Building AI Sales Reps: How ShowMe Orchestrates Voice, Video, and Multi-Agent Workflows to Close Deals

Summary

Podcast Summary: Just Now Possible

Host: Teresa Torres
Episode: Building AI Sales Reps: How ShowMe Orchestrates Voice, Video, and Multi-Agent Workflows to Close Deals
Date: February 19, 2026

---

Overview

In this episode, Teresa Torres interviews Yuri (CEO) and Kike (Lead Product Engineering & Customer Success) from ShowMe—a company building AI-powered digital sales reps, primarily for inbound sales organizations. The discussion chronicles how ShowMe spotted this opportunity, built their initial prototypes, iterated through real-world deployment, and architected their multi-agent, multi-modal workflows to close deals efficiently. The episode is a deep dive into product discovery, agent orchestration, technical trade-offs, and the ongoing challenge of evolving AI sales tools with the market.

---

Main Discussion Points

1. Origin Story and Product Focus

• AI-First from Day One:
  • ShowMe was founded in April 2025 as a natively AI-focused company, directly inspired by their prior experience noticing website visitors not converting unless a human interacted with them.

    "We actually started the company April last year, April 2025. The idea came from...a lot of people were coming to our website and...not converting. What we found is that when we were able to put one of our sales reps...in front of our visitors...they basically converted." — Yuri [04:00]

• Inbound > Outbound:
  • The team intentionally specialized in inbound sales, contrasting with "spammy" outbound practices that degrade user experience.
• Role of AI Digital Workers:
  • ShowMe’s sales reps have two skill areas: sales skills (video calls, calls, demos) and operational skills (reporting, performance metrics, etc.), designed to function as a co-worker within the client’s sales team.

---

2. First MVP and Early Insights

• MVP Construction:
  • Launched within weeks; focused on demoing products with basic video, voice Q&A, and manually ingested documentation.

    "You had some videos. You could select which video you wanted to see...and then there was an agent...talking about...the product...Very simple, Very clunky as well. But yeah, that was the first version." — Kike [08:26]

  • Early user surprise: even basic voice interaction was novel mid-2025.
• Manual Foundations, Iterative Process:
  • Early knowledge base was built using RAG (retrieval-augmented generation) from scraped docs, product recordings, and help centers.
  • Rapid learning: product needed to support more complex buyer journeys beyond demos—slide presentations, qualification, customer stories, and dynamic guidance.

---

3. Building Trust & Affordances in AI Sales Agents

• Overcoming Skepticism:
  • Users were initially unsure whether the AI was capable or trustworthy, treating it like a "dumb chatbot."
• Affordances & Interaction Design:
  • They experimented with making the agent look and act more human (e.g., with realistic video avatars via HeyGen, and replicating familiar UIs like Google Meet).

    "When they saw the avatar, it makes them think, hey, so this could have the knowledge of a real person. ...their brain instantly has that connection." — Yuri [15:20]

• Mimicking Human Behaviors:
  • Explicitly designed agents to deflect “connect me to a human” requests by convincingly demonstrating capability, thus winning user trust.

---

4. System Architecture & Orchestration

• Agent/Workflow Overview:
  • Agents: Primary interface with users—via voice, chat, WhatsApp, email.
  • Workflows: Deterministic processes above agents, handling multi-turn, multi-day engagement journeys (e.g., demo → follow-up email → call).
  • Movement toward more dynamic “orchestrator agents” for less rigid, adaptive workflows.

    "This workflows is like a layer that is above these agents and ... have both deterministic and indeterministic actions." — Kike [21:10] "We're step by step automating and automating each part of the process." — Kike [21:10]

• Coworker Metaphor:
  • Internally and with clients, ShowMe strives for a model where their AI rep behaves—and is interacted with—like a team member, especially via Slack.

---

5. Multi-Agent Approach: Specialization for Latency, Cost, and Quality

Conversation Agents (Sales Reps)

• Decomposition for Performance:

  • Complex tasks broken into smaller specialized agents for greetings/discovery, qualification, pitching, product walkthroughs, and closing.
  • Orchestration agent routes conversation based on call stage.

    "We build...different agents actually inside it with different prompts. One is going to take over the greetings...one is going to be expert in qualifying..." — Kike [34:16]

• Tool Sets Per Agent:

  • Each agent gets a tailored set of tools (e.g., slide presentation, calendar, Stripe, product videos, knowledge base access).

• Personalization/State Maintenance:

  • System collects and accumulates interaction context (a "dossier") to inform follow-ups and avoid redundant questions.

Creator Agent

• Prompt/Pipeline Construction:
  • Converts customer-provided docs, transcripts, call data into tailored prompts, knowledge base entries, and clean data for conversational agents.

    "The creator is to create the conversational agents...all the documentation...and creating the prompts...plus creating the documents..." — Kike [47:47]

Evaluator Agent

• Automated Assessment & QA:
  • Evaluates each conversation for qualification, next steps, roles, sentiment, and confidence levels.
• Determines When to Hand Off to Humans:
  • Real-time detection of uncertainty or frustration—triggering human intervention if thresholds are met.

---

6. Training Agents: Generic and Specific Sales Skills

• Two Layers of Skill:
  • Generic sales skills (prompted in): discovery, qualification, pitching, following up.
  • Company-specific skillset learned from call transcripts, onboarding materials, and sales enablement docs.

    "We ask our customers to share with us...transcripts of the calls from their sales teams...And from there...we create for the conversational agents...sales guidance skills for the company." — Yuri [39:43]

• Role-Specific Playbooks:
  • Agents learn which value props and slides to deploy based on customer role/ICP, using extracted rules from customer materials.

---

7. Evaluating Agent Quality & Customer Rollout Process

• Careful Staged Rollout:
  • Always proves efficacy via POCs, A/B tests, starting with low-value leads, then ramping when trust is established.
• Multi-Phase Evaluation:
  • Phase 1: Customer closely reviews most conversations, provides granular feedback and corrections.
  • Phase 2: Evaluation agent surfaces a subset of at-risk/low-confidence conversations for review, reducing manual effort over time to ~5%.

    "At the beginning, every conversation...is reviewed with the customer...A test is created, and the agent tried to go through that conversation again..." — Kike [52:58]

• Automated Recurring Testing:
  • Feedback is encoded as regression-style tests to prevent prompt regression as the system iterates.
• Reporting & Transparency:
  • Dashboards, Slack alerts, CRM integration allow client teams to monitor agent performance in real time.

---

8. Technical & Product Challenges

• Balancing Determinism and Flexibility:
  • Striving for enough guardrails to avoid hallucinations, but adding agents (orchestrators) to break rigid workflow logic where flexibility is required.
• Latency vs. Quality Tradeoffs:
  • Use of smaller, faster models for real-time voice; larger models for backend eval and context management.
• Scaling Manual-to-Automated:
  • Many processes began manual (“concierge”) and are incrementally automated based on customer usage and feedback.

