McKinsey's AI leader moved to head a $2B AI workforce - Matthew Fitzpatrick [Invisible]

0:00Data does tend to be the biggest impediment, particularly in large enterprises, to building anything with AI. Next iteration and generation of this will be anchored in kind of having a coherent set of platforms that allow you to get AI working in production. You know, 25 years ago, you had like literally a local loan officer that made a decision whether you got credit or not. And now 85 to 90 % of that is auto adjudicated by underwriting mo. Today on Billions, I'm sitting down with Matthew Fitzpatrick, the CEO who saw an opportunity others missed. leaving one of the best AI role in the world to run the companies that train the AI everyone else builds on.

0:36At McKinsey, he led 1 ,000 engineers for the firm's entire AI R &D arm called Quantum Black Labs. He was the person Fortune 500 CEO called when they didn't know what to do with AI. In January 2025, he took over in Visible Technologies as CEO and 8 months later raised $100 million at a$2 billion valuation. Matthew, thanks a lot for being here. Thank you for having me. Can you maybe walk us through why you decided to move from a very comfortable position at McKinsey to where you are right now? Yeah, look, I think first and foremost, I think I was just very excited about Invisible. I think the company had some really unique expertise and experience having trained all the large language models for the Frontier Labs and the large language model builders.

1:27I thought that was a, you know, the paradigm in the market today, back then, at least, of how do you ensure that kind of an AI model is working in production in an enterprise context? There was not really a way, and there's still, to some degree, not a perfectly well understood paradigm for how to do that. And I thought Invisible had really unique experience in how to do that. I thought the kind of caliber of the team that had been built around that as well was exceptional. I think those are the two things that really attracted me to it. And then I also think there was a unique point in the market where, you know, I'd spent a decade, 12 years at McKinsey at that point.

2:03And I did feel like I wanted to enter a world where I could raise capital to pursue the ideas I wanted to pursue. And I also think, you know, Invisible is a platform software company where we've, you know, the work I did at McKinsey and at Quantum Black, focusing on forward applied engineering was something I loved doing, but I also have a believer that the next iteration and generation of this will be anchored in kind of having a coherent set of platforms that allow you to get at working in production. And that was really, was I excited to build that. If you had to pitch like for people who don't know invisible, like if you had to pitch it to a two year old, what would that sound like?

2:41Yeah. So we're a modular AI software platform. We do two things. We train all the large language models for via reinforced learning. So think of that as effectively any type of the world, anything, anything you, any model you might build, we'll actually bring human expertise to then evaluate if that output is good. And then on the enterprise side, we serve 10 different sectors, but I said the most active ones being food and beverage, healthcare, asset management, and consumer retail. And we use the same suite of platforms across both. So we have Neuron, which is a data platform that brings together structurally bunch of data.

3:22Atomic, which is a process builder that can build out any operational process. Atson, which is an agentic platform that can orchestrate agents. And then Meridial, which is our expert marketplace. And we have 1.1 million experts at any one time. That's what we use to do a lot of the reinforced learning training. So we use that suite of four platforms. in the enterprise to build effectively hyper-personalized software using that common suite of platforms. And so everything from like inventory management to pricing, we're using those same suite of platforms to deliver. And I was going to say, and we wrap it all with forward deployed engineering as kind of the core delivery motion.

3:54So think of that as we use a lot of the really impressive kind of innovative GNI coding tools to actually deliver a lot of this much faster in that paradigm. And can you share like a specific use case of something that has been developed inside one of these industries? Yeah, sure. So I think one example I can cite recently is Swiss gear. So Swiss Army and luggage is an example. You know, one of the challenges that they had was inventory forecasting on a six-month lead time is quite challenging. And actually how knowing how you should forecast inventory. And obviously that has huge implications both on the amount of inventory you have to carry if you don't sell it, or if you don't carry enough inventory, you miss out on revenue as a result.

