The start-up fixing AI’s errors, with Invisible Technologies’ Francis Pedraza

FRANCIS PEDRAZA: So, about a century, more than a century ago, Henry Ford was innovating on the production of physical objects through the invention of an assembly line.

And there was automation, but the automation was in the augmentation of humans — right, there were humans in those factories. All the tooling and the process flows were to increase the productivity of those people. The same is now happening with AI. If you’re in the paradigm of AI completely eliminating humans, you’re actually misunderstanding the industrial application of AI.

RANA EL KALIOUBY: Ford created the assembly line, and now Francis Pedraza is digitizing it. His company, Invisible Technologies, is using a combination of human and AI workforce to help enterprises scale.

PEDRAZA: I think that you always sort of need the human in the loop. You always need some sort of human to QA the model and to train the model to get better. And without that feedback loop, it’s just talking to itself.

EL KALIOUBY: Today, Francis and I are talking about digital assembly lines, the roles people play in powering the AI industry, and why he’s made the decision to buy out some of his investors.

I’m Rana el Kaliouby and this is Pioneers of AI, a podcast taking you behind-the-scenes of the AI revolution.

[THEME MUSIC]

Hi, Francis. Welcome to Pioneers of AI.

PEDRAZA: It’s so good to be here. Thanks for having me on the show.

EL KALIOUBY: Before we dig in, I am just curious about your lifestyle. I know you’ve spent a lot of time traveling and living in various places around the world, including Bhutan. Does that give you a different perspective on your business? Like how is that kind of lifestyle?

PEDRAZA: I think like most founders, I spent many years just in my apartment looking at a screen. And so I started my career in San Francisco and then I moved to New York. And then, a couple years ago in early 2021, when I hired the first CEO from the outside to come in and I became chairman, I was thinking about my role and the future of the company.

And also what I wanted out of life. And it felt like one of the main things that I had traded off in being a founder was travel. And so I wanted to experience the world and also grow my network. And so I took my first vacation and I went to Bali and Thailand and then I traveled throughout Latin America and Europe. And now it’s just pretty fluid.

And I think entrepreneurs are pretty comfortable with the gray area between what is a work relationship and what’s a personal relationship. And if I look at the last decade of building the company, all of our best hires, best clients, best investors have all come through personal relationships. As I’ve traveled, I’ve kind of become like a cat. I can land on my feet anywhere I go — like, where can I get a healthy meal? Where can I get a workout in? And now I’ve got friends everywhere.

Copy LinkHow a digital assembly line rethinks enterprise operations

EL KALIOUBY: Yeah, I love that. That’s like hashtag goals. All right, so let’s get into what Invisible does. You’ve coined the term operations as a service and you also kind of liken what you guys do to Henry Ford’s industrialization of manufacturing. So tell us exactly what Invisible does and yeah, let’s dig into it.

PEDRAZA: Sure. So the metaphor that we’ve used is a digital assembly line, and we thought of the company as building the first digital assembly line. And the way it works is we break processes into little steps like Legos. We’ve integrated hundreds of AI and automation and software tools into our process builder software.

So we try to automate as many of those Legos as possible. But you can’t fully automate every step in a complex process. And so for the Legos that can’t be fully automated, you need humans in the loop. And so we have 6,000 contractors on our assembly line. These are PhDs and masters in almost every subject, speakers of almost every language. And so that allows us to deliver an end-to-end run and solve the last mile problem.

EL KALIOUBY: Define the last mile problem for some of our listeners who may not be familiar with that concept.

PEDRAZA: For sure. In shipping and supply chain, when Amazon wants to deliver to you, they now have this incredible fleet that allows them to go to your door and deliver that to your door. And it’s very difficult to solve for last mile. And that’s a big part of their advantage as a company.

In digital operations, the last mile is whatever software can’t do. And so for the last 25 years, almost every enterprise software company that has come out of Silicon Valley has had one business model: software as a service, SaaS. The first SaaS company was Salesforce, which was founded in 1999.

This has created a world where, from the customer’s point of view, everyone is selling tools. Nobody’s selling solutions. Imagine if you wanted a cake. Silicon Valley is not gonna sell you cake. They’re gonna sell you all the tools you need to make the cake.