---

9. Patterns & Industry Implications

• Emergence of Agent + Workflow Patterns:

  • ShowMe's architecture echoes that of other advanced AI teams (e.g., Gradient Labs) utilizing persistent state/workflow above specialized agent turns for multi-day tasks.

• Humans in the Loop:

  • Both for building/training the agents and for safety/quality guardrails at runtime.

• Future Differentiation Will Be Customer Understanding:

  • Technical capabilities are rapidly commoditizing; deep discovery and ongoing adaptation to real customer workflows is the durable moat.

---

Notable Quotes & Memorable Moments

• On demonstrating AI’s abilities:

  "Whenever someone really understood how to guide the AI, how to talk to her and so on, they were really good at extracting the value from it and they were converting... So how we solved this...by basically trying to replicate humans..." — Yuri [15:20]

• On the “affordance” of human-like UI:

  "If we try to make it like a video call, same as Google Meet...people are really used to it and they will understand this is like being in a call with a human.” — Yuri [18:30]

• On decomposing AI complexity:

  “Engineers have always had to be good at decomposition, but now we’re learning it in an LLM way... It's like a new skill of decomposition.” — Teresa [35:55]

• On prompt regression testing:

  “Automatically a test is created, and the agent tried to go through that conversation again with that...feedback...we run that test until it's passing that test. This way, we're creating...a battery of tests with...partial conversations that need to be passed.” — Kike [52:58]

---

Timestamps for Key Segments

• 00:58 — What ShowMe Does: AI Digital Sales Reps Overview
• 04:00 — Company Origin Story & Market Problem
• 08:26 — First MVP and Manual Knowledge Base Approach
• 13:41 — Expanding Beyond Simple Product Demos
• 15:20 — Building Trust and Human-like Affordances
• 21:10 — Architecture Overview and Workflow Orchestration
• 34:16 — Multi-Agent Decomposition for Sales Conversation
• 39:43 — Training for Sales Skills: Generic vs. Specific
• 47:47 — Evaluator & Creator Agents: Roles and Workflows
• 50:49 — Measuring Quality and A/B Rollout Strategies
• 52:58 — Two-Phase Quality Feedback Loop and Test Regression
• 58:35 — What’s Next: Smart Orchestrator, Automation, and PLG
• 61:22 — The Fast Pace of AI Industry Change and Customer-Centric Discovery

---

What’s Next for ShowMe

• Technical Roadmap:
  • Evolving the orchestrator for smarter, flexible workflows.
  • Building toward a self-serve, DIY/free trial motion (“PLG”).
  • Automating more of the agent creation pipeline.
• Market Vision:
  • “Digital workers” vision—beyond sales reps into customer success, support, and other roles as driven by customer demand.

---

Takeaways

• ShowMe exemplifies the “concierge-to-scale” playbook: Start manually, learn from each customer, and automate only what truly creates value.
• **Multi-agent systems and orchestration/workflow layers are emerging as a common AI product pattern for complex, multi-turn, real-world use cases.
• **Human-like affordances—both visual (avatars) and behavioral (sales process fidelity)—are crucial for user trust and adoption.
• **Technical excellence alone won’t win—deep customer intimacy and adaptability in product development are decisive differentiators in the fast-evolving AI landscape.

---

If you're building with AI and curious how ambitious agent-powered workflows are architected, evaluated, and evolved in production, this episode is essential listening.

Transcript

[A]

Welcome to Just Now Possible with Teresa Torres.

[B]

I'm Yuri, I'm the co founder and CEO at Xiaomi which is the company and product we will be talking about today and digging deeper into the technical stuff. Prior to Xiaomi like I used to work at AI companies specifically in AI for market intelligence.

[C]

Hello, I'm Kike, I'm the co founder and lead product engineering and part of customer success. My technical background in general working in product in AI both as pm as developed as well full products and have always been an enthusiast of entrepreneurship and have built my own side projects and different startups in the past. Yeah, happy to be here and share all our knowledge or part of it at least.

[A]

Yeah, excellent. I'm excited. Let's get into. Tell me a little bit about what Shomi does.

[B]

Yeah. So at show me we basically build like AI sales reps for companies and we decided to specialize for companies selling mainly via inbound channels, right. And the way we approach this is as digital workers that are basically teammates inside our customers sales teams and in the end just to keep it short like they have two type of skills, right? On the one side they have the sales skills and on the other end they have operational skills. Same as any teammate inside any team, right? On the same skill side they can do something similar to video calls and this is they can share their screen, they can speak via voice, via text, they can also demo a product. Then they can also obviously call phones, right? And they can also send messages. And the cool part I think about our product is that we also orchestrate all these tools right for the teammate to achieve their objective. This is on the end of their sales skills and then on the operational skills side they basically do same as a human employee which is basically reporting via Slack via email, also sharing metrics under performance and so on. And just on the outcome they basically create either qualified meetings for our customers or they try to close contacts contracts directly. So yeah, in the end and to summarize it, it's like digital workers, right? Specialized on the essence part especially for companies selling to via inbound as their main channel.

[A]

Yeah. There's a few things I like about this right off the bat. I like that you're focusing on inbound. I see a lot of companies creating what I think of as spam for all of us where they're just blasting everybody with outbound which I think is not contributing to the world. I guess it's generating some revenue for a company maybe. Okay. But for the rest of us creating a bad experience. I also Find it fascinating like the depth at which you're going to this. So your AI agents are doing video calls with people and they're making phone calls. So really pushing on this idea of your agent. As a coworker I've done some previous episodes with, I interviewed folks at a company called Nepal and they also are building agents as coworkers. Really focused on customer success but. But really very text based still. And then we've had a few episodes where we've explored voice agents and even one company perk. Their voice agents make something like 10,000 outgoing calls a week. Not sales related, more around like hotel reservation management. But this I think we're going to get into some more sophisticated orchestration than we've seen so far. So I'm excited about this. Let's start at the beginning. Tell me a little bit about like where did the first of all, are you an AI native company? Did you start the business with this in mind? And then tell me a little bit about what was your very initial idea, how did this. Your product sounds like it has a big footprint. Where did you start?

[B]

Yeah, so 100% AI native company. We actually started the company April last year, April 2025. And the idea basically came from something we saw in the previous company we worked for. So I was leading a go to market team there and what we saw is that a lot of people were coming to our website and for whatever reason they were not converting into close one customers. So there were multiple reasons. Some of them were not filling the typical demo form, some did, but they fall through the process and so on. And we started like basically digging it, digging in it to solve it. And then what we found is that when we were able to put one of our sales reps, human sales reps, in front of our visitors in the website, they basically converted, right? So it was like, okay, how do we basically get people to really understand the value of our product and removing the friction of putting them in front of humans. And so at that time we tried to basically solve it with the typical free trial motion which to be honest didn't work at all because we realized we had a pretty complex product and a product that the aha moment like was too late in the journey and you had to do a lot of things as to get to it. So free trial didn't really work for us. And then on the other hand we started seeing how AI was getting better and better and basically smarter. And also we started seeing how latency, especially for voice models was improving a lot. And this was around February last year. Right. And to summarize this, it was like, hey, so we know that whenever we are able to explain the product with a human and so on, it converts better. Right. And on the other end, we see how AI is getting smarter and latencies are improving a lot. So we could probably create something like a human to explain the product. And at that point, like, the idea came to us and it was like, hey, what if we basically try to create sales reps that we put directly in the website and they can basically explain the product directly to anyone without people having to talk to our sales team. Yeah. And at that moment, we basically built an mvp. We can go into the details now. And we started seeing how it works and that companies were really interested in it. And then Kiki and I decided to leave the company and basically build Xiaomi. So that's how the idea was born. Yeah.

