Pablo Palafox, HappyRobot | theCUBE + NYSE Wired: Mixture of Experts

Palo Alto Studio Connection Silicon Valley and Wall Street. I'm John B here with Dave Volante, my co-host. Welcome back to the Cube Studio here at the New York Stock Exchange. I'm Jemma Allen, host of NYC Wired. We are connecting Silicon Valley to Wall Street. Today we are talking mixture of experts and an expert in the space around how AI agents are moving beyond answering questions to actually doing the work. Happy Robot is building AI agents that handle the calls, emails, and messy coordination that keep business running from logistics to energy to telecom. They've just raised 150 million at a 1.2 billion valuation. Joining me now is their co-founder and CEO Pablo Palifox. Welcome Pablo. >> Thank you so much. Super excited to be here. So interesting company, interesting time. Maybe just to start level set with me. Give me the 101 on Happy Robot. >> Super. Uh we basically help enterprises put agents to work in some of the most complex environments. We work with Telos, utilities, financial services, and supply chain, which is actually where we started. We've been operating for a couple years. Uh raised over 200 million uh dollars in funding. And uh yeah, very excited to uh have announcer CC a few few weeks ago. So Siri, August 4th, I believe you guys announced that, but you have had a couple of raises in a pretty short space of time. Correct. >> Yes. We went through Y Combinator summer 23. We pivoted away from what we were doing back then. We started the business really 24. >> Wow. >> And we had our first sale April of 2024 and then we uh raised from Andress in Haritch, our series A, and we did our series B last year with B 10 partners. >> Wow. >> So a fun couple years. >> Two years post OC. First unicorn in the intake. >> Yes. first unicorn in our batch. Yes. >> First of all, congratulations. I mean, that's incredible. But let's get into what made you a unicorn. Let's talk about this company and this product. So, first, >> it's an interesting time in the world of AI and enterprise AI, right? We hear a lot from different companies, different models, different service providers claiming that they can solve all sorts of enterprise problems, automate, save money, save time, save headcount. You started this business so with kind of a unique enough problem case, right? You were looking at freight. >> Yeah. >> It has evolved. Talk to me about the trajectory and the decision to start something that had a very vertical focus. >> Yeah. [snorts] >> So there's a a very interesting take here which is intelligence is not the limiting factor. How you will that intelligence is the key. So basically LLMs are out there. You can use them in your in your company. theoretically everyone should be able to already automate all of their messy clunky processes with with AI right but the reality is different the reality is that these enterprises which is what we are focused on have a lot of coordination issues to put it in a way there's a lot of handover of information from one party to the other customers are reaching out you have an issue with a customer that customer service representative might actually have to turn around and reach to the financials team and figure out the problem with that customer then a let's say a trucking company that is delivering your parcel if we're in the supply chain space uh has to actually move that shipment. So there's all of this coordination in those industries that we kind of uh have under the umbrella of the real economy. No. And what we saw serving the logistics side is wow these enterprises the problems they have they're not really supply chain specific. They're just a coordination problem that an enterprise suffers from. And when we were working with folks like DHL, which is one of our early customers, and they connected us with folks like Doce Telecom, we started to learn that that coordination problem across the enterprise was what we could solve with AI with a platform really that could wield and coordinate these agents. So that's really all we're doing. We're coordinating agents, doing the work, executing, gathering the insights from the real world, and serving customers and really anyone across the uh the enterprise environment, employees, partners, So I want to get into some of those examples of customers you gave. But first, what was truly unique from the perspective of what you could standardize was the fact that it was human to human interactions that they were repetitive that there was a whole lot of data which is very live right very instantaneous data especially in the world of fright logistics that needs to be managed. What was it that you thought, okay, this is a aha moment. If we can solve for this, >> it it it's a transferable solution. [snorts] >> We probably built one of the first voice agents back in 2024, January 2024. That was the analog to a lot of the processes that traditionally have been run on phone. Today, we do any channel, boys, email, WhatsApp, chat bots. But voice was really the key unlock. We uh have put a lot of care into that. We actually have a research team building text to speech models uh for our own voice agents and that was the unlock because in particular freight moves a lot of the uh of of their shipments through voice through calls and email but that was the key unlock. Today we work with nine of the top 10 freight US freight forwarders and freight brokers sorry uh those folks have a lot of those communications through phone. So