :
So, like I said, as we started exploring these use cases, which span kind of on the upper range of the funnel. We started going after other, better customers, design partners and, went deeper and deeper along that customer journey. While understanding the market landscape, who's there, which players have been succeeding, whatnot. So once you've taken all of that, plus the first five, seven beta customers and design partners. You quickly start to understand where you're heading, actually, where you're heading home. I was definitely more enterprise and further down the customer journey, where it was very FTE heavy. It was very people heavy organizations, internal or externalized, and the ROI was instantaneous. And there are two moments there where it becomes very clear to you, as how significant, you know, once you have actually that level of ROI. Your customers are shouting about it, you know.

Pablo Srugo (00:17:16) :
Do you remember those moments, like those specific stories?

Roy Moussa (00:17:18) :
Yeah, I remember two, we can go through them. One is related to that, indeed. It was for a telecom. It was very quickly, eight weeks after we iterated quickly. We have a continuous learning engine for our agents, with teams as well behind that, supervising, guiding. You know, at first on there, when we first launched the first agent, and that's typically what we see anyways. Results were OK-ish. Then we quickly iterated, improved on the performance, et cetera. It was a Saturday afternoon, and they were operating around the clock. And we saw a huge peak in the metric, in what we were measuring. Which is first time resolution and it was right after an A-B testing, and a new AI agent we had just launched on Thursday or somewhat. And then we started looking at that peak, and I was like, is that for real? And, I called my co-founder and also our implementation guy. The guy, we were just a few people at the time, and we couldn't believe it. So we had to dig into the details and the data behind it. And really understand whether, and it was. And then on Monday morning, first thing we did was redo, redo with the customer directly. You know, what's the ROI exercise on that once we scale it to your full volume, et cetera, and that was a key moment there.

Pablo Srugo (00:18:39) :
Was this, just to go deeper on this. The AI at this point, was it just chat or was it taking actions on behalf? Like, in order to resolve a case or an issue, was it just a matter of Q&A, answering things and, just going back and forth with the customer? Or did it actually, was it able to go out and do things for them?

Roy Moussa (00:18:57) :
I'll just be very clear on that, and I'm very critical about it. For businesses, AI that just chats is useless. That's the way I see it and that's how our customers see it. Our system, the two worlds of process automation and conversation automation, or agents for CS. They have fully, they have merged, at least in our world.

Pablo Srugo (00:19:22) :
Is this the problem? Willow, because chatbots for customer success have been a thing, you know, you had Drift, you had Ada, you had a bunch of different ones. I mean, they got some traction, all those companies but it never seemed to really, say, replace humans, right? Was that the missing piece?

Roy Moussa (00:19:35) :
That's one of the missing pieces. In fact, I can tell you about the three others as well.

Pablo Srugo (00:19:39) :
Yeah, what are the others?

Roy Moussa (00:19:40) :
Indeed. One of the pieces is process and conversation together. You need to actually deliver the end value end to end. Otherwise, it's not worth the squeeze. You're paying one buck a resolution just to answer a question. Customers quickly realized that, hey, it was kind of exciting and innovative in the beginning. But once you do the numbers, it's not. It's not worth actually the one buck I pay just to answer a customer's question about a product that doesn't justify even that cost, right? So that was quickly on. It was very clear that we needed to embed or have a unified system of conversation and processes all in one. The other one is how complex of a customer journey you can actually handle or customer experience you can handle. Most chatbots and AI agents out there hit a ceiling and break past the ten percent simple cases, OK, call it twenty percent depending on the organization. And if you are truly in it to solve the hard problem, and actually bring value, and automate really what the customers are after. You have to actually build something that can handle the entire customer experience and then it's up to the customer. And this is the second component, how much automation and how much human do you need in there? It's up to each customer, and there's no one size fits all. That's the other piece. How complex can you do CX or CS or customer interactions? The third piece is there is no one size fits all. I don't know if you saw the scandal recently with Gap and Sierra. I don't know if you heard that in the news, but it definitely goes to show there was a massive hiccup with AI agents that Sierra launched for Gap and it really goes to show that you cannot build a one size fits all. Because the way Gap does CS in the US is not the same as Chase Bank does it in the UK, or you name it, right? So what we've built is a platform that allows our customers to build an agentic CX for them and what that means is they choose how much automation, and how much human is in there. How much determinism and how much generative is in there. So that's kind of their level, their comfort with risk, and the use case they're using, et cetera.

