Re-Air: The End of Busywork: How AI Transforms Productivity at Scale with Alberto Rizzoli of V7

Speaker A: Hey everyone. Before we dive in, we wanted to take a moment to thank you for listening and being part of our community. Today we're revisiting one of our most popular episodes in the archives, a conversation full of insights worth hearing. Again, we hope you enjoy it. And remember, you can stay up to date with the latest content and subscribe to the show@datastaxshow.com

Speaker B: hi, I'm Eric Dots.

Speaker C: And I'm John Wessel.

Speaker B: Welcome to the Datastax Show.

Speaker C: This the Datastax show is a podcast where we talk about the technical, business and human challenges involved in data work.

Speaker B: Join our casual conversations with innovators and data professionals to learn about new data technologies and how data teams are run at top companies. Before we dig into today's episode, we want to give a huge thanks to our presenting sponsor, ruddersack. They give us the equipment and time to to do this show week in, week out and provide you the valuable content. Ruddersac provides customer data infrastructure and is used by the world's most innovative companies to collect, transform and deliver their event data wherever it's needed, all in real time. You can learn more@, uh, Ruddersack.com welcome back to the Datastack Show. We are here with Alberto Rizzoli. And Alberto, I cannot wait to dig in on this episode because you've been working in AI for more than a decade actually. And so I think you're going to bring a, ah, wealth of knowledge to our guests. But give us the quick 1 minute background on yourself in a minute.

Speaker A: I'm Alberto Rizzoli, co founder and CEO of V7. I have started my journey in 2015 in AI. So shortly after the first Rusevsky paper that demonstrated that GPUs can train neural networks efficiently. And so I've seen all the iterations of AI hype and then AI hypization and we're obviously in the most exciting time, uh, ever in the field. And I think we have enough of a technology to change the world in the literal sense. And the way we're doing it at V7 is by enabling people to automate the most boring parts of their work. Everything that involves document understanding back office workflows that are slowing down enterprises massively and that usually involve dozens or hundreds of people doing work that they frankly would rather not do and consider to be repetitive and giving them the chance to develop systems to automate that on the run.

Speaker C: Um, Alberto, that's going to be really fun to talk about. So I'm super excited to talk infrastructure with you. We talked a Little bit before the show. There is a wide variety in infrastructure and AI today that people are selling under these same marketing campaigns, and it looks really different. So we're going to dig in a little bit into that topic. What's something you're excited about? Chatting.

Speaker A: I think I'm very fascinated by what the end state should look like. So today we are all dipping our toes into AI workflows within our own businesses. Some people are taking it to the extreme quite successfully. But what I'm interested in is, like, what will the Fortune 1000 look like in five years if they correctly implement today's AI paradigm? And what will the perceived roles that are affected by it need to do to adapt and successfully adapt to all these changes? And then, like, what are the things that AI is still not very good at doing that we probably should not bet our careers on?

Speaker C: Love this one.

Speaker B: Well, let's dig in because we have a ton to cover. Alberto, like I said in the intro, what a privilege to have someone who's been working really at, uh, the forefront of AI for over a decade. It sounds weird to say that because the industry is moving so quickly. But before we dig into your background and then, you know, talk all about Data infrastructure and AI and what you're building at V7, I have to ask more of a personal question. So you've been an entrepreneur since you were 19. You've been working in the forefront of AI, and that's so interesting because, you know, kind of what you've known in your professional career is being your own boss, essentially, and working with the latest and greatest technology. If you had to go get a job outside of tech, what would you do? Like, what interests you?

Speaker A: Great question. I've asked myself that, uh, many times before.

Speaker B: Probably hard to answer after so many decades in the saddle as a, as an entrepreneur.

Speaker A: Yeah, indeed. It's effectively, if I hadn't chosen the entrepreneurial career, I would still be building something. And the careers that I considered when I was in high school was either theoretical physics, which does not involve building something. And so I scraped that because you have to spend 15 years and then maybe you get mentioned in, in a good journal and maybe you make a good discovery and then you're done.

Speaker B: Essentially a teaching career.

Speaker A: Teaching and research. Yeah, yeah. If you can't succeed, you end up having to teach in any academic field. But the other thing I was really interested in was user experience design, specifically in the advent of the personal computer revolution. I think we had to figure out all these things that we take for granted. That have to do with user experience all the way from like the mouse and the graphical user interface. These were brilliant inventions at the time because there was no pre existing paradigm. And why I find AI so fun these days is that it's a repeat of that. We are finding out new ways to interact with machines that have this new paradigm of an LLM or like this token predictor. And I think we're maybe like 20% of the way there in really figuring out what uh, what in which ways we will interact with computers and intelligent machines in the more distant future. So in short, I would have probably been in the product side at a tech company if I had not picked a founder group.