[C]

I wanted to share. Can I share one important component to the insight that I think is very important and something that we realized as well is that in the end, when you think about it, Teresa, people are coming to your website. Some of them are super interested and they want to buy right away. So what we tried is, okay, let's try to put a human in front of them in the previous company as soon as possible. That worked. The problem is that some others, they just want to understand a little bit more and they're not qualified. Right. So in the end, what we saw is, okay, we're putting humans into this, but it's costing us a lot of money that we're investing in people that are not working. Right. And that's where we say, okay, AI can help you in this regard because it can filter that out and it can give a good experience to buyers that are already very interested and try to route them to the a correct person as soon as possible. But for the ones that are not so interesting for us, you can also filter them out.

[A]

You know, this pattern of like you were working together at a previous company, you got exposed to a very clear problem. We have a lot of people coming to our website. They're not all converting. We see if we get a human sales rep in front of them, they convert. There's this insight of maybe we can do this with AI, especially for these instances where they may not be ready to buy. It's too expensive for our human salespeople to engage with them. You leave the company, you start your new company. I think this pattern, I think that's this podcast is called just now possible, because I think People everywhere are starting to see this pattern of like, this problem is now solvable. Let's go and take the leap and go after it. That's amazing. Okay, so you. And it sounds like this is all like a year ago. You've taken the leap, you've started to build your sales agents. It sounds like you've covered a lot of ground in just less than a year. So tell me a little bit about what did that first version of your product look like?

[C]

So we focused a lot on the part of demoing the product. Okay. At the beginning, the first version of it, we built it in a couple of weeks. We actually first got our first customer that was. It was not our previous company. So we really wanted to see if someone really had the pain in order to put something that was clunky in their website. And we got that person, that head of marketing, and we went over, did the product in two weeks. That was basically, you could. You had some videos. You could select which video you wanted to see from the product. And then there was an agent that was going, a voice agent that was talking about regarding the product and like explaining that part of the product. So it was like, I select the part of the product I want to see instead of just seeing a video. I have an agent that is telling me the product and I can ask some questions and get some answers from it. So that was it. Very simple, Very clunky as well. But yeah, that was the first version.

[A]

So that very first version already video already Voice. Is that true?

[C]

Yep, yep.

[B]

It was mainly like a Q and A. Imagine the typical Q and A. Right. That you could basically ask things. But yeah, it was with Voice, which at that time really surprised people, to be honest. And things in AI go super fast. And right now you see anything that talks and can listen to you and is smart and it's. Yeah, I don't really know this works. But yeah, one year ago or one and a half years ago, it wasn't there. So it really surprised people at that time.

[A]

I would argue it's still surprising today. I can't say that I interact with Voice agents very often. Yeah. Other than like maybe the labs where you can just chat instead of. Yeah, I don't know. I don't know that. I can't think of a single service I use that like an agent other than the foundational labs is using Voice to engage with me.

[B]

Yeah, that's a great point. Maybe it's a thing of being really in the trenches. And I really perceive this when we go and Sell our product. Right at the beginning it was like wow, I have never seen something like this before. And right now it's okay. The voice part is okay, but the surprise pricing part is probably more the orchestration.

[C]

Right.

[B]

But yeah, the main insight is it's crazy how fast things go but still there is a lot of room for adoption, that's for sure.

[A]

Yeah. Okay, so this first MVP I can imagine like the agent has to be familiar with the product. If the customer is able to say pick a part of the product, walk me through it, answer my questions. Imagine even for that very first mvp you had to have a way of ingesting knowledge about the product, getting screenshots about the product. Tell me a little bit about what went into that.

[C]

Yeah. So at the beginning we were basically getting the credentials from the product of our customers and recording videos that's basically breaking them ourselves and then ingesting all the knowledge. We built a simple knowledge base with simple rag and that's what we were using. Basically give us all the documentation. We scraped the website, we scraped the help center and with that we build a basic knowledge base and from there we were able to start Very simple. Yeah. Ok. Manual.

[A]

Yeah, it sounds very manual. But this was enough to get you your first customer, correct?

[C]

Yes.

[A]

Okay. And then where did you feel like Customer?

[C]

No.

[A]

Excellent.

[C]

And from there we started to see, I would say a couple of insights that were very interesting. We started to see that seeing the product was just a part of it and many questions depending on the on the stage of the potential customer. If they're like just discovering the company, if they're ahead and they want to see more about the pricing and understand more things like it's just a part of it, understanding the product. And we were seeing that if we focused our product just on that, we were not able to be in the end inside the website of the customer that needs is exposed to all this different range of potential customers or different stages. Right. So we started to not only have product but also be able to show slides not only have documentation and information about the product but also on how the value proposition works and very importantly how to guide the user depending on where this one qualifies. For this part I'm going to show this part of the product but also I'm going to explain this customer success story or I'm going to explain this package. All of that we found about. And then the second big insight was that it's not only about one session and that's over and I'm going to Complete the next step that as you said at the beginning, is that qualified meeting or that sell, you need more. You need to probably go back to the user, try to get their interest back. And that was like what lead to the orchestration part that we will talk about as well.

[B]

And I would add a really interesting insight that has to do a lot with AI is what we also saw is that the website visitors seen our product embedded there, didn't really realize how powerful it was. Right. So we needed to basically to find a way to transmit to the visitor in the website that this could be like speaking to a real sales rep with all the knowledge about the company and that they could really help them, same as a human being. Right. And then we started iterating a lot of things to try to say to website visitors, hey, you can basically one speak to this AI, which is what's something people weren't realizing when they got there. And second, this is not like the previous chatbots. You have seen that they are not that smart and that they won't help you. So there was a lot of work also on the product front of making users basically understand that this really worked and how powerful it was.

[A]

Yeah, I like this insight. This is something that comes up a lot on this podcast is some of these ideas around, like how do you build trust in the AI? How do you expose that this isn't a dumb chatbot? And I can imagine, I know AI video is getting pretty good, but I'm guessing it's still pretty clear it's not a real human. And so there is that you probably have some of that skepticism of and maybe you don't even want to present it as a real human. But so I can imagine people do bring some of that skepticism of can you really help me and how do I get to a human as fast as possible? What are some of the things that you had to do to create that bridge for people?