that was the the key unlock. >> Give me some examples of some of your top customers. I know you're here in New York meeting with some folks just today alone. Give me your kind of best customer proof point and talk me through the before and the after of their relationship with Happy Robot. >> We work with a with a large uh parcel company who has to run anywhere from 20 to 40,000 customer phone calls and follow-up texts every day. So imagine to run 20 to 40,000 phone calls plus a follow-up text every day. you need a a lot of folks doing very repetitive stuff. And if you're working for one of those companies as a customer service rep, you probably don't want to spend your time making the repetitive and mundane stuff. You want to add value to the company. So that's really what we help these c our customers with. We help them with an execution layer really to put agents to work in these very repetitive and mundane and complex tasks many times. um really enabling their humans to elevate themselves as the new guardians of exception and manage the exception. So with these customer this partial company, those 20 to 40,000 calls, they're basically outbound to customers that haven't maybe paid duties on a parcel. >> Wow. >> That is very interesting because it unlocks a lot of ineffic reduces a lot of inefficiencies. When you haven't paid duties on a parcel and you're waiting for your parcel to arrive, you might not even know you have to pay those duties on that parcel, right? But you might just think that it's lost. >> In reality, it's actually waiting on a warehouse and you just forgot to pay those duties. If an agent actually follows up with you every day or whenever in a very formally fashion, orderly fashion, uh, and just checks in every every other day with you and you end up paying those duties, you end up getting your package. The customer or our customer also saves space in the warehouse and that package doesn't have to go to origin. That's one example in the supply chain space. Now other example in the uh telco and utility side uh we help our customers in the telco and utilities uh with b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b basic customer support. But that customer support is not a shallow customer support. It's actually a type of customer support that requires agents to turn around and figure out who what technician is going to go fix your router issue or your leaky boiler. That is a sort of coordination problem that our agents can help with. You call in with a leaky boiler, but in reality that problem needs some physical labor uh to go to your house and fix it. That is a coordination that an agent today can do very efficiently with humans as that escalation point which is really really interesting. How plug and play is this technology like talk me through the example of that parcel company right that's a huge ecosystem they're managing daily a lot of different workflows a lot of different humans also I'm sure all over the world how quickly can you integrate this and how quickly can you realize value from a product like copy role but I presume it takes a lot of API integrations a lot of talk me through it [snorts] >> fun fact this use case that I mentioned was deployed in less than four weeks >> wow >> so that that was a a pretty fast one. Um what we see typically is the faster whenever we can deploy the faster is when the customer is leaning in the most. Uh and you need to have executive leadership really leaning in first. So you need top down decision- making. Uh so that is really the key unlock for us to move fast. When you are aligned at the exec level, everyone is really going towards the same place. Our key proposition here is we bring a platform and a deployments team and this is when executives gets get really excited because many times they just get a platform for building agents let's say from one of the hyperscalers they just get a platform to build their agents and then they're left alone really to use that platform maybe train their teams to use it which is something they can definitely do but we do think that the deployments uh the deployment side of the house Having experts, what we call forward deployed engineers, which is now a very fancy term that everyone uses and palenteer pioneer. No, having a an engineer that can also think business onsite with the customer for weeks as much as needed. That is really what's unlocking speed in our deployments. We assign uh full-time forward deployed engineers to our customers to really be the drivers of the value and that is really what's unlocking the speed of execution. So some time ago, I used to work in supply chain for Microsoft Windows, right? We would bring on a new client or a new customer and it would feel like we reinvented the wheel every time because not everyone is as like strong in their governance and knowledge documentation as others, right? There's varied levels of documentation and context context capture. Yeah. Across organizations. >> How do you solve for that? Like you know you you mentioned four weeks but in the more complex scenarios where maybe the ducks haven't been in the exact role you would hope for especially from the perspective of structured data and even documenting workflows full stop. How how do you solve for that in these complex environments? >> Maybe going to the point before about FD that catalyst these forward deployed engineers are the catalyst to make things happen. that complex workflow you're mentioning that really requires someone to sit down