And they’re gonna say, go make the cake yourself. Because Silicon Valley has not wanted to solve the last mile problem, because the last mile problem is way too messy. And why is it so messy? Because organizations have humans in them. They have unique processes, they have unique problems, they want unique capabilities. There’s politics involved. It’s all very custom, and therefore they assume — we have assumed — that the only way to solve the last mile problems is to become a quote unquote shitty services business.

EL KALIOUBY: Mm. Mm-hmm.

PEDRAZA: So, by the way, shitty services businesses can be incredibly valuable. Accenture is a good example. They’re huge. These enterprise services companies have turned their job into implementing enterprise software. And so as an enterprise services company, they’ll integrate enterprise software and they’ll tell their clients, hey, we’re Accenture. We will help you use the latest SaaS tools and help you implement them inside of your company. But the reality is that Accenture is not a tech company and Silicon Valley has just not built services companies for the last 25 years. So in comes Invisible.

You’re not a SaaS company. We’re not a SaaS company. So they assumed that we were a shitty services company and not a tech company. And so the business model that we’ve innovated, we call the AI process platform, and an AI process platform you can think of like a triangle. Where there’s a horizontal platform, and that is very similar to SaaS — it is software and it’s built to abstract the problem and it can be reused in many industries or functions. But then to solve the last mile problem, you have forward deployed teams of engineers and field CTOs that go into the client’s companies and solve the messy problems, and then they build vertical specific business applications. The AI process platform business model is built to handle that messiness. And therefore it’s much more useful to most customers. They get the cake. They don’t just get tools.

Copy LinkWhy solving the last mile matters more than selling software

EL KALIOUBY: I love that. So let’s take a few examples to bring this to life a bit more. So one of your clients is DoorDash. And for a little bit of context, during the height of the pandemic in March of 2020, restaurants were scrambling to get onboarded on the DoorDash platform and you guys stepped in. So maybe you can walk us through how you helped DoorDash get through this challenge.

PEDRAZA: I remember getting an email from Tony Shu, the CEO of DoorDash, saying we really need help because all of the traditional enterprise services companies have shut down their offices. Because they had physical offices and we had a digital assembly line, so we had no physical offices. So we were ready for business.

And he said, we need help digitizing restaurant menus, and these restaurant menus can be a photograph of like a Korean menu or a spreadsheet of a pizza restaurant’s menu or a website link to a sushi menu or what have you. They come in all sorts of different formats, and you need to use a variety of approaches to figure out how to automate or streamline the process. And I’m gonna make up numbers. But that was a perfect example of something where no single software tool could solve the problem. It’s outside of DoorDash’s core competency. They need to focus on their overall marketplace, their overall platform, not solving this onboarding process of onboarding restaurants and vendors. And this was one of like 40 processes.

EL KALIOUBY: Right. Because if they wanted to really build it in-house, quote unquote, they would have to build specific software tools and they would have to hire a team that would be responsible for kind of doing all of this. And it’s just not their core competency. It makes zero sense for them to invest in that area. And I guess this is where you guys came in.

PEDRAZA: That’s right. And so most companies have to waste so much time and energy building functions or capabilities that aren’t really core to their platform. And this is a classic example, like DoorDash was not founded to go digitize every restaurant menu in the world.

They were founded to build a delivery network. And so that was like a side mission, but if they didn’t solve that problem, they wouldn’t have been able to scale. And so for us, being able to take the competency we built for DoorDash, we were able to build similar capabilities for Uber, GrubHub, Delivery Hero, Bolt, Rappi, Toast, and that vertical was like our first big enterprise vertical. And then it set us up to go hit a bunch of dominoes in other verticals as well.

EL KALIOUBY: But Invisible Technologies isn’t a company that’s just trying to streamline ordering on your favorite delivery service. They’re also working with some of the biggest AI companies to help train and validate their models. Basically – they’re the team mitigating hallucinations and ensuring the accuracy of models like ChatGPT.

We’ll get to that after a short break.

[AD BREAK]

Copy LinkHow human experts help train better AI models

You have another business basically under the Invisible Technologies umbrella. I call it like the behind the scenes of AI. Like you actually partner with companies like OpenAI, Cohere, Nvidia, where you validate the foundation models and the LLMs that we’re all using today.

