Everyone wants a piece of general purpose models. Instacart has deployed ChatGPT for recipes and meal planning. The Mayo Clinic is using it to summarize patient records. Schneider Electric is using an OpenAI LLM to generate sustainability reports.
With such powerful models, what’s the need for specialized models built for specific industries, especially in climate tech?
In this episode, Shayle talks to Sam Smith-Eppsteiner, partner at Innovation Endeavors. She recently wrote a blog post arguing that there may be a market that general purpose models struggle to meet: physical industries where training data is siloed, unstructured, and private. She talks through climate-relevant examples like Cadstrom’s copilot for electrical engineers, Hubflow’s automated trucker scheduling, WeaveBio’s AI-powered platform for regulatory approvals. Shayle and Sam also cover topics like:
- Applicable cases, like cross referencing complicated technical manuals, repetitive manual work that employees dislike, and technical compliance
- The technical knowledge lost when workers retire and how specialized AI could help
- What it takes to build specialized models, including data access, vector embedding, prompt engineering, and fine tuning
- What budget categories businesses might use to pay for specialized models
- Selling the technology (i.e. the traditional SAAS model) vs selling the work (i.e. answers informed by models)
Recommended resources
- Innovation Endeavors: Specialized brains for industry: the immense potential for domain-specific AI
- Innovation Endeavors: The next industrial unicorn: Where is AI rapidly transforming the physical economy
- Catalyst: The coming robotics wave
- Latitude: How utilities are designing and embedding AI operating models
Credits: Hosted by Shayle Kann. Produced and edited by Daniel Woldorff. Original music and engineering by Sean Marquand. Stephen Lacey is executive editor.
Catalyst is brought to you by Anza, a platform enabling solar and storage developers and buyers to save time, reduce risk, & increase profits in their equipment selection process. Anza gives clients access to pricing, technical, and risk data and tools that they’ve never had access to before. Learn more at go.anzarenewables.com/latitude.
Catalyst is brought to you by EnergyHub. EnergyHub helps utilities build next-generation virtual power plants that unlock reliable flexibility at every level of the grid. See how EnergyHub helps unlock the power of flexibility at scale, and deliver more value through cross-DER dispatch with their leading Edge DERMS platform, by visiting energyhub.com.
Transcript
Tag: Latitude Media: podcasts at the frontier of climate technology.
Shayle Kann: I’m Shale Khan and this is Catalyst.
Sam Smith-Eppsteiner: There are some sort of general purpose, really high growth AI tools that I hear are losing some deals to vertical specific players because they’re better suited to that context. But on the other hand, I know law firms that are using anthropic instead of Harvey or legal AI tools. So there’s definitely anecdote on both sides. I don’t think we know exactly where this is going to net out
Shayle Kann: Coming up: We ask the David-or-Goliath question on whether it makes sense for there to be specialized vertical AI solutions for physical industries.
I’m Shale Khan. I lead the frontier strategy at Energy Impact Partners. Welcome. So as most of you know, I think my job is to invest in early stage companies that are delivering some true step function improvement in technology in categories like energy and industry. And in other words, I’m primarily a hardware investor, at least there’s always some hardware in the loop in the types of things that I get involved with. And more broadly in climate, the sectors that matter, the ones that drive the emissions and thus matter from a climate perspective, are generally doing big physical things. They’re doing industrial chemical reactions or generating terawatts of primary energy or moving big things, long distances, et cetera. So obviously AI is going to transform some of these industrial processes or at least how they’re performed. And we talked a few weeks ago about one flavor of this, which is replacing some human labor with robotics, but another flavor is using AI to augment the humans and make them more productive or efficient or safer or whatever it might be.
And here there’s kind of an interesting question, which is do we need domain specific verticalized models, tools and companies or will the big foundation models, the names that you’ve been hearing about raising eye watering staggering sums of money, will they eventually just get good enough to essentially do all this stuff for us? My friend Sam Smith Epstein or has been thinking about this question a lot, she’s a partner in innovation endeavors, which is another early stage venture firm and she’s been investing in companies in this verticalized space that she calls specialized brains for industry. I think it’s super interesting, it’s a very dynamic question and Sam has some great thoughts on it. So let’s hear. Here’s Sam.