[B]

Yeah, what I would say is how we first realized about this is what. Because what we basically saw when analyzing conversations and so on, and we can get in front of how we analyze them later because I'm sure you probably want to do that. But what we saw is that whenever someone really understood how to guide the AI, how to talk to her and so on, they were really good at extracting the value from it and they were converting for our customers. So that's how we basically understood that we had to work on this and how we solved this, the way we have solved it for now is by basically trying to replicate Humans. And it's something that I didn't like at all. And I basically discussed this a lot with Kike, and my main point was like, hey, why are we trying to replicate, like, humans when people really know that this is an AI? We don't need to replicate humans. And I realized I was wrong, to be honest. What we basically did in terms of the video call that we were talking about before is we basically added an avatar there that is pretty realistic. So we use hey, Gen for this. It's pretty realistic. And to be honest, after doing several user sessions with users, we realized that when they saw the avatar, it makes them think, hey, so this could have the knowledge of a real person. And their brain instantly has that connection. Right. So that has helped a lot since we started, like, having the video calls with the avatar. Like, the quality of the conversations and the similarity of those conversations to those that they would have with humans is way higher. And again, I was completely like, I didn't want to do this, to be honest, because I had the insight of, hey, this is an AI. We don't need to tell people this is like a human. But, yeah, I was proven wrong by the team.

[A]

It's fascinating. Are you too familiar with the word affordances?

[B]

No.

[A]

This is a design concept. The easiest way to describe it is imagine this is going to sound like a total non sequitur. But stay with me. Imagine a door, right? And if the door has a flat panel, it's communicating to you. Push the door. If it has a knob or a pull handle, it's communicating pull. So the affordance is how the device communicates, how to use it. What I think is fascinating about this, and this has come up on previous episodes, it's like we're all learning how to interact with AI. We all have a little bit of discomfort about is this human like, and how human, like, do we want it to be? And I think we're learning the affordances of what is required to get behavior. So in your instance, you're seeing, you saw some shallow engagement. That was maybe the type of engagement you would get with a dumb chatbot. And the more fidelity you provided, whether that's through better voice or through a better avatar and through, like, more affordances that are, like, human interaction. The more people realize there is some intelligence behind this and I can ask my questions and really engage with it the same way I would engage with a human.

[C]

Yeah.

[B]

And actually see an example of a conversation which is one I really like. So some people used to get to the chat or to this kind of video call. And they were like, hey, connect me to a human. And the way we always responded is, hey, yeah, I can for sure connect you with a human, but I can also answer all your questions. And it basically then shared a slide. And the person was like, wow, this can really help me, and I don't need to talk to the human. And they basically continue continued the conversation and then they extracted all the value and they basically converted by just talking to the avatar there. So, yeah, it's a lot of the concept you shared and an additional insight that it's really interesting here. At the beginning, the way we designed, like the video call was like not trying to replicate the typical Google Meet or the typical zoom. We were not trying to replicate it on purpose. But then we realized, hey, if we try to make it like a video call, same as Google Meet, and so on, people are really used to it and they will understand that this is like being in a call with a human. Right. They can share stuff and so on. And again, it proves us right on that front when we added the video call, like, people started talking more and more to the AI as a human. So, yeah, basically reinforcing on the point you shared.

[A]

Yeah, yeah, I would be that person that literally would just start with, let me talk to a human. But I can also see how I would very quickly gain trust because I think, especially when we talk about support in sales, like, I'm the kind of person where I want to go get all my information myself. By the time I reach out to a company, especially if we're talking about technology, I usually have a very specific. Does your API do this thing in this specific instance? And if you can't tell me that, I don't want to talk to you. And like most salespeople can't tell me that, but I could imagine that your AI avatar could tell me that. And so I think I would start the skeptic and you would probably win me over. So that's fascinating. So here's what I've heard you describe. You have this video, like a way to engage with a video agent. It can show slides. It has a lot of expertise about the product. I heard you earlier, Yuri, say that you've trained it on how to do sales. And I want to dig into that. What does that look like? And it has operational skills. It reports on how it does and interacts with the team. This sounds like a product with a giant footprint, and you've been building it for less than a year. So give me just like the High level. If we were in front of a whiteboard, what does your high level architecture look like? How are you orchestrating all of this?

[C]

Okay, good question. All right, let's start by how it looks like. I think from a. Like to make it simple, we have the agents that are the ones that are interacting mainly with the users of the end users, the users of our customers. Right. And those are usually voice agents or agents that chat via WhatsApp or via email. Okay. And then we have what we call workflows. This workflows is like a layer that is above these agents and that basically have both deterministic and indeterministic actions. We've started more with the deterministic actions and this means, okay, so if a conversation, for example, happens in the website, what are the next steps that need to happen depending on the outcome of the conversation? And that can be a follow up email followed by a call in one day, if you respond, the follow up email, et cetera, et cetera. Right. So that's how we orchestrate the different actions that can happen and that are different depending on the outcome of each of the agents conversations. Okay. Now we're starting to add agents on top of these workflows so that it's able to not just follow always the same path and be able to change paths depending on what is happening. And that will add a lot of flexibility because in the end what we're doing is okay. When we're not able, when we get to an end to that dead end, what happens is that we hand off to the human. Right. And that's happening, I would say around 70, 70% of the times we're able to handle the workflow till the end, 30% we have to hand off to a human. So we're trying to reduce those handoff as much as possible. So that's kind of high level on the part of, I would say interaction with users, then on interaction with the customers. And talking here a little bit about ux, we also want to make it look and feel like a coworker as we were saying. Right. So we don't want you to have a back office or a tool where you enter and you basically go there and do all these workflows or do the prompting of the agents or whatever. What we want to this to look like is that you have a coworker and you communicate with it as you communicate with other coworkers. And if you have an intern or a junior that just joined, you're gonna talk to it on calls, you're Gonna talk to it on slides, you're gonna talk to it on the different communication channels that you have. We started with Slack. So you communicate with the agent via Slack also. This is very much in the works. So at the moment what is happening is that while we have already automated this all the first part like all the onboarding is happening in Slack. So it's able to manage the instructions that you get. And from there it's able to do our first version of the agents and the workflow that of course still is analyzed and supervised by one of our forward deployed engineers. And then we want to get into more. So what is into more those reporting that we were saying at the moment what is happening is that we're getting the reports of our agents and we're keeping it tidy and good looking for our customers and we send them. The idea is that the agents that like the employee is the one that is able to send those overs like on a weekly basis, et cetera, et cetera. So that's all the full map, lots of things. Some of them are still manual and done by forward deployed engineers. And we're step by step automating. And Teresa has how we started this in fake it till you make it. And we almost everything we did in manual and then we continue automating and automating each part of the process.

[A]

Yeah, yeah, I love this. There's already. I'm hearing some similarities with some of our previous episodes. So two patterns I want to highlight and then I want to get into the details. The first is this idea of you referred to it as a workflow that sits above the agent. And this is. We have a. We had an episode with Gradient Labs where they. Do they have agents? They do, they work with financial tech companies to help with things like collecting know your customer documentation, reaching out to get more information for fraud disputes. And it's a little more complicated than just reply to a customer ticket. Right. Because the agent has to be proactive to get more information from the customer. And they described it as they have a state. So this when you said you have your workflow, it almost was the analogous of each turn in a workflow might be an agent, but there's this state that they're maintaining across multiple days even which I imagine is the case in your situation. And then so it's not that there's one agent in a loop that's running forever over days and days, it's that they're maintaining the state and a turn might be agentic. So does that sound similar to what you're doing.