And for our listeners who are not familiar with this process, these companies are training these models, but they have to test them against hallucinations and bias and whatnot. This is where Invisible comes in. So I’d love for you to kind of take us behind the scenes and tell us more about how you work with a company like OpenAI.

PEDRAZA: Yeah, in early 2022, OpenAI came to us with a training challenge where, if you think about how do you make an AI model better, there’s your software engineering talent and the core platform that you’re building. There’s compute power, but then you can also test and train the model after it’s in production.

And this is where I’m sure most of your listeners have seen the models get better over the last few years. But we all remember some of the funny answers we got in the beginning, where you could convince the model that two plus two didn’t equal four, or it would just get some facts wrong and would state them as if they were true. So we were able to rapidly hire PhDs and masters in almost every subject, including very niche subjects. So if you wanted to train the model on 19th century French Beaux-Arts period art, how do you do that? You have to hire an expert.

EL KALIOUBY: Right. Because for everybody else, you wouldn’t know if the answer is correct or not. Like you really do need that domain expertise. Yeah.

PEDRAZA: You think about these general models. They have to be smart at everything. They have to be good at answering questions about financial modeling. They have to be good at American history. They have to be good at physics and chemistry and like every subject. And so the processes we developed in partnership with their research teams were processes that thought of more and more ways to break the model, make it answer incorrectly, and then grade the model, and then provide alternative responses. And then think through what inputs might — what data sets or what other inputs might result in the model having a better outcome on the data sets side.

Basically at this point we’ve downloaded the entire internet. We’ve trained our models off of all publicly available data. And so once you’ve run out of publicly available data, you realize, well, we still need more data, more training data. And so AI training is the industry that has basically been birthed in the last three years.

And Invisible, Scale AI and Turing are in a kind of oligopolistic market structure where most of these foundational model companies are using us in some combination to solve their training needs. And so if you’re bullish on AI, you should probably be bullish on AI training. And this creates a yin yang to our business where we are both training all the main AI models in the world and we’re helping enterprises actually use them in their business.

Copy LinkWhat it takes to build and manage a human in the loop workforce

EL KALIOUBY: It’s very cool. So I wanna actually talk about the humans in the loop. Because you’ve already kind of alluded to this massive workforce of human agents that are involved in these two parts of the businesses, whether it’s building the AI or helping enterprises AI-ify their companies. How do you find this workforce? How do you keep them motivated? What does their everyday look like?

PEDRAZA: I think these are the jobs of the future. We had 366,000 people apply to work at our company last year. Yeah. And I think we hired a thousand. So there’s a huge amount of supply relative to demand. There’s a lot of referrals. Last I checked, something like 50% of the people that we end up hiring are coming from a referral from another member of the team. And then we have an outbound process that’s very specific. So let’s just say, for example, that Anthropic’s model Claude is very good at Tibetan language work. So let’s just say they hired a bunch of Tibetan speakers. How do they do that? That’s an example of like a specific outbound process.

By the way, I’m not claiming that we did that work with them. I’m using it as an example. You end up needing to do very targeted searches. So maybe it’s reaching out to the physics department at a tier one university and saying, hey, is anyone interested in doing this type of work and training one of the leading AI models? We’re hiring. Here’s the rate card, here’s what you might make. And we’re paying very good wages, depending on what the complexity is.

EL KALIOUBY: I actually wanna push back on you a little bit there. There’s a growing concern that the quality of these jobs — that there isn’t like job security, or there’s low pay or poor working conditions. How do you think about these concerns and how do you address them?

PEDRAZA: Yeah, I’m taking this call from France, and as you know, this is an economy where labor laws are incredibly rigid and it’s almost impossible to hire — not just contractors, but even employees — because it’s very much an anti-employer regulatory environment.

There is definitely a political component to answering this question, which is that we have focused on hiring contractors in states that are pro gig work, pro contractor, pro employer. And then there’s a sociological component to this, which is that our human society has evolved from predominantly agricultural jobs in the 19th century to factory jobs, then from factory jobs to office jobs.