Speaker B: Yeah, I love it. I actually was reading recently, uh, just doing some research on the chat interface and some of the inherent limitations there and other things like that from a user experience standpoint. And one author mentioned that he said something very similar to what you said, which is that LLMs present the biggest user experience shift that we've had in 60 years and a sort of a major paradigm where you're moving towards sort of an outcome based, you know, you're telling the computer what outcome you want as opposed to issuing a command and then getting a response. And that being an exchange where there's direct input from, you know, a mouse or a keyboard or even you know, buttons that you click or other configurations that you make. But you're actually giving an open ended outcome and expressing your intent to the computer, which is fascinating.

Speaker A: It's hard to believe that 60 years ago was the 60s. For a second I thought like, oh, but the GUI was a much bigger fit. But actually it's around that time. Yeah, I think that's precisely it. We are no longer. We also are now interacting with machines that adapt themselves. They're almost like shape shifting software and it's able to adapt to the need or tool that we need to use at the time. And there's an avenue by which maybe the feature of software is software that codes itself and generates what the exact minimum viable code that we need to run to do something. And it's very similar to sort of a science fiction or fantasy magic tool that just turns into whatever we need at the time. Yeah, and yeah, that, that brings up a lot of different possibilities and also a bit more of a winner takes all approach to the world of software, which is a more concerning aspect.

Speaker B: Yeah, well, let's talk about science fiction a little bit because back in 2015 you were working on the cutting edge of computer vision technology and translating Images into, you know, into text, across multiple languages. And you know what really struck me about that? And I hope this doesn't come across the wrong way, but when I was looking at that technology and prepping for the show, in terms of the timeline, it wasn't that long ago, but relative to what we have today, it felt really primitive. It felt like a very early prototype of the types of things we have today. Could you give some perspective on that? Because back then it was really cutting edge. There were very few technologies that could do anything like that. You point your phone at an image, it speaks the word back to you across multiple languages. I mean, that was really wild back then, you know, but now with GPT's real time vision stuff, you know, it does kind of feel primitive. And that happened very quickly.

Speaker A: Very much so, yeah, it happened quite quickly. And most of that progress happened in the last three years. Yeah. So Ipoly was an app effectively to capture images with your smartphone in real time. So think of it as a video, and then every frame of the video was understood by AI and spoken back at you, so someone with a visual impairment could wave their phone around and hear what they had in front of them. And it had a vocabulary of about 5,000 words and objects. And if you were to try it, it would have a similar wow effect at the time, uh, as if you were to use GPT or Gemini's real time vision features today, but only if you were an engineer, because you understood the impressive capabilities required to make something like that work. And the actual proof points is whether your average person on the street would gain a lot of value from this type of experience. Which is why it was successful with very niche group of users, which are people that have very low vision and some enterprise use cases that knew exactly what types of objects they needed to understand. And therefore you could limit its capabilities. But it was a good proof point of why AI only succeeds when it's able to cover a lot of edge cases. A good way to think about AI is that it's always a distribution of data. And the AI model models that distribution. Think of it as a bell curve, where the total number of objects that you interact with on a daily basis is like the very middle of that bell curve. And if you have a product that doesn't capture the edges of the bell curve, which are the least common objects you interact with, you consider it to be a bit useless. In particular, because as the average person, you don't actually care about identifying a glass of water, you care about identifying that one component in your ac, which is broken. And we've only been able to enable that at scale. Before that, you had to spend months of data labeling time and R and D to build a custom model to understand the inner parts of an AC unit, which only made it viable to companies that had research budgets and were really focused on a big problem. That has completely changed. And the main reason is that the AI models are able to vacuum up most of the Internet and then train on them. And we discovered that if we gave a lot of money to a few entities that spend billions of dollars to train big models, we end up creating a better outcome than if a lot of small companies focus on their own domain and create bespoke models for that. It's unfortunately a kind uh, of a power law distribution aspect of the rise of AI.

Speaker B: Yeah, can we dig into the. You mentioned that AI is a data distribution and then LLMs are processing that. Let's talk about the infrastructure in the data layer. I know John has a number of questions here, but I'll kick us off with can you paint a picture of the differences between just the underlying infrastructure and I. It's really funny that you said, you know, this would be amazing, it would be a crazy experience, you know, but only for an engineer. For the average person it's like, okay, I point my phone at a glass of water and it says, this is a glass of water, that's cool. But an engineer would, you know, he would be, you know, amazed by this. Which I actually think as a side note is kind of an interesting description of AI in its best form ever, where, you know, it almost disappears and you have this sort of, you know, really fluid experience. But can you talk about the infrastructure then and what you and your team had to build and the data processing layer versus what you're doing today at Ah, V7? How is this the substrate of infrastructure changed?