[C]

Exactly, yeah. Because if you think about it like a call, there's always going to be an agent that makes the call, that runs the call, and then we have agents evaluating the outcome of the call and understanding what's the next path that in a deterministic way, it has to pick inside the workflow. Is the next path an email? What's the email? An agent is going to write that email. But you get my point, like, kind of the paths that can get are deterministic. And basically because we don't want, for the moment, any hallucination in that part, we imagine that we're calling someone three times each hour because that has hallucination. Right. So we're trying to have those strict guardrails. And step by step, we will add more and more agents on that part as well.

[A]

Yeah. I think what's fun is we're all converging on patterns. Like, this stuff is early enough that I feel as every team goes out and tries to build a complex behavior with LLMs, we start to see a lot of similar patterns bubble up. And to me, it's like, very fun and exciting to be like, we are probably inventing the patterns that will drive the future. The second pattern I heard, and this is similar to our incidentio episode. So incidentio does they have a swarm of basically, sres, so engineers that are responding to an incident and their human SREs interact with their SRE agents in Slack. So when both parties are investigating an incident, all the back and forth is through Slack. And they came up with that model because their team is interacting with each other in Slack during an incident. And this was a way to feed data back and forth to the agents that are also helping. And this metaphor of we're just going to treat the agent like a coworker and use that as a driver of the UI is also becoming a very common pattern. A little bit disturbing and also, I think, really cool.

[C]

Yeah. Why disturbing?

[A]

I think I. So my background is in cognitive psychology. And actually the program that I did focuses not just on human information processing, but also machine information processing. And so I feel like it was the perfect program for the era we live in right now. But I feel like I've always been really fascinated by these, like, philosophical questions about, like, intelligence and consciousness and who we are and what makes us distinct. And I think we Even today on LinkedIn, somebody asked me, I wrote an article about how ChatGPT works, and I tried to write it to be accessible for people that don't have math Background, and somebody read it, and they're like, I don't see where the intelligence is in this. And my response was like, where's the intelligence in the neurons in your brain? And this is a concept that I think humans have wrestled with for centuries. Like, this realm of philosophy goes back forever, but I think AI is making it very visible right in front of us. Like, interacting with this thing feels like interacting with a human. And how do I feel about that? And I actually. I think it's really cool, mostly. But there is a little bit of this. What does it mean to be human then? Okay, we don't have to get totally distracted by this. Okay, so let's get into your. One thing you said that I think sounds different than what the Gradient Labs team is doing is you have this workflow, I'm assuming a workflow in your model, correct me if this is wrong, is like somebody comes to the website, they're now a visitor. At some point, they get turned into a lead. Maybe they get turned into an opportunity. We're going all the way to closing a sale or being handed off to a human. Is that a correct assumption?

[C]

Yeah, exactly.

[A]

Yes. Okay. So the workflow is managing kind of the relationship with a visitor through to some end. Tell me a little bit about where agents come into play. And then the piece that I. That felt a little different from what Gradient Labs is doing, you talked about adding an agent even above the workflow, and so I'd love to learn about that as well.

[C]

Okay. So, yeah, right now, agents come into play, as I said, mainly in the conversations. Right. And then for this part of building the agent, there are also agents that extract all the information that we're getting from the customer and putting it in the correct place to create those prompts that will create, like, the conversational agents for each of the parts. And then we have agents that are evaluating each conversation. Right. And so we have the conversational agents, evaluators and creators. This is more. More or less how we call them. Okay. Those are the main ones. Then when I'm talking about agent that manages the workflow is like the orchestrator or everything, the orchestrator agent. And that the idea is that we break with that determinism. So example that is happening now for you to understand very easily. We called someone that was in the website, was interested, but didn't complete the different steps to get into a qualified meeting. So we called them the next day, and they told us, I'm busy, call me in a couple of hours. At the moment, as we don't have that capacity or that ability to be able to understand. Okay, so they want us to call them in two hours. I'm going to create a new action that is calling two hours. What is happening is that we're calling them the next day. That is what was pretty fine. You see what I mean?

[A]

Yes.

[C]

So that Orchestrator is the one that is break with that, I would say lack of flexibility that we have now in the actions that we're taking inside that those flows.

[A]

Yeah. Okay, I see. So one of the things that I love about sales as a product person is it's just a funnel. And I can imagine your deterministic workflow is very like a lead comes in, they do a demo on the website through our videos, we get some information to qualify them, whatever that order is. And I could see how you have very clear rules of like how you're moving through the funnel that is a sales opportunity. And then this Orchestrator agent that might sit on top of that is that that's the happy path. But not everybody goes through the happy path. So how do we adapt on the fly as we get there?

[C]

Yeah. And when you think about it, like, of course, our customers have their HubSpot, Salesforce, et cetera, that could do these parts, and they have their workflows and their automations there, what we're seeing is that we're able already to give that extra personalization in what we're getting with our evaluators, what's happening and what the outcome can be and how we personalize the next interactions. But it's key that we also personalize the next actions to happen when they really need to happen. So that's why we're building this.

[B]

And actually in the end, just to add on that note that Kike shared in the end thing, it's same as a human sales rep. Right. In the end, they assign to them several leads, several accounts, and they will basically use the tools they have to try to get to the goal, which is basically converting them, right?

[A]

Yeah.

[B]

And they will start sending emails, they will start sending WhatsApp messages, they will start calling them. And ideally, it really depends what you will do, depending on the status of the lead, who they are and so on. And what do you want to get to with them? Right. So ideally you would try to use them in a personalized way. And this is what Kike is mentioning. We are trying to explore the boundaries there, but also without getting into the part of not using the right tools at the right time and hallucinations and so on, because in the end, and this is something we have not mentioned, this is a critical task for most of the companies we work with. In the end, this is about sales, which is, hey, if we fail at any of this, they are losing revenue, which is something that makes us like experiment. Yes. But being really cautious with the experiments we run because we know the impact directly into their pdml.

[A]

Yeah, yeah, for sure. Okay, so let's dig in. We have a conversation agent, a creator agent and an evaluator agent. I can imagine. I can assume your creator agent is really supporting the conversation agent. So let's start with. Let's start with the conversation agent. It sounds like your very first version of this was your MVP of here's a knowledge base. Here's. You added an avatar. Have a conversation with the customer. The customer's driving it. How has that evolved?