Pablo Srugo (00:21:59) :
I'm going to ask you for a small favor, a tiny little favor. In fact, it's not even now that I think about it. It's not even really a favor for me. I'm actually trying to help you do a favor for you. Just hit the follow button. You won't miss out on the next episode. You'll see everything that we release. If you don't want to listen to an episode, you just skip it, but at least you don't miss out. Yeah, walk me through just maybe a question on that. How much is it, because you have the old chatbots, they try to do actions and it was purely deterministic, if this, then that and so you can have all these rules. And, you know, it's good because it'll never mess up but it's really bad because you get stuck in these. As a customer, as a user, you get stuck in these paths and all of a sudden for whatever reason, it breaks. Plus there's so many edge cases, it's just impossible to get them all correct, right? The flip side is if you go fully probabilistic, you know, you're going to have some situations that might happen that you didn't want to happen. So what is your solution to that? What's the best of both worlds?

Roy Moussa (00:22:54) :
I'm so happy you asked that. That's exactly one of our core USPs, and that's what we built this system for. And it's also because of the time that the company was launched. So that's, there's a lot about timing there because once we got going on GetVocal, it was GPT-3.5 out. Which was already starting to become a strong LLM model. So we got to test very quickly its capabilities, its limitations, understanding what it means to have a probabilistic model doing business interactions and business transactions. And the bottom line is you need the best of both worlds. And so you said the right term there. And we've built a system versus, you know, what is mostly out there. Especially post-LLM, which has been a lot of LLM wrappers. You get an LLM, you give it a UI, you try to put constraints, a lot of constraints on it, guardrails, et cetera. That's mostly what's out there and there are two core problems with that. One is what you said, you know, whether it doesn't behave freely as you want, and maybe it does depending on the use case, or if it's simple enough and frequently asked questions. It would in ninety eight percent of the cases, and then maybe it'll break in two. And for example, the Gap situation twenty days ago. That's one, and two, scalability really of that model, whether it's from a cost perspective or how do you scale it for enterprise wide. You know, as you go wider and wider, with the automation and agents for the enterprise. How do you actually scale that in a centralized way. What we've built really is a system that's deterministic first, actually. The front facing authority is deterministic because that's what you need. Business conversations are deterministic, and it makes no sense to put a pure probabilistic thing to do a mostly deterministic process.

Pablo Srugo (00:24:50) :
And to make this tangible, you're talking about. For example, talking about Telco, like, oh, I was billed $30 for international charges but I shouldn't have been billed. Well, there's a specific answer to that once you figure out what the issue is.

Roy Moussa (00:25:02) :
Exactly, or I want to, I just changed my address and I tried to do it online, and it didn't work. I need to update my contract or whatnot, right? And so the way we saw that and built an entire system, versus just put a model and wrap it. We put a deterministic AI based on the most intelligent and efficient way to do it. Which is graph based and go, and have a look. The world just woke up to that. We built this three years ago and we've been capitalizing on it. But over Christmas, a big term got coined, which is context graphs. I don't know if you've heard of that. Check it out. We saw that three years ago and built it, and we've been capitalizing on it ever since.

Pablo Srugo (00:25:43) :
What is a context graph and yeah, how does it apply here?

Roy Moussa (00:25:46) :
Are you familiar with knowledge graphs?

Pablo Srugo (00:25:48) :
Yes.

Roy Moussa (00:25:48) :
Knowledge graphs are, just to give you some context. Knowledge graphs are semantic databases that connect and hold relationships between words, concepts, and items. So this document talks about a cat, this is a cat, or this book has a picture of a cat in it. They're interconnected with relationships, and that's a semantic database, and that's a knowledge graph. Very powerful for many things, quickly you can search in it, you can identify patterns, you can identify stuff. Three years ago, we took that and we thought, OK, we built from the ground up. Purposely designed for processes and conversations, and we based it on something called the speech act theory. A new class of graphs that actually interconnects things not based on semantics, which is knowledge but based on intention. Which is, in any conversation, in any process, there's one single unique intention at a time and so it's a network of interconnected intentions. Anyways, we are getting far more technical than that, but that's the best way to have memory. Because that's what it does, it stores memory. Advanced reasoning capabilities because you connect different parts or steps of a conversation, or intentions of a conversation and a process.

Pablo Srugo (00:27:06) :
What's an intention for this specific purpose? Like, I want a refund or I want to change my contract. Is that the intention? 