Speaker A: Yeah, absolutely. When we first started V7, the majority of our workflows that we built were training workflows. So they were teaching an AI how to identify objects that general purpose AI did not know. And that aspect was essentially capturing thousands of images of unique objects and then loading them onto V7, which would use AI to pre label some of them. So it would segment relevant objects, it would understand what you're trying to identify there. And then having humans complete the rest of the work, add additional tags, adjust the labels, and then a machine learning scientist would train a bespoke model that can understand these objects. Say again, the parts of an AC or the specific components in A document like the shipping address and the consignee address and the unique reference number. And only then you could start using these models. And if you had to change anything in these models, you had to go back into the infrastructure, add additional labels, restart the GPUs and train a new model and then push it back into production. Now that has become a lot more fluid. A user can go into v7, drop the document that they're looking to automate, and simply explain in plain text what are the fields they want to extract. And if there's anything that AI is struggling with, you just explain in English a little bit more on, um, how to find those specific fields or what rules they need to follow in order to reason through, say what is an acceptable termination clause and what is not an acceptable termination clause in a contract. This has changed the user base from being very technical and scientifically minded users to people that are actually doing the job that AI was being built for. And I think that's a massively fundamental shift. And the stack or the infrastructure now has moved away from something that leads you to training a model to something that leads you to just showing an off the shelf model like GPT5, how to do a particular task that is unique to you or to your business. And effectively, we have moved the workflow from the end goal is to produce a model, to the end goal is you start with a model and then given the inputs that you give it, whether they're documents, a recording of this call, what do you want to get out of it, and then you're configuring this workflow so that it can call external apps, send an email, or just understand within this unstructured data context, what do I need to extract and what guidelines and rules do I need to follow to evaluate what I just extracted. This model is a lot closer to either standard operating procedures within a company or a back office. And it enables these very back offices to effectively replace the most annoying tasks with AI workflows. The infrastructure has therefore moved away from the lab and into the hands of general users. And there's many different types of AI infrastructure as well. This is just the one that V7 specifically caters to. But you can think of it as like the next generation of E commerce websites may no longer have a process by which they hire a photographer to go and take pictures of models wearing their particular dresses, but they start by generating them. And then if a product becomes evergreen, then they go through the effort of doing a more refined job, which is getting a real human person to do that task. And whether you like this process or not, from a fundamental ethics perspective of hiring the human to do the job, it is the most likely course of action that we will see. And so there's now AI infrastructure components for almost every operation within a business. And what business leaders need to ask themselves is what are the most likely ones that we can apply today and that will ultimately be fully solvable with AI so that it's worth investing the time and effort into it today.

Speaker C: Yeah, yeah, I think that makes a ton of sense. Um, and I think, and I do think that's going to be hard for people.

Speaker A: Right.

Speaker C: If you're in your example, if you, you know, somebody calls you up and said hey, I know that we normally use you like to you know, demo or try on our new products or whatever and then like we don't need you anymore, you're, we're replacing you with AI. I think that's going to happen to people and I think, and we can, we've got, we're going to dig into that topic a little bit more later. But on the infrastructure side, I want to talk. We've got a number of engineer uh, types that listen to the show. I'd really like to like paint a spectrum from people of. So let's start with the marketing. The marketing says like we have AI agents, we have agent swarms, we have like whatever key buzzword you want to use and it can do all these magical things. Like that's the marketing side of things. But then from an implementation side, from what I've seen and what I've talked to people, there's vastly different implementations. There is some very light handed implementations that are pretty simple workflow tools and you do a little custom prompting and you like, you know, stitch them together. And uh, you know, there's that, there's some very more complex implementations where you've got each component separate and you're doing you know, vectorization and custom tuning of the models or maybe building your own model. It'd be fun to kind of talk through your thoughts on that. So for, so when you talk to companies like I imagine a lot of these people, a lot of people that are kind of into AI, like are aware of some of the like consumer grade tools, like an N8N for example and not that like you, you can use it, you know, not that companies don't use it for some things, but I'd love them to talk through that tooling world from like a uh, you know, end user that's kind of like getting into AI versus like true production AI, like at scale for some of these like Fortune 500 type companies. And what's the difference? And what should people think about? When should they know, like, hey, I'm hitting a limitation of like my, you know, alpha project or my little, you know, project that I'm working on the side.