[C]

Yeah, so different things that happen here. First of all, complexity. As we were saying at the beginning, just the Q and A. Now it guides the conversation and that's very different. And we could talk about the sales skills and all the things that we've learned as well in this world and what needs to happen in each moment and how you unveil those pain points and from there you start to dig and extract information. You need to do the qualification and so on. So that's one. And that in the end, going to the more technical part, sometimes we build simpler agents that have a simpler product, et cetera, but most of the times they are more complex and you need to divide in different agents actually inside it with different prompts. And one is going to take over the greetings and first discovery. Then one is going to be the expert in qualifying and one is going to be the expert in pitching and showing off the product and so on. Why we did this is basically our limitation of the current models and a trade off of latency. The heavier the model, you cannot do in a conversation that is via voice, you cannot have high latency. So that made us use smaller models. And for that we needed to divide the problem in order to be able to have a conversation that made sense. And one thing that, for example, for you to understand, the agent was forgetting what they started asking at the beginning. So it started to pitch and then again started to ask the same questions from the beginning. So we needed to divide the agent into different ones. So that's, I think one of the main points.

[A]

Yeah, yeah. I love that even your conversation agent is actually multiple agents in sequence, which makes a lot of sense to me. I think this is A common pattern of if we want an LLM to do anything with complexity, it's how do we take a complex task and break it up into much smaller, simpler tasks, Both for cost reasons, as you mentioned, and latency reasons, as you mentioned, but also cost. Right. Smaller models are cheaper. But I also think this goes back to even that philosophical conversation we had about your brain and consciousness. And like we see everything you do on your computer is just bit. We see complexity emerge from really simple systems. And I think these tools are also putting this front and center of how do we. It's like a new skill of decomposition. Right. Engineers have always had to be good at decomposition, but now we're learning it in a LLM way. Okay, so is there an orchestrator agent for the conversation that's trying to figure out where are we? What stage are we in? Who do I route this to exactly?

[C]

Yes.

[A]

Okay, so I've got an agent that is managing the flow of the whole conversation. It's decided. We're at Greetings. It tackles the Greetings agent, we move on to qualifying. We probably get into like product exploration. Do each of these agent types have their own unique tool set? You mentioned skills. Tell me a little bit about how these agents are supported.

[C]

Yeah, same way, right? You like the we kind of have we or if we had all the tools for one single agent, it got into a mess and sometimes it was calling the tool that it was not the correct one. This is what we do. We divide again, the one that probably uses the more interesting tools is the one that pitches. And here you can have in the end, hey, I have to pull up this slide. I have to pull up this other slide. I have to pull up this part of the product and I need to have the context of this part of the process. I can explain it. And that's the best one, like the most complex one, I would say. And then you have on the one for the next steps, you have I need to pull up the calendar of this person or I need to pull up the calendar of this other person.

[B]

This is.

[C]

Or I have to pull up the stripe of this specific product. This is the other kind of interesting tools that we have. And then of course we have the knowledge base that I didn't mention about, but the knowledge base we continue using simple drag and embeddings extraction mechanisms. And the truth is that what we found out is one, to avoid complexity, two, to continue with good latencies. The key is cleaning the data and have very clean data we extract from. And that's also A big process that we run there with the creator agent, having all the ingested data that they pass us in PDFs, in slides and whatever, having that into something that is easy to grasp, but continuing to have the quality of the content that we need for each specific part. Is a lot there as well.

[A]

Yeah, yeah, let's. Okay. There's so much to dig into here. Let's talk a little bit about the sales skill because I'm curious about. I have a little bit of a background in sales. I was obviously sell for my own company. I sold. I was a CEO of a startup, so I managed a sales team there. My dad was in sales. I've read the sales books. There's definitely a, like, sales process that I think like all teams follow to some degree. But how a company sells, like what's in their pitch deck, what they're highlighting at different moments, how they qualify is very specific to the company. And so I imagine this is coming from the documents they're sharing with you. How do you manage these two things? How are you deciding what's specific to the company? Are there generic sales skills built into the agent? What's the intersection there?

[B]

As you shared, There are two main things here. One is the generics, the other one is the specifics for the company. So in terms of the generics, it's something we basically add into the prompt most of the times. And then in terms of the specifics, this is where the big part is or where the complex part is better set. So what we usually do is we ask our customers to share with us like transcripts of the calls from their sales teams. Right. And this is transcripts from their SDRs, this is transcripts from the AES and so on. And we ask them for as many conversations as possible. Right. And from there we start the process Kiko was describing before. So we will take the documentation and so on. And one of the documents we create for the. For the conversational agents is what we call the sales skills or the sales guidance skills for the company. Right. What does that document have? And Kike can go more into the details because I'm not in the technical details anymore. But it basically has things such as, hey, what's the aha moments for the company? What's the things they constantly. Or the main value proposition points they constantly go to. And all of those details are contained in that document. But what I would say generally is we try to take the real knowledge and we found that the best way to do that in terms of sales, it's Basically the conversation. So this is how the people in the company are pitching it. And then we try to make it as simple as possible as to pass it to the agent because that's when he really under or talking as a person, it really understands it. But yeah, Kiki can go more on the technical stuff, but that's how we do it.

[C]

Yeah. The key, Teresa, to be honest, is keep it simple. So we extract all the information, as Yuri was saying, from the transcripts, their typical training pitch decks that they have for new employees, et cetera. And from there we basically understand, okay, what we're going to do always is green ins and try to do a discovery. Okay. From there we're going to understand who we have in front of us. Right. And from there, if it's this icp, this other one or et cetera, we're going to follow a certain number of rules. So if this is the aicp, if I'm talking to the HR manager of this company and the HR manager has this pain, I'm going to follow these steps. And we have a tool basically that is okay, this is the person I have in front of me. Give me what I should do to sell or to give the value prop to this person and from them what we're getting is basically, okay, you need to sell this slide and you need to pitch this part of the product and that's what we do. Probably is not the best fit for this specific person. Okay, that's maybe 10% of the times, but let's go for the 90% of the times and avoid all the complexity of trying to make a case for everyone. So this is the way we do it for guidance.

[A]

Yeah. What I like about this is that it's like every head of sales dream. We're going to define a sales process. Here's your sales enablement documents, here's the comment objections always use this script. But no sales team follows all the rules every time. And you're basically building an agent that is going to follow the rules every time.

[C]

Yeah, yeah.

[B]

And what we also do, forgot to mention it, is when we are asking them for calls, it's usually also calls from different people in the team. What we used to do in the beginning it was like, hey, give us the calls, the five best calls from your best ae. So we were basically selling their best ae, right. And there are things that for sure they are the best for something, but then we were not getting the rest of them. Right. So this is one of those realizations that was key here. We need to get into different conversations from different people too. And then the other insight I wanted to share is something that we also like to ask them is, hey, give us the same materials you used to train your sales team. And a lot of times we get like the typical matrixes, spreadsheets and so on with the different ICPs and so on. And we try to codify it into the agent tool and that really helps. But we all. Something that we do that I really like about our product is we always try to come back to our value proposition of hey, this is a teammate, right? So treat us the same as a teammate. So how would you basically onboard one of your sales team members either being an SDR or an AE and give us the same documentation you will give them and we will try to ingest that and basically create one more sales rep inside your team?