Sam, welcome.
Sam Smith-Eppsteiner: Awesome, thanks for having me, Shayle.
Shayle Kann: Excited to finally have you on and talk about AI and physical industries in industrial sectors. I dunno, whatever we want to call it. Why don’t you start with just the high level thesis. You’ve been spending a lot of time in this space. Walk me through it. At the highest level, how are you thinking about where there is value in the applications of this new wave of AI to these big heavy industrial industries?
Sam Smith-Eppsteiner: Yeah, so I think the first thing we’ve been thinking about is whether and why and where there will be specialized vertical applied AI products versus using general tools here. So I think there’s no perfect answer. We’re hearing anecdata on both sides. I know there are some sort of general purpose, really high growth AI tools that I hear are losing some deals to vertical specific players because they’re better suited to that context. But on the other hand, I know law firms that are using philanthropic instead of Harvey or legal AI tools. So there’s definitely anecdata on both sides. I don’t think we know exactly where this is going to net out. I think my theory is that there is room for and a real need for actually vertical applied AI tools in the industrial and physical economy categories. I think there are a few reasons why I think most of them live on the supply side of this, of why technology and how technology will serve this sector will.
And this is mostly about the customer’s data. So if we think about the customers here across a bunch of different industrial categories, so think manufacturing natural resources like mining and materials construction in the built environment, supply chain and logistics, all these categories, what we’re talking about, the data has a few features for these customers. So first it’s fragmented and siloed, so it lives in these sort of legacy systems of record. So in manufacturing alone, you could imagine you have an ERP, an M-E-S-A-P-L-M and many more and these systems are often sort of legacy. They’re not contemporary software tools. Sometimes they might even be OnPrem. And again it’s across all of these different systems. So they have this very fragmented data structure where the data is conceptually linked, it has theoretical relation, but it’s not sort of pragmatically or practically tied. And beyond that there’s just tons of PDFs here. They live in SharePoint, they live in email, they live in totally different kinds of documentation storage. And so this data is just inherently fragmented and siloed.
Shayle Kann: That is especially true in physical industries. Is it more true that the data is siloed in PDFs relative to other industries or is it just that that’s true universally and so per that criterion there is an opportunity for verticalized things kind of anywhere?
Sam Smith-Eppsteiner: I think there’s probably some of it everywhere, but I think more of it here. And I think the tooling again is more legacy as opposed to, I don’t know if you imagine you’re a large tech company, you’ve probably built a lot of your own tooling or are using contemporary software as opposed to software. If you look at, I think we were looking at once companies that have built multi-billion category defining products selling to hardware engineers and the last one was started in the nineties, maybe even the eighties. So there’s just not a lot of contemporary products here compared to what you imagine the tech stack or sort of back office stack even looks like for a parallel technology company.
Shayle Kann: Right. Yeah, that makes sense to me. I mean they’re also just, I would bet the average age of the company is older than the average age of the company and obviously in software. So we’re getting through this list that you published a really good piece on a little while back, which is the list of characteristics. The way I think about it is it’s the list of characteristics that make it true or that would make it true for there to be an opportunity for a specialized verticalized AI solution for that industry. So the first one that we’ve been talking about is these legacy systems of record that exist and are siloed and have lots of PDFs and stuff like that. So that’s one. What’s the next one?
Sam Smith-Eppsteiner: So the next one is sort of related, which is a lot of that data is unstructured and so that has to do primarily with the PDF component of this and it’s a lot of textual data, which actually can work quite well with l lms such as maintenance logs and other things. But it’s also just a massive amount of highly technical and visual data. So you can imagine blueprint schematics, diagrams, even 3D models, all this kind of stuff, which looks quite different from a lot of what AI models have been trained on so far.
