Invisible Technologies CEO On Building AI Around Real Workflows, Not Hype

[00:00:00] A quick thank you to NordLayer for supporting the podcast and helping me make these daily conversations possible. And if you are listening and you're responsible for security or IT, you will know the reality. The reality that most of your risk now sits inside SaaS apps and browser activity. That gap is exactly what NordLayer is addressing with its new business browser.

[00:00:24] So instead of bolting security on from the outside, it builds it directly into the browser itself. This means you can control access, monitor activity, enforce policies and reduce shadow IT all from one single place. And most importantly, it does it without adding deployment headaches or complex onboarding. You get things like browser-based data loss prevention, SaaS access control and zero trust browsing.

[00:00:54] But delivered in a way that your team can actually use. So if you've been trying to simplify your stack while improving visibility, please check it out at nordlayer.com slash browser. But now it's time for me to introduce you to today's guest. AI is everywhere right now. It's in our headlines, our news feeds, our boardrooms, our smartphones and apparently even in our refrigerators.

[00:01:23] But despite all that noise, there is still one question that keeps coming up in almost every conversation that I have with business leaders and technologists. And that is, why do so many AI projects still struggle to move beyond the pilot stage? Well, my guest today believes the answer has very little to do with the models themselves and a lot more to do with everything around them. His name's Matt Fitzpatrick.

[00:01:50] He's the CEO of Invisible Technologies, a company that's focused on helping organizations actually make AI work in the real world. And he moves at a pretty fast rate. And since stepping into the role, he's already raised $100 million in funding and expanded the company's global footprint.

[00:02:12] And before all that, he spent years leading quantum black labs at McKinsey, where he's worked on large scale AI deployments long before most of us were talking about generative AI every day on LinkedIn. But in today's conversation, Matt will share why the last mile of enterprise AI might be the hardest part, but why messy data quietly derails more AI initiatives than many people realize.

[00:02:40] And why the future of AI will almost certainly include humans very firmly in the loop. But I'd love to wash all this down with some very real world examples today, whether it be turning ordinary broadcast basketball footage into advanced performance analytics, or cleaning up hundreds of fragmented data tables for a luggage brand in just a week,

[00:03:05] to helping physicians build a more complete picture of patient healthcare across wearables, records and clinical systems. Yeah, we've got so much to talk about today. And I think those real world examples really do bring this to life. And somewhere along the way, we'll also tackle some of the bigger questions too. What should boards really be worrying about when it comes to AI governance? Why are enterprises still moving cautiously while consumers race ahead?

[00:03:34] And will AI replace jobs or simply reshape the tasks that fill our working days? But enough from me. Let me introduce you to Matt so he can talk about what it actually takes to make AI deliver on those big promises. So thank you for joining me on the podcast today. Can you tell everyone listening a little about who you are and what you do, Matt? My name is Matt Fitzpatrick and I'm the CEO of Invisible Technologies.

[00:04:01] We are in a modular AI software platform where we have a series of tools. Neuron, which is a data platform for structured and unstructured data. Atomic, which is a process builder. Axon, which is an AI agent builder. And then an AI expert marketplace called Meridial where we source all sorts of human expertise around the world. And we use those modular suite of platforms to do two things. Build applications for enterprise and then train and fine-tune models for all the large language model builders.

[00:04:29] And Matt, you are also the epitome of hitting the ground running. Before you join me on the podcast today, I was doing a little quick research and you stepped into the CEO role at Invisible Technologies in January. You quickly raised $100 million, which is no small feat, especially in the current climate. So what was it that convinced investors that your approach to making AI work is different from the wave of AI platforms already in the market out there? Yeah, it's been a busy year.

[00:04:57] I think the interesting thing, if you take the market landscape for AI, is you have kind of three different buckets, I'll say. Like the large language models, which have obviously done amazingly well. And then you have all of the application layer companies, which are basically building one-off single-use applications anchored around agents, right? And there's contact center agents, legal services agents, whatever this might be. We're a very different bucket. We are kind of the last mile of enterprise adoption.