Speaker A: Yeah, great question. So I think the simple stuff will eventually be eaten by the LLM providers themselves. Yeah, and simple stuff is your typical like zap from three years ago, which was pull data from this app, apply some heuristics, push data into this app. And that is actually a lot of, a lot of tasks that are being done today. And just by nature they exist because they're simple to configure. What's much harder to configure is build a system that helps us decide whether to buy a building or not. Or like uh, an asset manager, which normally is something that you give to an analyst that is paid quite a lot and they have all this inherent information about how to do the job, what the company needs, what are the ways it evaluates a particular property and ultimately like who to go to if there's a decision to be made. And this just expands the number of nodes by quite a bit. And the infra ends up looking like you still have inputs which are usually a document reaching you by email or a person just saying evaluate this to an agent and dropping in a PDF. Then there's all these reasoning steps in the middle, which is the, probably the biggest value driver at the time, both for V7 and I think of AI in general. And then there's outputs. And a lot of startups tend to focus on either straight up output generation or making it easy to like zapier to take in inputs and then push out outputs because it's very amenable to consumers and it's a low hanging fruit for a lot of small companies. But in an enterprise it's very unlikely that you have like an ABC workflow. You have a lot of steps in the middle before an output is generated. And that output can be an IC memo telling people, hey, we should buy this building. It can be a uh, slide deck or it can just be a decision, yes, no, or like let's update the CRM to say that this is a deal to be made. But what happens in the middle of the sandwich is really complex. There's knowledge bases that need to be queried, there's chain of thoughts that need to be developed. And some of these are inherent to the process of the company. And some of them are crutches to just the way LLMs work and making them work more efficiently. The former is the equivalent of memory. You have a vector database, you pull information from it that helps you, uh, understand how to solve a particular task or subtask. The second is about breaking any reasoning task into a smaller subtask. And the reason why this is done is because LLMs are largely trained on question answer pairs that are part of short conversations. It's unlikely that there's a lot of training data in these models. That is of the nature of here's an offer memorandum for a building. Think like an investment manager and tell me whether you should buy it or not. And then it ends up producing a 200 page thesis document. Very unlikely. And so the best way to do it is to break it down into individual steps where every step is effectively a prompt and expected response with customized inputs. The input usually starts with the core document. As you move along, you may use some of these previous outputs like hey, let's evaluate the financial performance of this building from the past five years. Then you use that financial performance analysis as a downstream input to another property, in this case not a building property, but a data property to then at the end come up with a final decision or output after all this like forced reasoning and forced filtering. And if you don't force it, you probably are not going to get a really good result. That is up to spec to, you know, what a Fortune 1000 will do. The reason why we like these reasoning models so much is that they kind of do the reasoning flow for us and for every topic that they are smarter than us at or better than us at. It feels really good, but it's very frustrating. When you are a lawyer, you ask a legal question to the model and it starts making all of these assumptions that are not grounded in the way you actually do the job. And so a lot of this infra is about actually grounding the model into we are this business and this is how we run this specific process. If you follow these steps, you will be successful.

Speaker C: Okay, that makes a lot of sense. So if it's a simple. So let's think about like your original example. I'm using Zapier to, you know, currently to do this thing to like tag all my emails based on deterministic rules. I want to use an LLM to make this a little bit better. Now I'm going to use maybe Zapier to better, uh, categorize my, you know, email, very simple type task. So I Guess, I guess from there what I imagine other than the technical side. So we're going to kind of get into this like AI implementation, you know, topic a little bit. I imagine that a lot of the hard part here, there's some hard parts with like, you know, tuning models, getting context. Right. But a lot of the hard part here is still working with the humans to have SOPs of like what needs to get done anyways. If you've ever sat down with a team to like develop an SOP or even we were talking before the show about like clear objectives. It's actually a fair amount of work just to like map out a clear objective and like the let's use okrs, for example, like a clear objectives and key results. Like that's actually a lot of work to even do that. So I'm curious about strategies like maybe that you guys have developed as you help people think through. Like what are we actually doing here in a systems thinking where there's a bunch of people involved and it's not