4:37And so, you know, we mapped together about 100, 750 different data tables very, very rapidly to do that kind of force casting transformation. And the tool we built actually doubled the number of SKUs that they could forecast effectively, just as one example. Wow, that's nice. And do you like, what's the business model like? Like, do you build all this kind of like tools and models inside the company, and then you have like a fee to maintain it or is it like a one-off? Everything is a managed platform so yeah we can either host or or deploy as a container in the company's environment and there's no big upfront services fee what we're doing is deploying something on an ongoing basis that works and I think that's actually a really important part of this the model of like charge an enormous amount up front and then hand it over has really not worked in the general world partly because you just have like, I'll give an example.

5:30Let's say you're a engineering construction player and you build a model that should handle permitting, but then permitting regulations all change in three months. Like you actually need a way to an ongoing basis, make sure the model is right and maintained. And so, yeah, our, you know, if I take the Swiss gear example, we're using Neuron, our data platform to bring the data together. Then we're using Axon, our agenda platform to actually build out a lot of logic. And then that is maintained an ongoing basis. And so it's a managed platform structure. Nice. And the business model, would it be close to like SaaS?

6:03So it's like a monthly fee or yearly fee? SaaS is a controversial topic at the moment. What I would say is we have more different pricing models than we do have some SaaS pricing. So generally, I would think of our pricing on the enterprise side is a mix of three components um a platform fee some often kind of consumption based fee so that can be things like price per call and a contact center for example so it's tied to the variabilized usage and then the third is actually tied to outcomes so we do tie quite a bit of our fee structures to shared outcomes with our customers okay i think it's important i think that you know worlds where you just have a sas price per seat it's actually pretty hard to to make the ROI clear to folks.

6:48So it's actually been, I think, a core part of what we do. Nice. And, you know, we just touched base a little bit about SaaS and valuation and the kind of, you know, like a lot of people, and I think it's probably half the industry, think, you know, we're in an AI bubble with the valuation, with capital concentration. I'm curious to hear like your take on that. Yeah, I think the interesting thing about the F &I funding market is it's something like three companies are 30 % of the total funding, and then it's about 100 companies are about 70 % of the total funding. So you could argue, you can debate whether it's a bubble or not.

7:30I think there's all sorts of different questions about where else one would put money to generate better returns. But I think the main thing I would highlight is it is a focus on a small number of companies where there is scarcity. I mean, there's a lot more investors that want to put money into those hundred companies than is available space. And so I think that's the interesting dynamic of the current AI era is there's not a thousand companies doing this. It's a small number. It's where a lot of the market wants to invest. And so you have a scarcity dynamic and that's what's causing, you know, whatever valuations we have now.

8:02The other thing I highlight, though, is it's not a question of this era of companies is not like, you know, people bring up the dotcom era. It's not a bunch of eyeball companies. I mean, these companies have hundreds of millions, if not billions of dollars of revenue at a much faster pace than we've really ever seen scaling loss demonstrate before. I think it was like the number I saw recently, for example, is like this class of the most recent class of Y Combinator from a year ago has the highest aggregate revenue ever for its cohort a year in. And so you actually have like several different factors at work.

8:36You have scarcity and not that many companies that people can invest in. really unique scaling laws from a revenue standpoint in the whole space. And then I do think the macro makes this whole thing complicated as well, which is there are not that many areas right now where the returns are good. Areas like real estate that have traditionally been really reliable have been more complicated. And so you have a lot of capital trying to move into this space. And so you can debate what the right fair valuation is, but I think that's the fairly logical set of reasons why we are who we are. And from your perspective, because you mentioned the dot-com bubble, Like, what do you think is going to be the impact of AI versus like what we had with the Internet?

9:20So I saw an interesting article that Bloomberg published on Mary Beaker's technology predictions 20 years later. And so they basically looked at like what she said in 1999 and who she predicted would grow and at what rate. And the interesting thing is, and she was used like this insane optimist by any measure, she actually was underestimating how transformational the internet was. Meaning like the amount of e-commerce that developed as a percentage of total retail, the number of users that moved online, even by the wildest, most optimistic prediction you could take at that period, it was more adoption than that.