Shayle Kann: And is the premise there that, okay, so you have unstructured data in a bunch of different places and what we’re kind of asking ourselves here, is that a reason for there to be something specialized or could you take all that unstructured data and all those PDFs and just upload ’em to whatever generic LLM, upload ’em to anthropic or chat PT or whatever it’s going to be and it’ll synthesize them just as well. Is there a reason why unstructured data is better suited to a particular specialized model?
Sam Smith-Eppsteiner: Well, I think this relates maybe actually to the third point, which is most of this data is private, so most of this data is private and proprietary and lives on these customers cloud in a best case scenario, maybe on-prem, but certainly not in the public domain where OpenAI has trained on it. And so I think the reality is these models are just not that good today at understanding a very technical diagram, a blueprint because they haven’t seen enough of them. And so you can imagine a EC architecture, engineering construction is a particularly good example because there’s a lot of geometric and spatial representations. You need to understand that design intent, proprietary training data and most of that doesn’t look like the sort of normal tech space or video data that’s on the internet sort of pretty broadly.
Shayle Kann: Yeah, I thought that this one, the walled garden data is a very strong point, certainly from what I’ve seen and I experienced a lot of it in the energy industry, right? In the energy industry, so much of the data is private and this has been an issue in various software type businesses that have been tried to be built around the electricity sector in particular. It’s like, ah, but if only we had all of the utilities internal data, then we could optimize X, Y or Z. It’s like great, good luck getting utilities internal data, it’s super sensitive data, they’re not going to share it with anybody. So it’s definitely true and I can see how that is a hindrance to a generalized model because they don’t have access to any utilities data that’s like that. And so if you can get access to sufficient volume of it in a verticalized context, then you certainly have an advantage versus the generalized approach. Now still getting access to it in that context is no small feat, but if you’re building something for that sector and only for that sector and it’s private, then maybe you’ve got a better shot of it, I guess is the concept.
Sam Smith-Eppsteiner: Exactly, and I was just talking to a founder in one of these categories and she was stating that really important I think to her was from the beginning she made sure that access to that data and ability to train on it was a part of every MSA she signed and at the very beginning it was really challenging actually to get customers over the line on that, but now it’s become pretty boilerplate for her, but she feels like it was incredibly important part of her building a one performant and two defensible product for the category.
Shayle Kann: Yeah, that’s the barrier. You need to break through your thing that you’re building needs training data. If the training data is all private and walled, you got to be able to use your customer’s data to train. They’re not going to like that and you got to figure out how to get over that hump and if you can’t get over that hump, there’s kind of no way you end up building a product that scales or gets better.
Sam Smith-Eppsteiner: I think that’s likely true. Yeah.
Shayle Kann: Okay, so legacy systems of record, all this unstructured documentation and then the data is private, it’s a walled garden, what else?
Sam Smith-Eppsteiner: So those are the main things on the data side, which I think again, lead me to believe that a specialized product we can talk about that’s the model itself or something else will be more performant than the alternative. Again, sort of general tool, I think there are a couple of demand side factors, but I think really the main thrust is what we just talked about on the demand side from the customer piece. I think one thing we’re seeing is the great crew change, which applies to a lot of these categories where there’s a bunch of skilled experienced workers, whether that’s field engineers, field technicians, all these kinds of things who are reaching age of retirement and there’s just not enough folks to fill the gap behind them. And that means two things. One, we actually just need to do more work with fewer people in the future as we have a potentially smaller labor force to do the same work or that knowledge is actually retiring. And so we need a way to capture that expertise and sort of make sure we have a lot of that locked in into a product given that it today lives in people’s heads and that’s how these systems operate over time. So that’s one big piece.
Shayle Kann: Yeah, yeah, I mean obviously that’s true in a bunch of these sectors. It’s one of these things though that I do think is a big macro driver but is not, it’s like climate change in some ways it it’s going to take decades to play out and that’s one of the problems with it is that there’s no immediate, it’s like a painkiller, but it’s like a slow release painkiller, you know what I mean? So yeah, people are probably feeling the crunch of the lost expertise over time, but it just happens so slowly. You’re like a frog in boiling water.