[00:05:22] So we build hyper-personalized applications that use the specific knowledge of an enterprise or institution. And we build tools around that uniquely specific knowledge. And I think the interesting thing is very few folks are taking that approach. I mean, I think our approach is anchored in using four deployed engineers and a lot of customization for the individual customer.

[00:05:44] Most of the AI players that are not LLMs, most of the application layer companies, are trying to build reusable, consistent, single-use tooling effectively. And my personal belief is that that is an interesting approach. But I think the last mile is really where AI is going to need to go to reach adoption in the enterprise. And so I think it's almost a different approach.

[00:06:06] I think the number of other players doing kind of that four deployed delivery anchored, four deployed engineering anchored, highly specific workflow design. It's not many. It's probably less than five scalable players in the market doing that. And I think we have a really unique expertise from it. It's very similar. I spent the last decade before this at McKinsey where I was a senior partner and led Quantum Black Labs, which is the firm's global tech development group.

[00:06:29] And a lot of what I'm doing here is really anchored in similar approaches we did there, which was figuring out very specific problems that folks are trying to tackle and building very specific applications. The difference here is that this is pure software. So everything anchors around an application, an end-to-end tool set, and consistency of delivery in how we do it. And as you mentioned, this is not your first rodeo. You did previously lead Quantum Black Labs at McKinsey.

[00:06:56] So I'm curious, was there anything that you learned in those experiences there, shaping AI strategy, that would go on to influence how you build and scale a product-led AI company today? I suspect there was a few synergies there. I think the biggest thing is that I don't come out of the traditional Silicon Valley school of build a one-off scalable use case that you do repeatedly. I do believe that actually the customization in the last mile is the big driver here. And that's also because I spent a decade doing large-scale AI deployments.

[00:07:26] Back then, it was called machine learning. But you realize, like, machine learning has to be customized to the individual customer. It has to have the unique context embedded in it to make the models work. It's not like you can build something with third-party data and do it repeatedly. You need clear internal data that is organized and well-structured. You need hyper-tailored workflows to the individual institution. And you need clear line and business ownership. And so I think a lot of that experience has been a real driving factor for me.

[00:07:54] I think the other part of this, though, is there's a lot of similarities in the process of organizing a large team around a goal. I do think that over a decade of doing that at McKinsey, I did learn a lot about how to allow people creative outlets, the opportunity to move between different projects, what I would call rotational velocity, and ultimately the idea that ultimately recruiting the best people is the most important thing by far. I think that's the other thing I've really carried over from this is that for every hour I spend recruiting,

[00:08:22] it's a much better use of my time than almost anything else because ultimately the best engineers are the determinant of how much progress you make. I think Gartner found that 60% of AI projects fail because companies don't have AI-ready data. And it is such a powerful point. I hear a lot at tech conferences, no data, no AI. I do think that's been the other big directive for us is that I think one of the most interesting things about today is you've had a couple things happen at once. Models have improved by 60% to 100% over the last two years, so the models are a lot better.

[00:08:52] Consumer adoption has been exponential, so like 65% of consumers use various AI models once a week. But the enterprise adoption has not been there yet. Something like 5% of enterprise projects make it to production. And that big MIT study got into that. And so I think that's the really interesting thing over the next decade is how that evolves. Looking at a lot of your announcements, Invisible often also talks about human in the loop as the main deployment model. It's a big topic right now.

[00:09:20] I hear it a lot at tech conferences as well. But in practical terms, how do you design systems that expand what people can accomplish rather than quietly replacing them? Because, again, a huge topic right now, isn't it? Yeah, look, I mean, we're a very human-centric company. We actually believe that despite some of the kind of current media alarmism, actually humans are going to be the center of AI deployment for the next numerous decades to come and in perpetuity.