[A]

Yeah, this is something I've been learning. So I, my primary day job is I teach product teams discovery skills and I have a wide variety of courses and they're each rubric driven. And now I'm starting to build AI teaching tools. And what I'm learning is the exact same way I teach humans is how I'm teaching the AI. And sales is obviously an area where companies invest a lot in training and a lot in let's teach you how to follow our process. And so I can see how that same exact material could be used to teach an agent to do the same thing. Makes you wonder what parts of human work is not going to get eaten up by agents. I'm both excited and scared of that proposition. Okay, so we've our conversation agent has generic sales skills. You're ingesting a lot of transcripts and sales enablement material from the company and really tailoring it to the way they like to sell. Tell me. I still can't wrap my head around it. Especially when we're talking about voice. You're now doing video generation, like avatar generation. And it's all responding to a dynamic conversation. Like give me a sense of. There's got to be a lot of different models involved there latency. How are you coordinating all of it? Is it like does your customer really feel like this is a real time conversation? Are there things you have to do to make it feel that way? Tell me a little bit about that.

[C]

Yeah, so of course there's margin of improvement, as always. But yeah, the latency is very good at the moment and you feel that you are in a conversation, I would say. So the thing that we also do on the phone and those are Very good. They're not. The latency in the video call is not much more, just a little bit because of the video avatar that is there. What we're not doing at the moment, Theresa, maybe I didn't explain myself correctly, is we're not generating videos in real time for the product. What we have is we have a big library of videos of specific parts of the product and that's what we show.

[A]

Okay.

[C]

Yeah, exactly.

[A]

So the agent could give, from a voice agent standpoint, can have a dynamic conversation, but the video of the product is. It's got a library of clips that can choose to play.

[C]

Exactly. So imagine you have, I know this specific part of the product that is doing this specific part that can be very interesting for this customer. It gets it, it plays it and it tells what's happening in that part of the product and why this is interesting for them.

[A]

Okay, now I can imagine maybe this is going to get into your creator agent. I can imagine over the course of your workflow, which could be days, you have to capture knowledge about the prospect, where they are. Are you building out like a dossier on each prospect to then help drive future conversations? Because I think without that I could imagine an agent just asking the customer for the same thing over and over again.

[C]

Yeah, exactly. Yeah, yeah. Of course we're like keeping the information of the people and that is key for following follow ups. Yeah, yeah, that's key.

[A]

Is that what that, is that what that creator agent is doing is after

[C]

a call it's like figuring out what the evaluator agent.

[A]

The evaluator agent. Okay, let's get into these two. Tell me, what's the job of the evaluator? What's the job of the creator?

[C]

Okay, so the creator is to create the conversational agents. So grabbing all the context that all the documentation and everything that the customer is giving us and creating the prompts of these agents, plus creating the documents that go to the knowledge base of these agents. So doing all that cleaning of data. Okay. And then the evaluator is getting each conversation that had happened. And basically what we call custom fields is basically we evaluate and we say what's the next step? How this one qualified, what already happened and what's going to happen. And like whatever you can imagine. And actually we give our customers also the ability to add more custom fields in order for the evaluator to evaluate every each call. Like for example, I don't know. And what's the role of the person that you talk to? What's the name? Of course what's the email, the phone number, et cetera. But all of that is being captured by the evaluator.

[A]

Okay. And that happens at the end of every interaction?

[C]

Yes, correct.

[A]

Okay.

[C]

Correct. Which also adds something that I didn't mention. Confidence levels in questions that the agent tried to solve. So we give like a first hole of the confidence of in this conversation. How many questions had a confidence level above certain threshold or below certain threshold? And I'm missing one here. Yeah. And frustration from the customer, Frustration from the user. Okay. So yeah, the sentiment, basically this is

[A]

a perfect segue to what I was going to ask next, which is does your agent, especially your conversation agent, but maybe even your evaluator agent have when to hand off to a human Rules.

[C]

Exactly.

[A]

And like I can imagine both, like your conversational agent needs to even know in a real live conversation, like, hey, I can't handle this, it's time to go to a human. And then I imagine that's also something the evaluator is looking for.

[C]

Yeah, we do it like we do it even more in the conversation agent because it's more important in real time. You realize, okay, I'm not being able to answer this question. But it also happens that sometimes it's just trying to answer even though you have the guardrail not to answer to these specific questions or if you don't know about these specific questions that they're posing, just try to hand off to a human. So that's why we have that confidence level for the answers that the conversational agent gave.

[A]

Yeah, yeah. Okay. I want to go back to what Yuri said earlier about this is revenue generating activity. Like it's critical to the company. And so I want to get into. You've clearly built a multi agent like system that is simulating what inbound sales rep is doing. How are you evaluating quality throughout?

[B]

So in reality, when, always when we start working with customers, we always do a poc. Right. So we really work on trying to understand from the customer standpoint what are they trying to validate. Right. And we try to find the best environment as for them to test our agents and our technology in the way that is the least, the least riskiest possible. So what we usually do is we do the rollout plan is not going from zero to one. It's usually going with an A B test sometimes, sometimes it's passing some of their leads, the lowest quality leads through this part and then they start to go farther and farther. So it's really a thing of not just them trying it out and seeing that it works great, but also a thing of them seeing it live with real customers. And this, to be honest, is one of the parts that we are still trying to solve. We don't have the answers to. But yeah, it really takes time, right. And usually it's a thing of one month, two months until they are 100% rolled out on all of their leads. But yeah, it's a thing of building confidence. And then one thing that I really interesting insight here is one thing is that conversations are being really good and that the quality of the interactions are good. And another thing is that your customers know that those conversations are being really good. Right? So at the beginning we just focused on the first one which was like, yeah, all of the interactions are great and so on. And we quickly realized that, hey, yeah, we know they are great, but our customers don't really have the visibility into that. How are they going to know this? So we really try to work on all these reporting skills of the teammate which is basically, hey, sharing the conversations via Slack, logging everything into the CRM and so on, having dashboards with full visibility. And that allowed us to basically for customers to see the quality and also to be able to track the quality of the interactions. I'm more on the how we evaluate quality. On the technical side, I believe Kike will have way more context than me.

[C]

Yeah. Before getting to that, like the success metrics, Theresa is as probably every other human sales man, it's basically conversion and whether you're making those qualified meeting that you promised and whether you're doing the sales that you promised, right. So that's like the high level and important metric that we have with every customer. But of course then we get into the nitty gritty and see what's happening with each other of the agents and how we can improve to impact into those metrics. And for that what we do, we basically do two phases. The first phase usually is around one month where we really need the customer to be involved in the evaluation of what's happening. And we almost like every conversation and depends on the volume, of course. But almost every conversation is reviewed with the customer reviews. Almost all of the conversations that are happening, okay. And they're giving the feedback, okay, so this is happening in this conversation, this is happening. What happens with that feedback? Automatically a test is created and the agent tried to go through that conversation again with that. Like you said this, you need to do this. This is what the customer feedback is. And we run that test until it's passing that test. This way, while we're creating in our conversation agents, we're creating a battery of a lot of tests with a lot of different conversations or partial conversations that need to be passed. And that's how every time we're improving our agents, making sure that we're still passing the tests of the previous feedback. Okay, so this is how we're trying to get determinism of code inside the prompting. Indeterminism of prompting. Right. So that's happening in phase first, in the second phase, usually around that or after that first month. What we're doing is, okay, your customer, you already did a very good job. It's not going to be like this forever. We want to make it low effort for you. So what we're doing is with that evaluator confidence level that it's giving, we're getting basically, or spotting for the customer, which are the conversations that they need to evaluate. And that reduces over time. So at the beginning it's a percentage, and then it continues reducing until it's almost like 5% of the conversations. Okay, so that's how we're doing it. And of course, at the beginning we did a lot manual. We now do a lot less manual. But we still have to do some changes in the prompts ourselves. Our AI engineers are doing those in order to pass the tests that are created.