Sam Smith-Eppsteiner: I mean we were looking at something in oil and gas in this space that in talking to customers, it was highlighted as a major, major pain point for them of a real fear. So I think we are at a point where at least customers are voicing to us in some categories a demand for product based on that need. But I hear your point that it’s sort of a very long-term trend.
Shayle Kann: And then the last one you listed in the piece was a good point, which is more the question of can you build it, can you build it economically I suppose is the question. And when I say it, I mean something specialized and verticalized and good enough to compete with whatever, however the performance would be of a generalized solution trying to do the same thing. So how do you think about the sort of cost to build something bespoke like this?
Sam Smith-Eppsteiner: Yeah, so I think the question is how you build it and what needs to be true there. So I think on the technical side, there’s no perfect answer, but in talking to founders who are actually building, it seems like prompt engineering can get you quite far, right? So it’s widely variable, but let’s say you can get to 70 to 90% of where you need to be from a performance perspective. I think the question is whether that’s sufficient for the task at hand and whether you’re going to consume the AI internally or the customers are going to fill the gap with human in the loop because prompt engineering obviously cheap, scalable, all those kinds of things. I think the layers beyond that would be are you embedding, are you doing vector embedding? Are you doing fine tuning or you actually building your own model? I think I’m hearing really widely variable things there.
I’m hearing founders say embedding is important, we’re doing it, but it doesn’t work very well for nont textual data. I’m hearing some folks say, fine tuning is a natural evolution of prompt engineering and I hear other folks saying, we’re going to build our own model and that’s just part of what we need to do in this category. I think the reality is we’re the very early innings of this and there are lots of ways to sort break up these problems. One company that two accomplish a pretty simple task actually has nine separate agents doing micro parts of the task and doing compliance and testing of the answer on the task in the first place. So you can imagine this is actually it’s, it’s not like you have one mega monolith model to solve even a relatively small workflow. You might break this up into many different pieces and build your own model for a tiny portion of it, fine tune something else and just prompt engineer another part of it. So I think folks are getting creative and will continue to do so in terms of how exactly you build this and it’s going to be different based on the problem, but also I think we’re going to see a lot of evolution here and my hope is that deep seeking others have shown there is a path to lower cost smaller models that are viable for startup to build.
Shayle Kann: Okay, so let’s assume the premise holds. There is reason to exist for specialized verticalized AI tools of various kinds in physical industries. The question then is what kinds of companies can you build basically? And so I want to talk through because you’ve also spent a bunch of time thinking through what are the categories of types of companies to be built here. The first one that I know you’ve referenced is I think maybe the most obvious one, which is it’s like the knowledge base. It’s like using it as using AI as a better knowledge base to solve the problem of all the data and all the information being siloed in a million different places or held by people who are retiring or whatever. So I think of this as just being a really, really smart encyclopedia that is specialized to the individual needs of the individual sector or the individual customer. Am I thinking about that one, right?
Sam Smith-Eppsteiner: I think it’s more than encyclopedia. I think it’s making sense of complexity. So maybe wanting to think about it is I think people are just wasting a lot of time today searching for information manually cross-referencing sources or making decisions based on complete data. And I think a really important part of that is again, data across systems and marrying that. So maybe one very simple example that I’ll give in construction land is say you’re on a construction project and your question is when are the light fixtures that are going on the second floor coming in? It’s a really actually a quite easy question conceptually, but today quite hard to answer because we need to do a few things. We need to process that natural language of understanding what light fixtures on the second floor means and referencing that against likely blueprints or schematics or whatever that show exactly what fixture we’re referring to and then taking that understanding of which skew we’re referring to and looking that up in your supply chain system.