[00:09:47] And I think the implication of that, if I go back to the description I described earlier of applications versus last mile, if you're building something to require no human involvement, so like fully AI-centric, you probably need to design a closed system one way that is infinitely scalable in the way it does it. But if you actually take the totally opposite approach, which is what we do, which is to say, how do you do a process today? Where are there points of that that require human judgment, thought, expertise?

[00:10:15] And then you build a process around that. You can stand something up much faster that anchors in the expertise of the folks involved in it, but actually just makes their lives a lot more efficient. So like I'll give an example. Let's say you take a commercial credit agreement. This is one and one of many. But let's say a commercial credit agreement is 100 pages long or 50 pages long.

[00:10:36] And if I wanted to build a tool that could scale commercial credit agreements forever for everyone, it would probably require 50 million, 100 million of investment to build a perfect agent to do that. But the problem is you actually run into a lot of companies want to use those credit agreements differently. So we take a different approach, which is we'd say, OK, you, institution one or bank one, how does your process look today? Let me think about the different components in a commercial credit agreement.

[00:10:59] Let me think about the parts of this you actually can use agents and data foundations for very quickly, like core ratios, core reporting of company information. And so you can actually start to get a lot of stuff that market summary context. You can get to about 70 percent of the agreement is automated. And actually, it's the highly repetitive, not that thought provoking, not that complex parts of the agreement.

[00:11:24] And then what you leave is the rest of it, which is actually much harder to predict and create consistency for. You create a kind of a context for human input as well. And so the way I would think about it is there's a really good proxy for this, which is mortgage loans. If you look over the last 20 years, mortgage loans are about 20 years into the journey. About 80 percent of mortgage loans are determined by auto adjudication. They're determined by a machine learning model.

[00:11:52] But about 20 percent still get escalated to humans. And they're the most complex 20 percent that require a lot of additional context or review. And it's the same thing with the commercial credit example I'm sharing here, which is you're going to be able to take a lot of the highly repetitive tasks and automate that. And ultimately, that's probably a good thing to make the process smoother and easier to operate. But you're going to need human judgment for the stuff where you have not seen precedent formation.

[00:12:18] And before joining me today, I was also reading about the Charlotte Hornets project that turns broadcast footage into high frequency XY tracking data, even for leagues without MBA level style infrastructure. So tell me more about that. And also, what did the engagement teach you about building AI systems in data scarce environments? Love to find out more about this. Well, yeah, I mean, you know, in that case, we were taking broadcast feeds.

[00:12:45] And, you know, the typical way that people build what's called a computer vision video model is you would create a really complex set of camera installs. You would train for years and years and years to go after a really specific use case. And after some number of years, you get to a high enough degree of accuracy. And there's a bunch of players in the market do that now. We have not previously been a sports company. But what we figured out was that a lot of the work we do in AI training could be really quickly translated to computer vision.

[00:13:11] And so we were able to fine tune and train a model specifically for the Hornets that basically looked at a series of video inputs and was able to extract very specific data sets they wanted out of that video data by training that model very specifically for them.

[00:13:27] And so think of it as, again, the difference between with large language models, with a lot of the new computer vision tooling, rather than training a model specifically for one thing over five years, you can take a very specific use case and get it up and running in a couple of weeks, which is a totally different paradigm for technology. And in another case for Neuron, Swiss Gear, Neuron cleaned and structured 750 scattered tables in a week.

[00:13:53] And for people listening, just to put that figure into perspective, it's something a 20-person engineering team might have taken over a year to tackle there. So what is the real bottleneck in enterprise AI today? Is it model performance or data readiness or is it something else? I actually think data is the big bottleneck for AI in most cases. I think if you've got 50 fragmented systems that don't talk to each other, it's very hard. You can't build good AI off of bad data.

[00:14:20] So if you don't have a consistent customer record or product record or inventory record, the AI is effectively useless. And so I think what was interesting about this particular use case was the speed at which we were able to build all the data structures. The fact that we're able to do it with a very lean team and minimal input and get to a place in a couple of weeks where we were actually able to start building really sophisticated models on top of it.