Roy Moussa (00:27:11) :
That's an intention, yes, sir. When we're talking together right now, we're exchanging intentions. You just asked me a question with the intention of understanding more and I'm answering you. Every single thing anybody says, there's an intention behind it, and a unique one. And this is extremely powerful for the sake of conversation automation and process automation. And so within that same graph, there's advanced reasoning, memory, process automation. Because processes are within the graph, and it's generating conversation. What that means for our customers is that A, they have a deterministic authority doing the decision making, conversing, triggering processes, reasoning, storing memory, and doing the decision making in a very traceable, and auditable way. Because the graph leaves a trace and a timestamp, and everything in it. And at the same time, it goes and calls upon the generative whenever it's needed. And what generative is great at is fluency, contextualization, realism. It makes it feel human. So we call upon generative when needed in the conversation.

Pablo Srugo (00:28:17) :
How do you build actually? How do you build a context graph for a specific customer? Is that where you go in, part of onboarding, you take in all their chat history and you kind of build up to that or how does that happen?

Roy Moussa (00:28:26) :
That's the beauty and the secret sauce here. So we go in and take all their knowledge bases, as messy as it is, and we take what's critical in terms of processes. You know, in a typical process or a conversation process, for lack of a better term, in merging those two.

Pablo Srugo (00:28:44) :
These are documents that an enterprise would have? Like, standard operating procedures, whatever, right?

Roy Moussa (00:28:48) :
Yeah, yeah, exactly. Oftentimes it's messy or redundant, or from multiple sources and by the way, that's one of the huge sources of problems with it. Just a standard LLM doing RAG onto these knowledge sources and trying to answer questions. As often you have conflicting answers from two different sources because that's the reality of enterprise building up knowledge bases across, you know, a huge company. We generate from that, and that's where offline we can leverage a lot of the generative capabilities, et cetera, plus team supervision. We generate from that a centralized context graph for that company and from there it can serve as, it’s continuously and dynamically being grown, trimmed, pruned, reinforced on some paths.

Pablo Srugo (00:29:36) :
How understandable is that? If I'm the end customer and I want to make sure that I know what this system is going to do when it's confronted with a certain situation. Is that readable to me?

Roy Moussa (00:29:47) :
That's exactly it. Also, I'm glad you ask great questions, Pablo. There are many ways to read the way we've built our graphs. One is, have you seen, you know, you've seen knowledge graphs where it's radial and there's different bubbles of different sizes, and they're interconnected, et cetera. It would look like that except that in each node, there's not only a concept like a cat. It's actually, there's an intention, a conversation, a process, or an action, or a series of actions that are being triggered, et cetera. So you can read it as such, or we have just layers of information on it and it actually is transformed into a typical knowledge base format that you would actually read. And it's continuously being updated with every agent contributing to it. And that borrows from generative whenever it's needed, and the beautiful thing about this is you can start with a very simple context graph. Which is, or what we coined three years ago, conversation graph. You can start with just one critical process and leave everything else to generative. But what's happening with every conversation is whenever a pattern is detected, a process that is defined. It's being repeated, we're servicing that case every time in the same way. Why keep it probabilistic and call upon a model every time? It has become deterministic.

Pablo Srugo (00:31:08) :
Does that also mean you could, you know, because this thing is being updated and codified, and basically documented. You could, for example, as an enterprise decide, you look at your analytics and you say, oh, we're giving ten times as many refunds in this situation. Or a lot of our refunds are coming from this situation. Let's analyze that, oh no, you know what? In this case, we shouldn't give a refund. So then you go in, you change that, and then.

Roy Moussa (00:31:28) :
One hundred percent, the system itself is analyzing this and is measuring. Because of the way it's interacted, we measure at every turn and every word that the AI chats or voices. We're measuring a whole lot of performance stamps that are actually our performance metrics that are then stamped onto that step. I realized we got very technical, and so what that means is we're understanding the performance of every syllable that the AI says.

Pablo Srugo (00:31:58) :
You know, and just to be clear, on the technical piece. I typically don't go this deep into product, but this is an area where we're talking about customer support. If you take it too superficial with this, it's like, yeah, OK, AI for customer support, you know, it should have happened a long time ago. They try to happen, now you did it, now it happened, you know, how. The only way to know why this is taking off, right? And Decagon, which I'm going to bring up in a minute. Why they're taking off is almost the specifics of the product that maybe you got right this time.