So we’re talking about probably a few different sort of types of documents, a few different systems and marrying that data to answer what’s a relatively simple question actually, but today probably requires multiple people to look into multiple systems to figure out what those light fixtures are, find someone else who’s in procurement to go into the supply chain system and figure out when is coming in. That’s a not hard question, but it is challenging actually to answer given the way the data is structured and permissions and stuff across all those different systems and understanding. So I think it’s more than just search or encyclopedia, but it is understanding and making sense of complexity from all the mess that is this data here?
Shayle Kann: Yeah, it’s like a sentient encyclopedia or something. I can Encyclopedia brown, not an encyclopedia. There you go. Reference my favorite book as a kid. Okay, so that’s one category. Then the next category I guess goes a step further, which is more in the age agentic world. It’s like the ai, you call it AI that does the work basically. So that also seems, I guess on its face fairly straightforward. It’s like, I guess in your example it’s like, I dunno, go buy me another set of light fixtures for the second floor
Or something like that and it goes off and does it for you, right? Which is, it seems like that’s the general direction of travel of AI is toward figuring out what these agentic solutions are going to look like and it feels like we’re kind of on the cusp there, but not in the broader AI world like consumer AI world. We’re on the cusp and it’s starting to pop up a little bit. We haven’t totally cracked it. I still can’t have somebody go book me a trip with all the reservations that I need, that kind of thing.
Sam Smith-Eppsteiner: Yeah, I mean, so this is AI agents, agentic AI. You can think about it as a layer system of action, I think manual task broadly here, but in my mind there is something interesting here. I think the value could be in cutting time or cost but also potentially improving the end outcome. And I think what’s a good fit is anything where the work is annoying, where the person who’s doing it actually doesn’t like doing it, finds it frustrating for whatever reason where the work is already error written. I think that’s very common where you have to reference lots of data sort of across systems and it’s a multi-step process and where it’s highly variable or time sensitive, if you have super seasonal swings in the amount of labor you need to do something that seems like a natural fit for something you wouldn’t want to step up against.
So I see this more thinking about especially repetitive manual things. So one particular example on the automating repetitive manual tasks piece is a company called Hub Flow. What they’re doing is automating scheduling and supply chain and logistics, so automating that scheduling between truckers and receivers. Trucks need to come into warehouses at certain times, they’re need to be resources deployed against that today. That’s done through emails, texts, whatever systems folks live in today. And what’s interesting is Hub Flow is automating that for the trucker, but the third party, the receiver actually has the same exact workflow to them it’s invisible, which I think is an interesting part of what AI agents can do here, which is really play a role for one party in a transaction. There actually is a company conduit automating on the other side. So we may in fact have conduit and hub flow sort are chatting to schedule an appointment, but that’s the kind of example of the kind of workflow we expect to get automated.
Really you can imagine there’s tons of these up and down the stack of work at that gets done in all these industrial sectors, whether it’s payment, scheduling, this kind of automation here. So we just think there’s huge amounts of opportunity. I think the question will be often the workflows folks are starting with are narrow in nature. So the question is do you sort of build depth in capturing value in that workflow? Do you build additional workflows? You can think about this as just much, much better and more flexible and less customized RPA, right? Just much better performance.
Shayle Kann: Yep. Okay. So then the next category that you’ve laid out is the one maybe that we’ve talked about the most actually on this podcast, which is basically, I guess you call ’em copilots for discovery for engineering and discovery. And we’ve talked about it more in a discovery context because we spent some time on AI for materials discovery, which is an interesting sub niche of this whole universe less on the engineering side where I know you spent a lot of time and I find pretty interesting as well. But yeah, I mean how bullish are you on this using AI to do better discovery or I guess design is maybe the other component to this?
Sam Smith-Eppsteiner: Yeah, I’m quite bullish. I mean you just look at what’s happening in software engineering, right? I mean I think the way Cursor and tons of other software engineering tools are taking off is pretty wild and I think it actually speaks to, we may even see multiple big companies in a given category in software engineering are seeing folks who are helping you with your existing code base folks who are helping you with new projects, folks who are helping you translate from one language to another that those might all be different sort of companies doing that. So I think this is really similar to agentic AI of how can we do work better, faster, cheaper, and actually can we take the parts of the engineering work that engineers don’t enjoy doing off their plate? So I have an early investment in a company called Strom. They’re building a copilot for electrical engineers on the PCB design side.