[00:14:41] And ultimately, we were able to double the number of skews they could forecast accurately and basically expand category coverage of forecasting by about 30% for them, which has a really material economic impact. And so I think – but I would say that's the secondary part of it. The first part of this is getting the messy data cleaned up. And I think that's a lot of what our neuron platform does is we do think that when good AI meets bad data, the data wins. And so we start with making sure the data is clean.

[00:15:08] Another example with Lifespan MD, you're building a 360-degree patient view across fragmented EHRs, wearables and practice systems and so many more examples there. But I'm curious, how do you balance that clinical insight with data privacy and governance in such a sensitive domain? Because there are so many opportunities in this field, but they've always been a little cautious around data privacy, governance, et cetera.

[00:15:35] Yeah, look, I mean, the good news is having done this – having worked in this space for a long time, there are pretty clearly well-accepted paradigms of how to handle that. And so like in the healthcare context, you need to set up what's called a HIPAA-compliant multi-tenant cloud where the individual data for each practice does not leave – the patient data for that practice does not leave that practice. And so you create a really secure data environment where you can share very limited insights that are more population-based.

[00:15:59] This is a pretty well understood – there are several public institutions that have done this, but you start to collect patient information like longevity tests for men age 36 to 50 that you can use at KPI or access across similar wearables to understand longevity impact. So think of it as de-identified data that enables next-generation clinical insights, but keeping the actual patient data highly secure and anonymized in its local context.

[00:16:26] And I think that's what's really powerful is when you can start to get population-level health data on which weight loss protocols work, which peptides show impact, and how the top physicians are improving long-term outcomes. That's amazing for patients. But you want to make sure the patient data itself is kept highly secure and anonymized as part of that. And one of the reasons I wanted to bring that topic up is because I think Gartner and Deloitte have both pointed to governance gaps in rising AI incidents as adoption continues to accelerate.

[00:16:57] So presumably we're going to see much more examples likely. So how should boards be thinking about agentic AI risk when only a minority of organizations have those governance frameworks that are mature enough to handle what we're talking about here and launching hundreds if not thousands of agents out there into the wild?

[00:17:16] Yeah, look, I think one of the misconceptions is that you can regulate – that you should regulate the inputs, meaning like model building, as opposed to the outputs like an objective. And I'll give an example. Take model risk management. I mentioned mortgage underwriting earlier. The government put in place a bunch of really logical standards around things like redlining to say you have to ensure there's no bias. What is the mortgage underwriting process supposed to look like and what's the output you're going for?

[00:17:43] In that case, I think it's actually been a very positive impact on U.S. mortgage availability for the whole population. So I think in terms of regulation, it is most effective when you're focusing on an output. I think on the input side, it's very hard to regulate something that evolves every two months, and that is the challenge. And so I think where we've seen there is a paradigm for where this has worked in other sectors, and it's technical standards and frameworks. So you could think of aerospace.

[00:18:12] And aviation has a really good example of this. Cybersecurity. There are already a bunch of global principles that exist like this. Like OECD AI principles have 47 governments that have now adopted them, right? The Council of Europe's AI convention has support from, I think, 50 countries at this point. So you have a bunch of different examples where you set common standards of how stuff should be built. And then you allow – one of the other problems with trying to regulate is you have different countries that will be approaching it different ways.

[00:18:40] And so by setting standards, you create interoperability and flexibility in how things work across those regions. You know, I actually think in a strange way, cryptocurrency has been a pretty good example of this in that it's created frameworks of what is the accepted ways you want to build. And then with those standards, protocols, and frameworks, it makes it a lot easier for everyone involved to build according to the big problems people are trying to solve. So, you know, I think model context protocol or MCP is a really good example of this.