Roy Moussa (00:32:26) :
Exactly, it has failed to properly pick up until you actually figured out how to build the right system for it to actually work and so that's quite powerful. Because you're continuously registering things, whether it's using probabilistic models, LLMs that are picking up on, they're very good with exploration, right? That's what LLMs are great at. They're picking up on certain patterns, and then you ingrain them just like you would do in a chip. You ingrain them then, and they become a deterministic business process, as business processes typically are and so it's continuously growing that base or that graph. And so, you can rely less and less as you go on token heavy probabilistic models for deterministic business processes anyway. So, and by doing that, it's extremely traceable, extremely visible for you on what your CX is becoming or what that process is becoming. What is the performance being measured? How's the sentiment? When we talk about, you know, when there's refunds, we refused in those cases, we gave refunds over twenty. What was the sentiment like? What was the customer intent like? What was all of those at that particular point? And that actually helps feedback to improve your customer service and improve your product, by the way. Because, as a lot of my customers say, the best customer service is no customer service when your product actually doesn't require it. So our whole point with that as well is to feed back and create that feedback loop in a very performance driven, metrics driven way back to CS and beyond, and product teams, and the rest of the organization. That's what we're very much succeeding with. The other moment on PMF we touched upon, I want to talk about that other moment quickly, and then we can move on to your other questions. The other moment is when we had deployed an agent for a delivery for a courier in Europe, they're called Glovo. The equivalent of DoorDash in Europe and we deployed for one use case, showed a lot of value, a lot of ROI. We doubled their uptime in the restaurants they serve and groceries, and a bunch of other KPIs like that. Then all of a sudden, it almost became contagious with it and because it's horizontal. And it's built in a very, very appetizing and inviting way just for the other team members to pick it up. It became very contagious in the organization and so it spread without us doing anything about it. You know, and then another team wanted on, and they took it on, and they started creating agents not just for CS, but for onboarding new restaurants. And then another team took it, and it was for commercial, and for sending reminders, et cetera. And that team took it, and all of a sudden, within eight weeks or so, there were like eighty agents live within that organization completely being orchestrated. We have a fleet management system that actually helps you manage your fleet of AI agents. Just like typically you would have managed a fleet of people with some management systems and that's another piece, by the way. And we'll talk about it, I mentioned I've unveiled two pieces now. I'll talk about the other pieces as I go and that was another aha moment for PMF. When your customer and across teams, and across functions, and across personas are gobbling it up.

Pablo Srugo (00:35:46) :
Yeah, one piece is kind of that. I mean, part of it is qualitative but it's almost like that demo to close rate. When you show somebody, how bad do they want it and the other piece is expansion, right? Once you're in there, how much does that just inside? Especially at enterprise, right? And so when you got both, you're in a good place. Walk me through onboarding and especially selling this type of solution to enterprises. How much do you have for deployed engineers? How much customization do you need to do? Because you're talking about eighty agents and them kind of creating it themselves. It sounds easy, but I mean, there's so much behind that on the integration side. How do you make that happen?

Roy Moussa (00:36:18) :
Yeah, I wish it was very easy and at the same time, I'm happy it's not. Yeah, that is obviously one of the challenges, right? The reason being at enterprise is there are all those barriers or hurdles that you need to help the customer overcome and crush for this to actually be properly deployed, and for it to show maximum value, and ROI for the customer. And it's your business to do it as well with the customer. And it doesn't sound as cumbersome as, it isn't as cumbersome as I just made it sound. But the more you have, you know, you tie the customer function with product, and you remove as much of that friction in your product. Obviously the easier you'll have it for deploying, expanding, showing value super quick. There's a few things to say about that really, but we do have teams for deployed teams or solution engineer teams. We do have, we've scaled a lot as well, our back office to do a lot of the services that are needed to get this properly at scale within the enterprise in a very, very streamlined way. So we have plenty of teams that are working around the clock. They're helping do the various tasks that are needed along the life cycle of an AI agent. Once you look at it, whether it's in the design, or obviously it goes right from solution engineering and selling to enterprise. You actually, you know, you have to understand where the value lies with the customer, help them orient on which use case brings the most value. What's the lowest hanging fruit to actually do that, that's early on, and then designing the right agent for it. Helping obviously integrate in the right systems, helping with the continuous learning and the continuous improvement of the agent. That's a crucial part of it and a lot of companies seem to have missed that. Expecting something from the get go to be the highest performing thing is like getting somebody from the street, giving them a piece of paper for instructions, and telling them, OK, now you're doing this job. Good luck, you have to perform by tomorrow. So there's that continuous improvement piece, and this is what we've built a lot into our system in a very scientific way. How do you improve the agent's performance? And so teams that are actually responsible for that, but we've done it, and here's what we've learned in our past in building systems and scaling machine learning, AI, et cetera, computer vision. We were, you know, in the first company, analyzing like eight million images every month and sending actions for every single customer. You know, for a given particular customer it was eight million images analyzed and sending actions. We've already been through that and had to build scalable systems and services behind to ensure quality, to ensure continuous improvement, et cetera. In a very scalable way and that goes to product, and how you build your back office. And how you actually make it extremely scalable from day one versus having to do that as an afterthought, and having to roll back, and go back, and do it.