10:00now the flip side of that is i know this is kind of the most interesting part of that set of projections was she thought the players that would capture a lot of that growth were like the blockbusters of the world and so where where i think she was off was that it was not the established players that figured out how to monitor there are some like walmart etc that did figure out how to evolve but actually it was a lot of new entrants that kind of sprang up out of uh from scratch and figured out how to take advantage of the new market i think ai will be a very similar and by the way again that's all the context that it took the internet you know north of 10 years to really probably north of 20 years to get to real maturity i think it would be a very similar paradigm and almost every new technology has had that sort of a curve it will take longer so you know i think everyone's been of the view that everything's going to change in two years look i think enterprise adoption right now you know depending on the numbers you look at something like 5 % of JDI projects make it to production.

10:59You know, there's an interesting stat I heard the other day, like 98 % of inference compute is like 37 companies. Most of them are coding apps. So you have like, we're pretty early in the cycle on enterprise. Hugely advanced to where we are in consumer adoption. So like 70 % of consumers use this tech at least weekly. So we're there on consumer, but we're early in enterprise. And I think enterprise is going to take north of a decade for real adoption. And I think the most interesting question, if you take the Internet as an exact parallel, is if that all happens, is it that old quote of like, do the new players get distribution before the existing players build the tech?

11:36I think that will be the big question of the next decade is, do all the new businesses developed around AI become the new players in the spaces or do the existing players figure out how to transform? So I've got two questions for you. The first one, why do you think there's such a gap between enterprise adoption versus like consumer adoption? And I'll tell the second one afterwards. Yeah, I think it's partly the, as a consumer, when you use one of the models, you want it to be broadly right and broadly healthy. And it actually is always broadly right and always broadly healthy. So if you ask it a question about like where to make a restaurant reservation or, you know, even a new prototype you want to build for a business you're thinking about, they're unbelievably healthy.

12:22I think the more complicated dynamic for enterprise is most, call it particularly Fortune 1000 use cases, rely on extreme precision. So if you take anything from contact centers to medical coding to credit underwriting, you need 100 % accuracy. You need auditability. You need really clear data that underlies it. And so I think that adoption paradigm is just different. Like these are probabilistic models that are not always perfectly correct. They have hallucinations. And so the actual process of taking a model, designing a system around it, quality controlling it, testing it, that's just a newer motion that I think enterprise is going to take some time to come up the curve on.

13:08Like the phrase often uses, it's not a technology gap, it's a capability gap. And so I think this is a motion that folks are, I do think you've seen an acceleration in the market in the last year or so where like people are starting to adopt for the first time in material ways but i think that journey is going to take real time i mean the analogy i'll give because i think it's probably the best use case of ai kind of in american society would be if you took um underwriting of mortgages like single family home mortgages right um you know 25 years ago when you went to go do that you had like literally a local loan officer that made a decision whether you got credit or not right and so that was like a decision and one person ransom numbers came back.

13:50And now 85 to 90 % of that is auto adjudicated by underwriting models. And so that's been a hugely positive thing for society in that you have much fairer, less biased access to credit. Your speed of getting access to credit is much faster. And that's really powered a lot of the housing market. But that did not happen overnight. That took 15 years, maybe 20 years for people to actually understand how to redesign the processes, how to resign the intake forms, how to quality control it, how to show regulators, like via model risk management, why it was fair, why it was not redlining, why it was unbiased.

14:29And so I think a very similar pattern will exist for GMI. And how long do you think, because you mentioned, you know, like it took about 20 years, you know, for the internet to reach, let's say like the maturity stage. we see that things are moving a lot faster with the ai do you feel like the adoption is limited simply by human nature because you know like changes usually like take time or do you feel like it's a capability you know like the fact that the way the ai is evolving like for international it was really like capability oriented we can think of speed we can be the coverage we can about this in terms of speed?

15:12Well, one thing I'll say is I think it will be uneven in its proliferation. So I think you will see certain areas of society that will proliferate much faster. I think contact centers, for example, is one that it's definitely not there yet. And there's been a fair amount of bumpiness, but I could see that making huge inroads the next couple of years. I think media is another one where everything, video creation, I think you're going going to see really rapid adoption of that as an example. I think there are other areas like financial services that will definitely take longer, particularly any regulated industry is going to take longer.

15:50And I think by the way, that's different than the internet because the internet is a broad-based technology. It's a network. In many of these cases, you could be super advanced and I could not be advanced at all, and that we could both be using the same technology. So the adoption paradigm here is more individual to individual institutions, individual companies, individual sectors. And so that means I think in some cases it will advance faster than the internet business precedents. In some cases it will advance slower. But I do think it's going to take longer than people expect, even though you see hyper adoption at the individual level.