[00:19:07] This has now become a commonly accepted paradigm of how to link different systems together. And so I think what you need – the balance here is you need a real focus on trust and governance. So I think Deloitte came out with this stat that 77% of leaders consider governance as system origin when selecting a vendor. So governance is hugely important to whoever it wants. But at the same time, if you don't create frameworks for people to creatively build,

[00:19:34] we will not be able to address a lot of the things that I think AI can be really helpful for. I mean, one example I give often is so if you look at the U.S. population over 60 right now, that will go over the next 30 years from 12% of the population to 22% of the population, right? So you have a huge population shift that has major impact for healthcare costs in particular. So think about like the implication of the U.S. healthcare costs if you move that percentage of population to a retiree age.

[00:20:02] And then within that context, 30% of total healthcare costs in the U.S. right now is administrative cost. And so I think what becomes really interesting is if you can actually mobilize a lot of this new technology to solve things like administrative cost, building per day for critical infrastructure, it's a really positive thing for society. But to do that, you just need common standards and frameworks by which everyone can accept you should be able to route. And at the beginning of our conversation, we were talking about how you took the reins in January,

[00:20:31] quickly raised $100 million. But this is just the beginning of your journey, really, because Invisible has already expanded its global footprint. Phenomenal's pace this year as well. So I've got to ask, how do you keep global teams connected? And where do you see operations branching out next? Because you do seem to be enjoying so much success at the moment. Yeah, you know, it's interesting. There's been a lot of debates, I think, publicly whether in office or remote is the way to go.

[00:20:59] And we've very clearly taken a hybrid approach, which is I think if you are fully remote, you lose the kind of connectivity and organizational fabric that's so important for growth. So it's very hard to have an organization really evolving and building quickly if people don't know each other in any form day to day. Now, at the same time, there's a ton of research that actually global organizations are far more effective than just siloed local organizations.

[00:21:27] So like I think the Harvard Business Review's recent research that high performing globally distributed organizations materially outperform makes a ton of sense to me. You get a better mix of skill sets, language, expertise, culture. I mean, it's just a much better way to operate to have all of the best global expertise at your fingertips. But then the challenge becomes, well, how do you make sure you have connectivity and organizational alignment and culture?

[00:21:52] And so I think the approach that we've done, which we kind of, I think, it just kind of happened organically. I just started to notice that every time we wanted to do a big sprint initiative with our tech team, for example, they would end up co-locating somewhere to get to make progress. And so what we've done is we've now opened offices in New York, San Francisco, D.C., Austin, Texas. I was just at that office last week. London and Poland. We'll probably open up one or two more in the next couple of months.

[00:22:19] But those are not locations where people have to be in five days a week. They're places where we want co-location. As people move between cities, they do go to those offices. And I'd say most of the time folks are in the office three to four days a week with about, let's say, 30% of the company is still fully wrote. But even that group still comes to these offices occasionally. So it creates kind of cultural cornerstones of interaction, which I do think is really important.

[00:22:43] I think if you don't have the human connection and the fabric of an organization, it's just really hard to build strong culture. And so I think in my mind, if a really talented person wants to work from a beach on Friday and they are high performing and they work a lot, I'm fine if they want to work wherever they are on a Friday. But if they're never in the office, it's very hard to create consistency in the organization. So I think we've found a really good balance with hybrid.

[00:23:10] And you've already mentioned a few today, but one of the things I always ask by guests, if there are any myths and misconceptions about your industry, about your area of expertise that you might often see when scrolling through LinkedIn or anywhere where you're looking for information, you're right in the heart of this space. You're enjoying success. Any myths and misconceptions that we can lay to rest today before I let you go? Yeah. You know, I think the biggest one by far is the alarmism about job displacement. And I'm not saying to be clear on that.

[00:23:39] I don't think that means that no jobs disappear. I think the I think actually the World Economic Forum's research on this has been the best I've seen, which is it's something like 40 million of net new job creation will come from AI. And that would be 130 million new jobs and 90 million jobs displaced. And I think that that's that's how I think about this, actually, is that if you think about the biggest barriers to growth and kind of GDP productivity, it is bureaucracy, administrative time, a lack of clear information.