[A]

Yeah, yeah. Okay. So there's a few things I heard from both of you. One is this real challenge of onboarding a new customer. Right. Like, it takes a lot of work to get to the aha moment of this is a sales agent that's going to close deals. For me, it's almost like they're hiring an SDR and have to train them and get them up to speed before they see value from it. I guess there's good and bad to that. The good is that they're used to it taking time to get value from a sales rep. The bad of it is that's pretty tough from a software standpoint. The other thing that I'm really interested in, Kike, that you just brought up was this idea of these tests. It's almost. I know a lot of people talk about like curated data sets for evals. It's like your sales conversations are almost like that data set of we gotta make sure that as we change prompts, we're not regressing at all. Is it. Do you. I think at the very end, you said your AI engineers are iterating on the prompts based on those tests. So it's not that the agent is updating instructions to try to pass the tests?

[C]

Not yet.

[A]

Not yet. Yeah. Yeah, I can see good and bad in that. Right. It could get real spin off the wheels quickly. The other thing I love in your responses is how customer driven this is. Right. Like customers are providing the eval feedback and that you're able to get it to the point where they do a lot to start. And I know a lot of customer success agents work the same way. Review every ticket over time. You can do fewer and fewer, but it's never going to zero. They're always reviewing some always making sure quality is being maintained. Do you do any like aggregate scoring for them to give them some view of which converse like, how are you surfacing that 5%? What does that look like?

[C]

Yeah, in the end, like we have those scores of quality of conversations and we make transparent the evaluations that the evaluator agency is giving. And you can see over time like how that confidence score of the conversation is getting better and better over time. And that's a good sign. And here like there are two, two main pieces. There's one piece on the guidance, hey, you should have guided this user towards this path and you didn't. And there's evaluation on the knowledge. So you answered this to this question and was completely wrong. You needed to answer this one. The first one on guidance. Usually with the first phase, like this first month, all of it is resolved. But the knowledge is so difficult to have everything covered and you will continue to see that over time and over time. But still getting to that 5% is already a great result. Yeah. For our customers.

[A]

I imagine even with humans that's the case. Right. Like company knowledge bases just get out of sync. And humans tell customers the wrong thing because of it.

[C]

Yeah, totally.

[A]

All right, what's. Well, I think I can guess the answer. But let me ask what's next?

[B]

That's a great one. So first I'd say keep growing the company and this is not technical, but yeah, keep growing the company and keep getting more of our digital sales reps to more customers worldwide. More on the technical end, I'd say that the priorities are like probably the smart orchestrator, right. Which is not like the. These workflows that are predefined but make the agent pick the tools it needs to pick as to achieve the objective. And then also technical, we are working on a do it yourself motion on a free trial motion, PLG motion, whatever way we want to call it. And to be honest, it's super difficult. But that's something we really like because that's Something that would set us apart. So I believe one of the opportunities in the market right now is to try to get a platform where you can basically build an agent that is production ready in a way that it's also fast and that it's delivering results for critical tasks such as sales. And obviously it's really complicated, but we are trying to basically understand what we are doing in the background and how this could be automated. So something we are working on too, and then more on the future. And our vision for the company is to build more digital workers. So we basically started with sales reps because we saw the pain also. We really liked that it was like top line of a company, which is revenue. But customers keep asking us for more digital workers, and it makes sense. In the end, we are becoming like experts in their products and they are asking us to become like customer success managers to do customer support and so on, and they are pushing us to those markets. So that's what we are seeing. But as the journey might change a lot, what we do is basically we are listening to our customers, we are talking to them, and let's see where it takes us. But that's what we see for now.

[A]

Yeah. Amazing. The reason why I said I might guess what's next is I think you're the epitome of figure out what to build by providing almost concierge services and then build it and really just getting close to the customer, having the customer very involved, and it's clear it's driving good results.

[C]

Yeah. And the big challenge here, Teresa, is how do we make it much lower efforts from our customers to build all this. Right. But at the same time, now, all that effort, as you're saying, it's giving us so much insight on how we can improve it and automate it that we believe that we're in the right path and we just have to go step by step and automating stuff until it's almost like you just go talking slack to your new sales rep on digital sales rep and you have the sales rep created and it's working perfectly. But, yeah, we'll get there.

[B]

Something that I would probably add that was probably a lot different five years ago is that I think that in a year there would be probably solutions to some of the problems we are facing today that we can't even imagine of. And this is something that has happened to us through the journey of the company. If you could both have asked us this question, like six months ago. There have appeared a lot of new models, a lot of new Solutions that have solved a lot of the problems we had at that time and we couldn't even imagine they would appear. So this is something that is really interesting for us too. And we think things are going that fast, that we are really optimistic on all of the challenges we are facing because, yeah, we'll find a solution 100% and it will come faster than we can even think of. So it's also a thing of keeping the eyes open and seeing what's happening in the market.

[C]

Yeah,

[A]

yeah. I think this is something we're learning is as this technology moves so fast, like the differentiator can't be our solution today. It has to be our understanding of our customer and their needs. And this is actually a big part of what got me interested in this stuff, is that I think we're going to see discovery just gets more and more important and the solutions are going to evolve so quickly. So the companies that are going to run ahead are going to be the ones that can very quickly match those solutions to a clear need.

[B]

100%. And if you think about it like maybe three years ago, the bottleneck was on the speed of your engineering team on reacting to the insights. Right now, and we have faced this situation, the bottleneck is more on the insights that you want to share. So it's crazy how things have changed.

[A]

Yeah, it is absolutely crazy. All right. This has been amazing. I've really enjoyed getting to hear your story and to dig into your product. It's fun to see patterns emerge across all these products and learn how we're getting complexity to emerge from LLMs, even if it requires breaking it up into lots of simple tasks. And I wish you the best of luck. I actually want to go and see if there's a demo of your sales avatar on your website to play with.

[B]

Yeah, play with it and let us know your feedback. Yeah, it will be super helpful for sure. And thanks for having us, Teresa. Really appreciate it.

[A]

If you enjoyed this conversation, please subscribe in your favorite podcast app and give us a rating as it helps others find the show. Thanks, I appreciate it.