And now we’re moving into this AI economy, which is increasingly entrepreneurial. Organizations are gonna get smaller, and we’re not gonna have the sort of jobs that, say, my grandfather had. He worked for Procter and Gamble, and then later he transitioned to Pepsi. This type of job with corporate benefits and a corporate pension plan and a 401k — I think that is going away. And I do think this is gonna result in society transforming, but it doesn’t mean we’re gonna live in a society where there’s no opportunity, no income, and no wealth creation.

It just means it’s gonna be a much more dynamic and flexible market in society. And our politics, our regulatory regime, our welfare state, our education environment, our markets — all of them will adapt to the shift. I think the desirable outcome is one in which companies like Invisible that are responsible employers are not overly burdened by the regulatory regime and can hire and transition as needed in order to create the capabilities we need.

And where the education system produces people with the right skills for the AI economy, and where entrepreneurs have a ready supply of people that are excited to join their startups. Because as organizations become smaller, there’s a ton of opportunity to create new companies and for people to make a lot of money in these jobs. And I just think that’s a market structure shift.

Copy LinkWhich jobs and company structures AI may create next

EL KALIOUBY: Can you name some of the top jobs that the AI economy is creating?

PEDRAZA: AI training jobs are one of them. But I think this is a superstar economy. It’s been said that as AI increases productivity per person, the best engineer can get 10 to a hundred times more done than they were able to get done before. And I think that’s true.

Across fields. And then people say, okay, if that’s the case, then are we gonna live in mass unemployment? I don’t think so. Here’s why. If you woke up tomorrow and your sales team was twice as productive, would you fire half of your salespeople?

EL KALIOUBY: Yeah, probably not. You would deploy them to scale more, do more, right.

PEDRAZA: Yeah, you’d probably just enjoy the fact that you have twice as much revenue. You’d probably hire more salespeople is what you would do, because your model’s working better.

And this is where it becomes a conversation that gets a little bit technical for people, but it’s important to get a little bit technical. There was an economist a hundred years ago who I consider to be a prophet. He was like the father of modern microeconomics. I really believe he sort of saw AI in this crystal ball coming a hundred years from now. His name is Ronald Coase.

And Coase understood that transaction costs, coordination costs, switching costs, discovery costs, integration costs — all of these frictions that exist between supply and demand — are what determine the shape of most companies and most markets, and the marginal economics inside of them.

So, here’s an easy one for people. How many cooks can you put inside of a kitchen where the specialization gains — because every additional cook allows you to specialize more and more — don’t get canceled by the coordination costs of all the cooks having to work together.

That is a classic Coasian intersection of coordination costs and specialization gains. And AI pushes these frontiers out. So maybe with AI in the kitchens of the future, you actually have a hundred cooks working together because it’s so easy for the cooks to coordinate. And that’s how AI is playing with our microeconomics.

So you might end up with, on the margins, huge organizations. But maybe you end up with future mega corporations that are basically networks or marketplaces, like Invisible could be one of them.

And then you end up with the rest of the economy being relatively much smaller because they’re tapping into these labor pools as needed.

Copy LinkWhy AI could make smaller and even autonomous companies possible

EL KALIOUBY: Yeah, because it’s kind of interesting, we also talk about the one person billion dollar company. And I wonder if that’s gonna be a possibility too, like one person with a whole team of AI agents. Do you subscribe to that vision?

PEDRAZA: I do. I actually think it’s possible that you end up with no-person companies, although it’s a real mind bender. But I think in practice it’s not about a one person company, it’s about smaller companies. We’ve already seen some incredible examples of this, like Instagram and WhatsApp, which were incredibly small companies when they were acquired for incredibly large amounts of money. And those are Coasian companies in the sense that their microeconomics were just insane. For such a small team to get so much done and create so much value is so incredible. And the future is already here. We’re already seeing companies like this emerge.

EL KALIOUBY: We’re going to take a short break. When we come back, why Francis decided to buy out some of his investors. Stay with us.