[00:24:08] And so I actually think what I would do is massively streamline the ability to build new things, create new things, sell more. And so at a full industry level, you will see massive growth from AI. I mean, by the way, I was one third of GDP growth last year, as an example. But the best way I've heard this explained is AI does not reduce work at the role level. It cleans up at the task level.

[00:24:33] And so the important thing about that is it doesn't mean like the idea of changing a job or taking away a job makes no sense. If you think about every job is a mix of 55 tasks that interact with other tasks. What it actually does is say, what are big buckets of repetitive work that happen in a system that I can then streamline? And usually what that does, it frees up time for people to spend their efforts on much more high value activities. So like I'll take health care as an example.

[00:24:59] You will, I think, see a lot less time spent on medical coding and a lot more time spent on being around patients and focusing on standards of care. I think that's a really good thing for society. I mean, my favorite example of this is bank tellers. So if you look 40 to 50 years ago, I think it was 30 years ago, a bunch of tech came out that basically took all of the functional work the bank tellers did and automated it. So you actually didn't really need nearly as much time spent doing manual work as a bank teller.

[00:25:29] Now, what's the number of bank tellers today versus back then? There's actually more today. And the reason for that is that suddenly that made the economics of the bank branch pretty attractive because you didn't have to, it wasn't so manual. More bank branches opened and people figured out that having humans in bank branches built a real personal connection with customers. And so they opened up more bank branches, hired more bank tellers. And there's more today, despite reducing a lot of the manual work.

[00:25:52] I think that is the parallel I see for society as a whole is we will spend a lot more time on high value activities, building relationships, developing products, being creative, selling things. But we will spend very little time on the process of gathering information, which was never really that valuable to begin with. And it never really something that should have been as painful as it is. And I think that is a powerful moment to end on.

[00:26:18] But for anyone listening wanting to find out more details on anything we talked about today, we've talked about some big announcements. I suspect there'll be many more to come throughout the year. So where can I point everyone listening that wants to find out more information about anything we talked about today? Yeah. So invisible tech dot AI is our website. I encourage everyone to check it out and then follow me or invisible on LinkedIn for the latest updates. We're both pretty active on that. Well, I'll add links to everything.

[00:26:44] And I'll also post a few links to some of the research we talked about today. Thanks for joining me today. Thank you. So much in that conversation today that will stay with me. I think Matt said it best when he said AI rarely replaces an entire job. What it actually changes are the individual tasks that make up that job. So when you think about work that way, I think it almost reframes the entire debate.

[00:27:09] Instead of asking whether AI will replace people, the more interesting question becomes how it might free people up from repetitive work so they can spend more time on things that genuinely require real human judgment. And of course, none of this happens without good data.

[00:27:27] And if there's one theme that kept coming up throughout our discussion, it was that AI success still begins with the same unglamorous work that many organizations have been carefully avoiding for many years. Yeah. Cleaning up legacy debt, messy systems, connecting fragmented and siloed data and creating the foundations that allow AI to deliver meaningful results.

[00:27:52] So it might not be the most exciting part of the AI story, but I think we can all agree it is one of the most important. So if today's conversation sparked any curiosity in you, you can learn more about Matt and the work happening at Invisible Technologies by visiting invisibletech.ai. And as always, if you enjoyed today's episode, share it with someone in your network who is also trying to make sense of where AI is heading next.

[00:28:21] But before I let you go, I'll leave you with one final thought to reflect on. And that is, if the biggest barrier to AI success isn't the technology, but the way your organization structures your data, your workflows and your teams, how ready are you and your organization for the AI future that everyone keeps talking about? As always, let me know too. TechTalksNetwork.com. Myriad of ways you can get hold of me there.

[00:28:50] And I'll also speak to you again, hopefully tomorrow morning. Bye for now. Bye for now.