16:28There's a study I really like by the NBR, National Bureau of Economics, and they basically look at individual productivity use cases versus company level. And it's like the average, they look at like 60 ,000 employees or something like that. And the average individual improves in productivity like 15 to 20%. But the average company improves in productivity, I think it's like one to 3%. And I think that is really the paradigm right now is people using all these tools and they're helpful, but they don't really redesign the way the work is done. And I do think that takes time. Like there's an analogy that like the automobile took 40 years for mass adoption and it actually took the redesigning of the factory to get that.

17:10Like you actually had to redesign factories around the assembly line to bring costs down the automobile to use them. And I think there's a similar analogy here, which is we're not really changing the nature of the work in the factory. We're just having everyone use the tools. And I think that will take time to play through. So we're five years into the journey. Now everyone is using a lot of the technology but i think the nature of work will take another you know we're still i think five to ten years away from material like the amount of use cases being shaped i will say though this is where um digitally native companies are different than enterprise and i think that's um if you are a company founded in the last 10 years you are sorry last two years anchored around all the cool, all the digitally native kind of tooling, coding apps, et cetera.

17:58I think the acceleration of those businesses has been vertical. I think it's very different if you were a 10 year or 20 or 30 year old company re-engineering yourself around that tech. And I think that is a much harder thing to do. And that's most of the scaled companies in the world. So do you feel like to build up on what you were saying with, you know, the initial prediction of having the existing blockbuster of the world like take all the market share. We can compare this to some of the existing giants today. Do you feel like the fact that they have distribution, if you take Google, for example, even though Google is working on AI and they're shipping a lot of things, I think they've been a bit more quiet compared to what OpenAI and Enthropic, etc.

18:46have done. do you think they still have a chance like to to get like tons of market share or do you feel like it's a lot harder per your points because they haven't been built as ai company first oh i think they absolutely were built as an ai company i think they're the original and perhaps you know greatest ai company in my way so i what i'll um without coming on the tech landscape because i think that's a um i think that's a different dynamic that's a bunch of companies all investing heavily in tech. I think what I meant by that comment is more to take like, let's take banking, for example. Like if you take all the neobanks, like a Revolut, I think the question is the adoption rate of AI in a business like that versus traditional banks, how does that change the paradigm?

19:31And I think you can extrapolate that to almost every sector, which is for companies that become hyper adopters of AI, their product theory should increase materially. And so how many new businesses like that do you see that spring up by sector? And contact center is another good example that we've already seen some new entries grow incredibly quickly. Okay. And you were mentioning, I think that the SaaS playbook is kind of like dead. So how exactly do you think it's going to evolve from where it is right now? Yeah, let me clarify that. So I think the repricing in SaaS companies is not what the market, I don't think at least, is not assuming that they will all decline in revenue.

20:15They're not assuming something even gets ripped out. I mean, I think the rough multiples on SaaS was in 2021, the average public SaaS model was like 20X revenue. Recently, it's been on the order of 10X revenue, right? And maybe in many cases, 15X revenue. And so the repricing I think you've seen is a view that they should be trading much lower than that as a multiple because in the future, their growth will be slower and the terminal value will be less. And so the repricing is not a view that they're going to contract. It's just a view that they will grow less slowly because of a competition from AI companies.

20:50And there was a result of that pricing competition as well. And, um, and I think that is a fair assumption in many ways. Like, I think that the way that people buy software and technology is going to change. Like, I think in the last 10 years, you kind of simplistically, the way you bought software was you bought a box, whatever that box is. So like you bought something that did like an insure tech that did claims, for example, and then you have an SI come in and customize that thing incredibly for exactly what you needed. And so what you ended up with after, you know, two years of customizations was like a kind of an ill fitting thing that sort of fit what you needed, but was probably very different than the core box that you started with.