next to the operators but also next to the exec team and kind of bridge the gaps. >> The reason why even companies like our size we're like a 200 people company and we internally realize we almost need forward deployed engineers to bridge a lot of the >> the missing pieces that we that we had internally if that makes sense. Imagine if if a small company like ours needs some form of catalyst to make things happen. Imagine a 200,000 people company, you need to have some form of bridge, some form of catalyst between the operators, what's going down in the what's going on in the field, what's going on in that warehouse, what's going on in the in the in the energy plant and bring it back to leadership and then make a decision. So that is where a deployment motion like ours um really unlocks the value of AI because again intelligence is not the limiting factor. You cannot just throw LLMs at the problem, right? >> Absolutely. >> You you need someone to wield that intelligence and make it useful and rethink processes. That's actually another piece that is interesting for customers. Just rethinking the process might even allow you to to to to make it more efficient on its own. you don't you might not even need AI for certain things which is also a bit of a contradictory or or interesting take no >> so what's the business model here I mean what you're describing is a certain amount of service spend too right it is you know advisory spend as well as platform spend explain it to me is it consumption based is it seatbased how do you what's the market position for this >> we talk about a transformation we talk about hey customers uh Mr. customer. This is a transformation and as such we're going to drive you somewhere. We're going to drive somewhere together. So our model really is a combination of three pillars. We have a platform. We have um the driver of that platform which is the forward deployed engineer the services if you will and then the consumption the credits. >> Okay. >> The platform is that expensive car that is going to drive you somewhere. The credits is that gas that you need just to move somewhere. It's it ends up being an afterthought really for our customers. And then the driver is someone that's going to teach you drive that car, that platform somewhere. Eventually, you're going to be able to drive it on your own. So, we actually try to create a lot of these workshops with customers to bring them all in the same room, teach their data scientists and engineers and business folks how to use the product. So, yes, initially it is uh heavy on um our deployment motion. We don't really see as consulting. It's it's a deployment motion because we're deploying agents across the company and then we try to again give the keys to the to the to the Ferrari to the customer so that they can start building on their own. >> Let's go under the hood of the platform for a second from a tech perspective. What you're building is it 100% proprietary? Do you use various other models within your stack? You know, I guess it's totally cloud-based. Are you seeing cases where companies want to have a certain amount of data sovereignty for particular workloads? Break that down for me a little bit. >> So let's look at the platform from the three layers that compose it. We have the execution layer. That's agentic workflows. >> Okay, that's we'll do a place where our agents live or the agentic workflows live >> because agents is a bit of a too f too fancy of a word. >> We try to demystify a little bit. It's just workflows. Everything in a in a company really is a combination of some workflow, some process, uh some data, and uh really executing on that. So, we we we start with that workflow layer. That's where you build your agents, maybe an agent that is fielding, uh an inbox, an email inbox, and looking at what's coming in. That's a workflow. You can have an agent that is fielding all of the inbound phone calls coming into a a warehouse or coming into your utility or telco and navigating that conversation and doing whatever the customer needs to do. Is it sending a technician over to you? That's the workflows. Now agents or agentic workflows need data. So that data layer sits on top. That's what we call uh twin in our case. Twin is really our data layer, >> our context layer to put it more fancy. That context layer integrates to systems of record of a customer, your CRM, your um data links, whatever existing system of record you you already have data in and you want to keep data in. So we have two layers right now. The third layer is how do you surface the insights that your agents are gathering and how do you know how to improve that human in the loop almost that's our interface layer we call that apps uh in our platform you can go and vibe code applications UIs to put it simply interfaces to show the work that agents are doing so that human teams can actually you know either guide agents or learn from what agents are doing. Wow. >> So those are the three layers >> for our proprietary like what is proprietary today for us is our voice agents. We've built as I mentioned before on our own uh compute uh texttospech models uh which are really fantastic and we have a small seven people team in our research team which is fantastic today with a mix of transformer models and open source models. I come from a research background. I was doing my PhD in deep learning and transformers were just getting started back in 2017 when they started. Today you can have a small team of uh seven machine