Speaker A: just one person's workflow, it's universally hated and the rules of construction are written in blood. And if you speak to anyone in the construction field, they're intelligent people who just want to do the thing the way they know how to do the thing. And usually processes are only documented when something bad happens and someone reacts by saying, okay, let's implement a way to do this properly every time. Somewhere there's also an insurance company that enforces it so that you can get covered. They still remain universally hated exercises. And I think one way to hate it less is that as you develop the sop, which we see it just as a list of prompts, effectively you're actually seeing the work get done in real time. So one way we sort of solved it is that as you develop an agent in V7GO, you're effectively adding all these rules in English language, something that looks a lot like a standard operating procedure. And every time you add something it recalculates your workflow and then you can see what the end results end up looking like. So you're almost like simulating the job over and over until you're happy with covering all the edge cases through the sop. But the problem remains that almost no process, especially in enterprise, is known by one single person. And we, we're sort of like biased by usually being in startups or being in small businesses where you have that one person that kind of knows everything and knows how everything should be done. But companies are a bit more like asml where or large companies where like there's no person that actually knows, sorry, uh, a TSMC more so than asml, but probably both of them. No person actually knows how developing chip actually works end to end. It's sort of like this wisdom of many people that know their specific niche and it's quite similar in, even though allegedly simpler in enterprises. And the difficult part is to get the right people to put the work into documenting how their specific slice of the pie works. And I think the only way to do it effectively is to give an incentive for people to just go into a platform and document this thing as a set of prompts. And I think the wrong way of doing it is to have effectively someone that goes and does all these user interviews like your typical for deployed engineer until they build a full mental model of the system. Yes, it works, but it doesn't work organically. And I uh, don't have a full answer on what the best organic model for it is. But I think there is a future by which any company's biggest asset is its knowledge base. And it is the aggregation of what AI has been able to observe across any of our work streams and document effectively their own opinion on how processes are run. And that can be used as a knowledge base for running any other process. Hypothetically, if you had an AI that was monitoring your laptop 24 7, I don't think that is a viable solution. But hypothetically he would be able to keep writing to a document to say, this is how John does business. And then if it does it for John, for Eric's, for Brooks, for everyone, then over time it starts to develop a pretty good thesis of how processes are run today. You kind of have to force that process it, you kind of have to get the humans to go and do that documentation.

Speaker C: Yeah, yeah, well that's exactly what it, where I was kind of going with this is, you're right, like people, most people really hate making, stopping, slowing down, making sops, making documentation. Most people really hate it. My question was going to be along the lines like, well, you know, how do we use AI to do that? And to your point, like maybe there is a future state where you can essentially have the AI, you know, recording everything. Right, but in the mean, but that's probably not practical yet. So in the meantime, I saw an article this weekend and this is what made me think about this topic, uh, where there's some of these robotics companies targeting smaller manufacturers where you're able to essentially record like with your hands, like what you'd like the robot to do and then the robot like learns from that versus having to program the robot with code. What do we have that exists today that you think is useful with that same maybe mechanism with an LLM other than like a prompt? Obviously. But yeah, there are there some like similar things that would feel a little more natural to people, you think?

Speaker A: Something that we tried is to use screen recordings as training data effectively to turn it into prompts. And it works surprisingly well. But it doesn't cover every edge case. But it is a good way to like we do have some customers that come to us say we don't have a documented process, but we have a lot of training videos on how to do this process. So we put the training videos through an LLM and then through that we generate effectively a JSON that can be imported into V7 to turn to create a workflow. And that works surprisingly well. And maybe like the closest proxy to that in uh, in the research side is that computer use within LLMs is largely being taught as this like set of steps that a human is doing. Like you know, order a burger on doordash depending on where my address is and here's like a list of my preferences and a human would go and try to fulfill that specific task and create training data for it. I think it's a good way to do that, but it's still, I don't know, I think it still doesn't solve the problem of not everyone knows how a specific process is run. But one thing that we do at D7 internally a lot is we use loom to teach a man to fish for almost any process. So if someone asks like hey, how do I register time off or request a budget to do this particular event? Then we just send them a loom on how to do that. Sure. And that that very loom could be used to just teach a model that can then straight up answer or do that task.

Speaker C: Yeah, yeah, that makes sense.

Speaker B: Yeah, it's super interesting because the. I agree that we're in the early stages of. Is there a uh, really, you know, sort of elegant, you know, one click way to do this? You know, that's still to be seen however, as I think about the types of data, or let me be more specific, the data formats in which those processes exist or in which they can be extracted. Right. So we have a loom video, we have, you know, Notion or Google Docs, you know, that someone creates to explain the process step by step. You could even, you know, have someone explain it in Spokane Word and all of that's, you know, sort of, you know, longer form unstructured data in various different formats, which is, happens to be like a, uh, you know, a marvelous input for a large language model. And so I'm pretty excited because I think that's going to rapidly become easier and easier because the formats are not a problem. Whereas before that was a major, you know, super expensive technical problem. And that's not the case anymore. Right. And so I'm pretty excited about what we'll see happen in the next couple of years.

Speaker A: Yeah, 100%. It's like the field of meta learning and research is like learning how to learn might be the next, the next big hill decline for any of these models.

Speaker B: We're going to take a quick break from the episode to talk about our sponsor, rudderstack. Now, I could say a bunch of nice things, as if I found a fancy new tool, but John has been implementing rudderstack for over half a decade. John, you work with customer event data every day and you know how hard it can be to make sure that data is clean and then to stream it everywhere it needs to go.