Pablo Srugo (00:39:34) :
I don't want to go too deep into this, but just because we had Ashwin from Decagon on this podcast not long ago. Is it direct competition? Do you go after a different ICP? Is there any one liner difference between the two of you? What's the quick answer to that?

Roy Moussa (00:39:48) :
From a market perspective, there isn't. It's the same ICP. We're not in the US at the moment, though soon we'll be, expectedly. So from the market, it's the same. From a product, I think we took a very different approach on the two first pieces I mentioned, whether it's deterministic on the forefront and get from generative. The context graph at a base, the technical stack we've built and we've engineered is definitely different. And just that approach on it is very different on how to build the best of both worlds really from the get go and not as an afterthought when it starts failing it. When it hits, you know, when it hits that ceiling. So that's one, human centric AI is another. We didn't talk about it, but it's a massive piece. How do you actually build, and I was so happy to see a new startup that got launched with some heavyweights as founders, human. And it was three months in, it raised, I think, $300 million to $400 million for like a $4 billion valuation or something. And that's a huge validation as well of what is actually needed to scale AI. You actually need the right kind of system that puts humans and AI in a very interactive way. A two way interaction system where AI can get the assistance of humans who are needed, and vice versa. And that's actually feeding, and coaching the AI's incremental performance. And its incremental increase in scope, and what is automated along the way. And that's a vision we had from the get go. Again, with that thirteen years of experience, you start understanding how not to expect, no matter how shiny on the surface a demo looks of an LLM in 2023. How not to expect it overnight to scale across and automate eighty percent of an enterprise function, and actually it's a glide path to it. And you have to actually have a good system that accompanies that with the humans at the center of it.

Pablo Srugo (00:41:46) :
Perfect, well listen let me stop it there. I'll ask the three questions you always end on. The first one is different because we talked about the moments of PMF, but let me ask you, by the time you decided to do AI for customer support. How fast did you hit a million ARR?

Roy Moussa (00:41:58) :
We went with one salesperson from zero to one in five months and two weeks, or something like that.

Pablo Srugo (00:42:06) :
Crazy and then now I guess you're on this one to ten path I would assume. Was there ever a time on the flip side, where you thought things might not work out could just fail.

Roy Moussa (00:42:17) :
What founder doesn't ask that question to himself. I have to, you know, I have to say though. Your second, third time as entrepreneurs, I guess they feel that all the time. You're having less of those moments. You still have them, just to be real but you have less, way less of those moments. Because as you build knowledge and wisdom, with all humility. You build confidence, and that's a virtuous circle.

Pablo Srugo (00:42:47) :
And what would be your number one piece of advice for an early stage founder that's looking for product market fit? 

Roy Moussa (00:42:53) :
My number one piece of advice is not to take advice, or at least take advice with a grain of salt and figure out what works for you. There's way too much advice being thrown out there, but you have to actually figure out what works. That's one but agility and focus. Agility is a word that was thrown around way too much in the days, but in today's world everything is a moving target. Really, go to market is a moving target, everything is shifting all the time. Focus, know your customers inside out. That's probably the strongest piece of advice. Understand who you're selling to, understand their pain points, understand what they're willing to pay for. The urgency, the deeper you understand your customers, the easier you'll have it and focus on solving one core thing, and doing it well.

Pablo Srugo (00:43:44) :
Perfect, well, Roy. Thanks so much for jumping on the show, man. It's been great having you on.

Roy Moussa (00:43:47) :
Thanks a lot, Pablo. I loved it.

Pablo Srugo (00:43:49) :
Wow, what an episode. You're probably in awe. You're in absolute shock. You're like, that helped me so much. So guess what? Now it's your turn to help someone else. Share the episode in the WhatsApp group you have with founders. Share it on that Slack channel. Send it to your founder friends and help them out. Trust me, they will love you for it.