21:32And that's been why I think enterprise software has been so frustrating for so many is that a lot of the tech was really not fit for purpose. And now with AI, I think the question a lot of folks are asking is, rather than doing it that way, shall I take a variety of different tools that now exist in me and build something or use something that is pretty customized to exactly what I want? And I do think that will be a change. I think that'll be a direction a lot of folks will go. And that's kind of effectively what Invisible does. We use a series of foundational infrastructure to deliver hyper-personalized software.

22:04And I do think that's a much better user experience than buying a box that is kind of ineffectively customized. Makes a lot of sense. And do you feel like, because you touch a little bit on this about the fact that usage was going to change. And I feel right now when I look at most AI software, I mean, a lot of people are talking about AI agent and things like that. But actually, it's more or less the same capabilities, you know, as a normal software or a workflow that you can do. In some cases, they add intelligence. But in most cases, it's like what I would call like just custom software or like just normal software.

22:45so in in what way do you think interfaces are gonna change to kind of like match the new needs of consumers or uh enterprise yeah it's it's an interesting question so i think what you're asking in some ways is everything is called an agent at the moment but a lot of it looks still like traditional software i think that is accurate by the way i think there's um i've seen a bunch of interesting data on what percentage of quote-unquote agents are actually agents and you see that a huge chunk like 30 are just basic like scripts for example so so so um i think that's been one of the other reasons that adoption has been slower than uh estimated is actually the prices of building like a task-based agent that is actually autonomous is not something that a lot of folks like what i would call like a multi-agent system where you have an agent that orchestrates numerous single task-based agents a lot of people talk about that um very few companies I've seen have been able to implement that effectively.

23:43And so I think that is a hard thing to do. And I think that's been one part of it. But I think when this is all, wait, let's take five years from now for further along the shorting curve, you will realize that the answer for everything, there are numerous tools that you might consider. Deterministic software, machine learning, agents, a probabilistic LM, any of those. And actually a lot of solutions are going to be a mix of those four things, not just one thing. And so you've got to be very directive about like which of those best solves a particular problem you're looking at. There are many cases where an agent might be worse than a software product.

24:23And so in that case, you might just want to build software faster that solves a need. You want to be purely deterministic. There are many other cases, like I think contact centers is probably the cleanest one where an agent is absolutely where you want to go because actually deterministic flow of like an IVR, which is how you do routing in a contact center. It's just very bad. And so you're much better off with an agent that can make judgments and respond. But there are other cases where if you want like much more like deterministic routing than traditional software is better. Okay. Yeah. I like there is the reason SAS has come down.

24:56I think software is going to be more and more popular in the next 10 years, but it won't be out of the box. Like kind of one, one function software like much of the SAS world has been. And it will allow for a lot more customization. That is the benefit of many new coding tools. And do you believe customization can be done by AI down the line? Meaning that you would have software that kind of fork themselves into a unique version for each person using it? Or do you feel like there must be a human working and a human in the loop to actually create what fits more of a group of people rather than an individual?

25:42I do think there will be variants of that that will be AI enabled for sure. I think that the reality is if I'm using a piece of software every day and it's able to see my usage patterns and my flows, then it should adapt to what I'm doing. I think that the challenge is, and people have asked, can that happen today? You have a bit of a threshold that has to be cross first, which is you need to get people using the technology first and you need to get adoption. And so what I'd say to you, if you take an example, let's say there are a hundred people at a company and you want to develop a piece of software for them.

26:15Even if AI could observe a lot of it, the process of bringing those people along and getting their feedback and the kind of process of building something in the way they want it is incredibly important for adoption. And so I think there still has to be a very material human element to get people excited about using much of this new tech. But I do think down the line, more of it, you will be able to have AI kind of evolve the software dynamically. And right now in the way you work, because I think at some point you mentioned that if you let your IT department, you know, like have the control over AI in building, you know, like this prototype, basically everything gets stuck in pilots.

27:01So how, you know, like, would you go about building like an AI project within a company? Like who should be involved? Like how should that be structured? And how do you like build everything you've done, you know, so far? Yeah. So the MIT study I mentioned earlier that talks about, you know, 5 % of AI projects making into production, they cite in there that they look at the companies that bring AI all the way to production most consistently. And one of the main characteristics they cite is line ownership of AI actions. And alongside that, very clearly divine KPIs of performance tied to that. And so what I say about that is it makes a lot of sense, right?