Speaker C: Yeah, Eric, as uh, you know, customer data can get messy and if you've ever seen a tag manager, you know how messy it can get. So rudderstack has really been one of my team's secret weapons. We can collect and standardize data from anywhere, web, mobile, even server side, and then send it to our downstream tools.

Speaker B: Now, rumor has it that you have implemented the longest running production instance of rudderstack at six years and going, yes,

Speaker C: I can confirm that. And one of the reasons we picked rudderstack was that it does not store the data and we can live stream data to our downstream tools.

Speaker B: One of the things about the implementation that has been so common over all the years and with so many rudderstack customers is that it wasn't a wholesale replacement of your stack. It fit right into your existing tool set.

Speaker C: Yeah, and even with technical tools, Eric, things like Kafka or Pub Sub, but you don't have to have all that complicated customer data infrastructure.

Speaker B: Well, if you need to stream clean customer data to your entire stack, including your data infrastructure tools, head over to rudderstack.com to learn more. Alberto, uh, let's dig a little bit deeper into V7 in this context. So you've brought up a couple of use cases, but I'd love for you to talk about what are the examples of where V7 is creating just immense, you know, sort of asymmetric leverage for a business. And I know you you work in multiple verticals. I'd love a couple of those examples. And then follow that up with, you know, here's why we do not, here's what we're not building because we don't think that V7 could be really great and create asymmetrical leverage, you know, for these other use cases. Excellent.

Speaker A: Yeah. So we work in several industries and maybe in order of size of business it's financial services, insurance, real estate, tax and legal. And then a long tail from there that includes logistics. And what all of these industries have in common is that they handle a lot of paperwork. They have a general idea of how the paperwork should be handled and they uh, and very few people at these businesses enjoy that specific task. And the use cases themselves range from high complexity, high risk use cases that we can handle very well. For example in finance, it's a confidential information memorandum which is basically a pitch deck for acquiring the business. Requires five to seven hours to process. From an analyst perspective to turn it into an IC memo or like something that gets given to the investment committee and an agent built. Ruby 7 takes about 15 minutes of processing time, including some review time by the human. So that's a 15x improvement on the actual time dedicated to it all the way to much higher volume use cases like understanding tax claims, understanding logistics slips or even insurance slips, which usually take you know, five to 10 minutes to process manually. But they are huge in volume and they're usually a big cost carrier to the actual service that the company is fulfilling. I'm calling from the UK and the NHS spends about 40% of its budget in administrative costs. That's our healthcare system. All these administrative tasks by and large can be taken care of by AI to reduce that timeframe quite significantly. So there's a big cost saving in general and we're doing things rather inefficiently today when you imagine that AI is possible. But on the other side, I think one of the biggest points of our OI is not easily measurable and it is the lead time between uh, an intent and getting a response back. So a good example of that is that we've all been in a pinch and needing something to be reviewed by an expert. An NDA could be an uh, easy example that many can relate to. If you're doing a deal. 48 hours is a huge timeframe and like 24 hour turnovers for NDA review is already pretty darn fast. But if you start to consider almost every process in a company to now be a, uh, five minute turnaround Time then the full aspect of work and productivity changes. For anyone who perceives this lagging indicator. And if you imagine a business where almost every process requires two days to get back to you any IP request, any. And that's fairly common outside of a startup where you can tap someone on the shoulder versus a competitor who has this system that is just as reliable but it only takes five minutes and that you get your response back, that becomes a massive quality of life improvement to work that I think will encourage people to. It's not really necessarily about doing a lot more of the same thing, but being able to do more things because there is a lot more room for interpersonal tasks and creative tasks. And maybe to answer what do we not do particularly well are just those that we have that I've described. We don't want to get into the way of human creativity and human relationship. I think fundamentally I would love for a very smart 11 Labs agent to answer my customer support queries but that's generally because I don't need to develop a relationship with customer support. But we don't want to get in the way of humans developing relationships with their own customers or with their own people or getting in the way of what is truly creative work. Whilst AI is amazing at producing creative outputs from art to images to poetry, it is that good because it's read a lot of it and in a way it stunts our true creativity because it's not that good at creating something completely net new. And I think a lot of creative aspects is about wowing the world by creating something entirely net new. Even if it's like designing a new way of doing something, designing a new process. It can be a useful copilot to validate your biases and assumptions. But I think that is the work of the future is designing workflows in the most efficient and fun way possible and the execution of them will be what we use machines for.