They’re starting on the verification and validation end of things like checking downstream as opposed to design upstream. What’s good about that is the engineers often don’t enjoy doing that. I have to spend a lot of time error checking their stuff. So it does delight the user in that way and I think the layer cake they’re building here is actually quite replicable to other engineering domains as well, which is they’re understanding the intent of the engineer but then parsing the information on the individual components of the board and what the limits are. They’re then combining all that to understand how a board would theoretically work, and then from there, understanding what simulations need to be run to make sure that the system actually does operate that way. And then two, actually automating those simulations, that same sort of layer cake of work I think could be applied to lots of other spaces. So I mean you can see when it comes to electrical system design, we imagine there’ll be plenty more here in the real world, so structural engineering and sort of other domains as well.
Shayle Kann: I like your framing of just look where the work is annoying and you’ve probably got that’s a good beachhead at a minimum and there’s a lot of annoying work in these sectors for sure.
Sam Smith-Eppsteiner: Yeah, I think what’s hard is there isn’t natural tension here if you are selling technology to the end user where they don’t want to be automated away. So I think it’s a natural starting point where it’s like augment them by starting with a thing they don’t want to do and maybe over time you do unlock more and more of the work, but it’s like how do you get that end user sort of delighted and bought in as opposed to afraid.
Shayle Kann: Right. Right. Okay. And then the last category is one we haven’t talked about much here, but I’ve seen certainly firsthand in the energy sector as well, which is more AI for compliance and broader risk mitigation I suppose as well where compliance in a lot of these sectors, but particularly the heavily regulated ones of which energy is one, is a big hidden burden. A lot of people, a lot of work, a lot of time, a lot of documentation, and so it seems one to me that is kind of obviously ripe for some degree of automation to the extent that it is trustworthy.
Sam Smith-Eppsteiner: Agreed. I mean industrial companies just have huge amounts of compliance, both to your point on the regulatory side, but also just even internal. It’s like what are your standards and policies and operating procedures that you have to comply with even internally, and it’s a total cost center and bottleneck headache, and I think it’s a really good fit actually for ai. It’s really textual in nature. So I think the work here is to sort of parse and understand the rules and then some combination of understanding your work and mapping it against those rules, realtime monitoring and flags or management of operations and then potentially also relevant for design support on new projects. So you can imagine, I’m sure you’ve seen plenty of us yourself in energy development. This is relevant in thinking about citing optimization and permitting of how do I even choose the right sites where I feel like I could get away with a permit here there’s a good path forward, let alone downstream and actually managing it.
Shayle Kann: That example is a good one to raise a broader question that I often have, which is there are some places where intuitively it makes a lot of sense to introduce these kinds of solutions because it is labor intensive and annoying and so on. But on the other hand, they don’t seem like that big an opportunity. This is how I feel about the citing optimization thing for developers. It’s like, great, I do think there is value in that. I don’t think you can build an enormous venture grade business just doing that. And so the question is always, am I missing something about like, look, this is a beachhead. You’re entering with this thing and then you’re going to broaden out with those existing customers or with other customers, you’re going to build some kind of data moat by doing what you do or is what we’re going to end up with this suite of hundreds of relatively small companies solving a real problem, but not with a big enough market opportunity to be generational category defining type companies.
Sam Smith-Eppsteiner: I don’t know, and I think we’ll have varied answers here by sector, but I think one way we’ve thought about it too is how sticky are you? I think what’s hard about the upstream stuff like citing and permitting is most of the data you’re using is external, right? It’s third party to understand the land and permitting situation as opposed to internal to the customer. And when we looked at that category when we talked to customers, it sounded like they were sort of readily changing tools, which is less likely to happen if you’re sort of deeply integrated with a tool and they’re building a model off of your own data that is already highly performant. It’d be hard to transition to another player. I also think it’s a question of just how big that specific opportunity is and how much build opportunity there is on top of it.