[AD BREAK]

Copy LinkHow the sovereignty game changes startup funding

EL KALIOUBY: So you talked about the different ways you’re disrupting the status quo, but I think there’s an additional way you’re doing so, and it’s in how you fund the company. So you’re venture backed, you’ve also raised outside capital, but you have this model where every so often you give current investors the opportunity to cash out, which allows you to stay private and privately owned for a lot longer. Say more about that approach.

PEDRAZA: Sure. We realized that the typical way of backing a SaaS company was to raise a series A, B, C, and IPO and M&A in five to seven years. And we—

EL KALIOUBY: Is the path I’ve been on. Right.

PEDRAZA: Yeah, most entrepreneurs have done that. And the venture model, the venture game, has created some incredible success stories. But it has a number of disadvantages over time. The majority of the company ends up being owned by investors, and the majority of the board ends up being controlled by investors.

So the entrepreneur eventually ends up losing a lot of both managerial and board control and influence. We started playing a game and then we realized we needed a name for it and we decided to call it the Sovereignty Game. And it was an alternative to the venture game. In the Sovereignty Game, you raise as little capital as possible instead of as much capital as possible, and you focus on long-term shareholder value instead of short to medium term shareholder value.

And you try to minimize dilution along the way. And you focus on raising money from your team, so to speak, by giving your team lower cash comp than they would otherwise get at any given stage, but more equity. And so this allows you to create a culture of what we call an ownership mindset or an entrepreneurial mindset where everyone’s really motivated and thinks like a shareholder, instead of just thinking like an employee or operating like an employee.

We raised only $7 million before we were profitable and we became profitable at an 11 mil run rate with a million dollars of annual profit in 2021. But the point is we were able to scale profitably without further dilution. We took 16 million last year to strengthen the balance sheet. And that was the first primary capital we’d raised since 2021. And then we used debt to create liquidity for our investors and to buy them out.

So when we started doing this, investors owned almost 50% of the company. And then we were able to buy back about 20%. So we got to a cap table that was 70% owned by the team, 30% owned by investors when we did a secondary round last year, and now it’s more like 65/35. And so we ended up switching strategy.

And so now we’re in a position where we actually need to raise capital to strengthen our moat and invest a huge amount in R&D.

Copy LinkWhat AI reveals about creativity and being human

EL KALIOUBY: Yeah. Amazing. Alright, Francis, last question, and it’s one I like to ask of all our guests on the show. You’re spiritual too. And with everything we’re seeing with AI — AI becoming smarter and more creative — and I’ve spent a lot of years building AI that’s empathetic. What does it mean to be human in the age of AI?

PEDRAZA: Martin Heidegger, who is one of the most important philosophers of the last century, wrote an essay called The Question Concerning Technology. And he came to the conclusion that the nature of technology and human nature are in a symbiotic relationship where they reveal each other over time.

So you can think of it like a wizard and a wand. As the wizard makes the wand more powerful, the wand helps the wizard realize who the wizard is. And so we are definitely not the capabilities that we build into our technologies. For example, if you are stacking books on a shelf and that’s your job all day, every day, and then you build a machine to stack books on a shelf, you realize, okay, human nature was definitely not book stacking. Human nature is beyond that. So I think that our creativity, our creative potential, ends up really coming into focus as one of our superpowers. And that creative potential is enhanced with AI.

That act of imagination and manifestation is very uniquely human. And the tools just really allow us to manifest, allow us to do those things. And so I think that we become more powerful, and the technology can only help us get from point A to point B.

EL KALIOUBY: Yeah. I love it. I love this visual of, as you enhance the wand, the wand reveals more of the wizard. I’m gonna take that away with me. Thank you, Francis. This was—

PEDRAZA: Thank you, Rana. I enjoyed the conversation as well. I appreciate you having me on the podcast.

EL KALIOUBY: I have two takeaways from my conversation with Francis. One, to drive the most value out of AI today you need this partnership between humans and AI. AI should augment and amplify our human abilities, and as someone who invests in human-centered AI companies, I think that framework is really key.

And my second takeaway is about business strategies. For years, SaaS companies dominated the tech landscape. And now with AI, they are ripe for disruption. Invisible Technologies is just one of the many companies that are seizing this opportunity. Is the era of SaaS over? Maybe not, but there is so much potential for AI to create new business models and become the central driver of our future.