Speaker B: Yeah, that makes total sense. And to drill in a little bit on something you said, the difficulty in measuring the ROI of, you know, I mean, okay, if you just look at the raw percentage of. I go from, you know, a two day review on an NDA or a contract with a vendor which anyone who's been in a large organization and has had to go through that, you realize that there are, you know, there's this sort of class of people where they're really good at their job, you know, in terms of just the raw skills that were on the job description. But they're very effective at their job because they know how to navigate the system more, uh, with more agility than other people, right? Oh, hey, I know this person down in legal and I did a favor for them or kids play soccer together or you know, whatever it is, right? I know their schedule really well, you know, or I've done this a hundred times and so I know the process. And so they're just able to get stuff done more quickly. Right? But that's a horrible use of their time, that's a horrible use of their like, energy is trying to navigate a really complex system to do something, you know, that could be done in five minutes. But I think the reason that's hard to measure is because of the compound effect, right? Like in an isolated one time thing, it's like, okay, well that's great. Like my week at work was better. But if you multiply that happening a thousand times a week across the entire organization, the time you get back, the energy you get back is it really compounds, right? That's really difficult to measure. But I think to your point, like, well, great. If I don't have to spend, you know, eight out of the next 16 working hours over the next two days, I just slot that off for navigating the system. It's like, well, I have that time back, right? It's like, well that's really interesting 100%.

Speaker A: And I think, uh, like if you think of about a sales career, the best salespeople in the world embodied that they find other avenues to remove friction from their ability to close a deal that are not the standard operating procedure. They meet with their clients, they, you know, remember personal details about them. They put like, they do extra work for them that goes outside of what one would normally do and then they eventually become very successful. And the parts that they don't want to do though are like, I need to wait two days to get this deal approved or this statement of work approved. Because, uh, and like the way they go around it is to, I don't know, charm their way through it, to go and bother people, to go and push for things. But these are all things that AI is not very good at doing, very good at enforcing itself. Yeah, yeah.

Speaker B: Well, we're close to time here, but one thing I want to get your thoughts on, especially as you're implementing V7 in these companies, you're removing these barriers. Who is most impacted and what types of roles are at risk. I mean, you know, let's talk about the salesperson that you just mentioned. You know, they're really good at navigating the system. They're really good at, you know, figuring out how to, I don't want to say cut corners, but they know the process well enough and they've built up enough institutional knowledge of how things operate in the organization to where they can just get things done more quickly than other people, you know, and unfortunately, the cost of that is whatever, 40, 50% of their job is navigating, not actually selling or building relationships for that person. This is going to be, you know, you know, sort of like a miracle drug of productivity. But they were already super productive, right? Do you see that being the case where you sort of have this subset, let's say, you know, half of a team, you know, over indexes for being productive already? Are they the primary beneficiaries of this, you know, who's going to get the most asymmetric value? And I guess I say who's going to. Who is, you know. And how does this play out practically on a team, you know, at a company that that V7 is implemented at?

Speaker A: Great question. I think that there's effectively three groups of people in the world that follow any process. They're the ones that have a different way of seeing the world and it is very hard to replicate it. And they're almost like an artist and an operator. And that would be the example of a salesperson that figures out that, uh, their biggest client goes to a specific golf course, so they learn how to play golf to get there. It's a bit of a crazy idea, but they're thinking out of the box. Uh, then there's like efficient, or maybe sometimes less efficient operators that are following the rules and they're doing quote well. And then there's folks that understand the actual operation at first principles to the point where they could teach it. And the ones number one and three are the most secure. Number two is the least secure. Because they are, no matter how efficiently they're following their process, which could be developing leads and then doing sales, if they're not adding something super new to the process each time and really thinking outside the box, they're at risk of automation. The ones that really understand the process, the first principles have a good opportunity of actually developing those automations. Software is becoming very easy to use and so they could be the ones that can actually automate. Group number two. Group number one. The creatives will likely not be automated for a very long time, if not ever. And a good proxy to this is manufacturing. We had only artisans in the past. Today we have people that design manufacturing processes. They help build factories, sometimes they're engineers, but sometimes not. We have super skilled creative artisans that will create something that a machine could not make. And they just have this very unique way of seeing the world. And then we no longer have artisans that do something similar to what an automated process can do, at least in modern countries. So, yeah, I think effectively, like, you're either a crazy person that just does things completely out of the box, or you should start to think of how to turn what you're doing on a regular basis into a machine.

Speaker B: Yep. I love that one thing that we've talked about before, John and I have actually talked about this and sort of explaining, you know, different types of people within an organization and kind of how to navigate things. But this is a dramatic oversimplification. But, you know, one way you can look at people within an organization is people who create information and then people who move that information around.

Speaker C: Right.