learning engineers build your own text to speech models. We're looking into SLM, small language models to distill a lot of the LLM knowledge into the day-to-day work. If you think about it, you don't need a PhD to be calling a driver to see where they're at. >> And the key advantage there in terms of small language models be what? cost efficacy. You don't need to spend like the token usage would be lower, the commute lower. Yeah. Okay. So, >> ownership. >> Absolutely. >> To the customer. >> That's huge. >> That's that's another point you brought it up before on the data side. We sometimes deploy uh single uh single tenants really to to customers that want to make sure that this we have a multi-tenant solution, a cloud solution where you can also deploy on prem for a customer. So, yeah, data ownership is is key. >> Let me ask you a question. And I ask a lot of folks that come on the show that have started in the space which is quite virtually aligned like like you and your team Happy Robot have. When you think about some of the conversations that are happening right now around SAS apocalypse and this one huge orchestrator right there it's rumored that this week Darede went on the record and said Anthropic might be the only private company that exists in 20 years. No one can actually verify that source. A lot of headlines are running it. How do you think about the competitive force of these large language models? You know, once they break into enterprise, they start to control certain workflows. They're very horizontally aligned. How do you think about that from a competitive perspective? >> I think it goes back to the point that uh intelligence itself is not going to autodeploy itself in a business. >> Maybe unfortunately, I mean, maybe we would all be better off if we could just have an LLM explore everything in a comp. The reality is that you do need a platform to build that intelligence and today's intelligence from Antropic might be better or worse than today's intelligence from OpenAI or Gemini. So you you don't want to marry yourself to a an intelligence provider. You want to marry yourself, I guess, to potentially multiple orchestration platforms. We we acknowledge we're not going to be the only orchestration platform or agent platform our customers are going to use. I mean I would love that but like we acknowledge we need to be interoperable. Uh what we do best is X maybe company Y does something else really well. We all are going to serve the industries that we tackle or we serve uh in different ways. And I think as such the the orchestration concept is key. It's not again back to your point oh you just throw an LLM from entropic to the problem and it magically fixes everything. And Pablo in that scenario the stickiness the moat of 5 10 years from now is it that early customer loyalty early context a real niche understanding of a specific workflow within an enterprise you know you're kind of there you're like really in the weeds of it or you know is it cost like how do you think about what makes it sticky in in a world like you said where people can move between models quite easily right it's not like the world of cloud it's a different beast Agreed. We talk about uh intelligence compounding in your enterprise. When you build agent one maybe takes let's say four weeks but be building agent N plus1 takes less time because you already have a lot of the the company context. You already have the integrations done. So that speaking as comes really from earning the right to do more. We we talk a lot about how do we earn the right to do more for our customers because we've demonstrated that building a certain agent for customer support is adding value. Well, how do we move on to building a sales agent? We do a lot of sales for our customers. Now, interestingly enough, the sales agent will be talking to the same person than the customer support agent. That's why we don't see ourselves as a point solution for any particular function but rather as an orchestration platform across functions because we believe that all agents in your company should actually tackle or or tap into the same context. So to your point uh really the the stickiness comes from earning the right to do more and serving our customers in a better way. You know, I love the term earning the right because I feel like tech, especially this wave of AI, can be quite self-righteous, you know, and quite, you know, I guess in some respects sometimes almost patronizing from the perspective of enterprise. So, I think that's a great message to lead with. Pablo, last question to you. 150 million in the bank. Nice little runway there, I'm sure. What's ahead? What are you going to spend that money on? What does the next 6 12 months look like for you and the team? >> Products and deployment. >> I love uh the the product team continue growing that team and and building better models focus a lot on the SLM side of house side of the house to distill a lot of the intelligence into SLMs and deployments again the catalyst in the in the in the enterprise is our deployments team they they make the the magic happen they uncover value they deliver that value and they make sure that we are earning the right to do more >> love it well Pablo also love to see a span in the US doing so well thank you so much for joining us on the cube and >> wire thank you so much. >> I'm Gemma Allen here at the Cube Studio at the New York Stock Exchange. This is Mix Your Experts, one of our programs with NYC Wired. Thanks for watching.