Speaker B: And if all you do is move other people's information around, I agree. I mean, you should be worried, right. If you can't teach, uh, like you said, they have an understanding of the process to the point where they can teach it, but that's actually creating, you know, new, valuable information. So that's. Yeah, uh, yeah, super interesting.

Speaker C: And I've looked at it. I mean, very similar framing, but essentially two. Two jobs, and they definitely overlap. One is that like, creativity, creative endeavor job, and one is that like, continuous improvement job. And you can do both for sure. But those are. Those feel like the, uh. And those are not specific to like a domain. Like, sure, you'll be applying that in like, marketing or operations or whatever. But I feel like that's what I've told a lot of people. Like, if, like continuous improvement, and that's like you. If you've got AI that can do this, then like, how can AI do it better? And then for their creativity, like that, you know, has a million different angles. You can take it. But other than that, like, I think there's a lot of change.

Speaker B: A lot of change coming. A lot of change coming. Well, uh, Alberto, we're at the buzzer, as you like to say, but one more question for you. What is like an app or a product that you've, uh, used recently that's maybe new to you, that really caught your attention. I'm just always interested for people working this stuff every day who are building it at a foundational level. And, you know, especially with your interest in user experience, is there something you've used recently, even a physical product or even an experience you've had where it Just really caught your attention.

Speaker A: I have, I think like many people have started to use video models more and it's sort of uh, as someone that's been in the research side of AI for some time, it kind of turns into a light bulb moment where you realize that a next frame prediction engine can do a lot more than just create videos of the Bigfoot, but it can actually create full demo experiences down the road. And that could be a nice little game changer that are also interactive. On the interactive side, I forgot what DeepMind's name for those virtual worlds but there's many startups that do it. One for example is Odyssey Systems and it's effectively a uh, navigable 3D world that is entirely AI generated on the floor fly. And you can think of it as like a video game that, that continuously renders new things on the spot. And those will be very useful for training certain models that need to be embodied like self driving cars and embodied robots. But I think they will also have a tremendous impact on entertainment on VR. And their genre is quite fun because they're exactly like the human mind. We're imagining a world in our heads and then we're navigating through it and we use that prefrontal lobe experience to think about whether our decisions are going to be good or bad. It's actually what differentiates us from a lot of animals that we picture scenarios in our head and we imagine uh, like if I do this is going to happen, if I do that it's going to happen and we can do it for quite a long period. And so it might even give AI a much better way of reasoning that is not like just spitting tokens out and talking to itself. And yeah, I uh, would say probably these are the two most exciting that are completely outside of our domain because we don't really deal with the generation of image assets. Another thing that we are experimenting with a lot that I'm having fun with is when you look at V7 it looks a lot like a spreadsheet, but it's actually a database. It's a big table and every row is an agent run and various steps that it's taken and information that is written. Yeah, we're experimenting. What would that look like if it was actually a true spreadsheet where you're not bound to a pure grid where you have columns and rows, but rather something that's a bit more fluid? And I think it's the perfect scratch pad for AI to operate in. But it's a thesis that we have not yet Fully validated. But I think of all documents that are going to change quite significantly with AI, I think spreadsheets are going to be the one with the biggest impact because you're no longer doing regular expression. You're doing AI functions and app tool calls and running agents within a cell. So it comes back with a particular piece of information. So there's a lot of UX changes that are going to occur there.

Speaker B: Yeah, yeah, it's super exciting. I heard a product leader one time say, you know, you could essentially boil all software down to like forms on a database. You know, that's basically what software is, right? Yeah, uh, for sure. Which, you know, you're like, whoa, that's overly simplistic. And then you think about it and you're like, yeah, that's actually true. Right. But I think what's interesting about what you're saying is that, well, and the point the guy made was like, so it seems like you could, you know, like, well, if Salesforce is just forms on database, like someone's going to build another Salesforce, it's like, no, the business logic is where they create the lock in.

Speaker C: Right.

Speaker B: And so I think what's really fascinating about V7 is you're abstracting the business logic, you're abstracting that out and turning that into an AI layer, you know, on top of the database. And so, uh, it's going to be really interesting to see what happens the next couple years.

Speaker A: Yeah, exactly. That's one way to explain it to an engineer. It's. We are a multi database system like Salesforce that is built for AI, uh, from the ground up. And as we built this thing, we started to realize that there's maybe a chapter beyond this framework because of AI's capabilities to break out of the mold of an SQL table.

Speaker B: Yep. I love it. All right, well, uh, Alberto, thank you so much for the time it flew by and we'd love to have you back on soon to hear about what's new with V7.

Speaker A: Thank you. Thanks for having me.

Speaker B: The Datastock show is brought to you by Rudderstack. Learn more@rudderstack.com.