So this is a very analog from a different part of our portfolio, but we have a portfolio company called Weave Bio. So from the health and bio space that for pharma and Therapeutics is helping them to build INDs, which is a step in the first stage of the FDA approval process, and they’re significantly speeding up the time it takes to prepare one of these media applications and improve quality. We actually feel like, well one that’s again heavily integrated into your own data and systems in a different way, but two, that is actually unto itself a big large opportunity. There are other plays we’ve seen though where that first workflow, whether it’s compliance or agentic AI seems small, but it’s like if you could embed into the payments layer that sits on top of that, then maybe that actually becomes quite significant. So coming back to for example, scheduling between trucking and receivers, my understanding is that there are fines for if you’re late for missing your delivery appointment, can you actually then automatically schedule those fines and actually apply them and then capture some of that flow through. If you can start to embed into transaction layers that sit on top of the workflows you’re on, then all of a sudden, obviously you’re talking about a much larger TAM right there, and additionally, you potentially could build a sort of obviously other workflows on top of that. So I think the question is just how big is the problem? What can you build on top of it? And as it always comes back to it, it’s like what’s the ambition of the team and where do they want to take it?
Shayle Kann: Yeah, I think that often it does come down to what is the, clearly the team as well, but it often comes down to what would be the next thing, what can you imagine coming after this and is there a next thing with this customer base or do you have to believe in some other version of expansion that it’s harder to picture?
Sam Smith-Eppsteiner: Totally.
Shayle Kann: Okay. So I guess final question for you, stepping back, I mean the other thing that’s interesting about this revolution of all these new AI companies is that there is some innovation on the business model and what they’re actually selling because they’re coming from a world where software companies sold. There was a pretty clear thing software companies would do, and it’s usually software as a service and you charge a subscription and it’s seat based or utilization based or whatever, and then you stack up your a RR numbers and you’ve got all your metrics. In AI world, there’s at least some discussion of turning the pricing model at least on its head. I’m curious what you’ve been seeing from all these companies that are out there doing AI in the physical world. Is it a standard software as a service type offering or is there innovation there?
Sam Smith-Eppsteiner: Yeah, so I think the choices in a very black and white sense are sort of sell the work or sell the technology. So selling technology is sort of the traditional SaaS model you’re talking about of offering your AI product to the user to use on their own selling. The work is saying, Hey, I’m just going to actually do this whole task for you. I’m going to consume the AI myself and sort of obfuscate whatever human in the loop or services are needed on top of that as part of that service. I think where we’re finding that to be most interesting to customers is where they’re already outsourcing. So if you’re already outsourcing your compliance work, for example, to a third party consultant, you’re probably sort of willing to outsource that to a different company instead and especially if that company is going to staff up.
So that’s why the question, what kind of services are you actually providing on top of the technology? But if you’re actually going to provide me an account manager or a case manager and someone who’s experienced in the space and is going to effectively be my new consultant, but by the way they have technology behind them now, then maybe that could work. I think this will have interesting implications to margin profile actually as will the technology broadly, and I think this is something we should all potentially be thinking and talking about more, but inference cost is non-trivial. I think for a bunch of the AI companies both in our portfolio and that we’ve seen, and this is not going to be a 90% margin SaaS, we may be talking about 40 to 70% gross margin based on inference costs, and obviously there are ways to manage that and hopefully costs will go down over time, but that alone I think will make these businesses actually different than the companies we’ve seen before.
But coming back to sort of business model, if you were also on top of now we have inference costs, significant layering on services to deliver a full service to your customer, this could be a much lower profile margin profile business than we’ve seen before. That being said, obviously the trade off typically between a full stack business and a technology business is what you trade in margin. You get in top line, so if you’re selling services as opposed to selling technology, there’s often a much bigger dollar value at stake in doing that, which I think relates to a couple of themes I’m seeing for adoption here that are interesting. I think this is maybe one of the first times where hopefully, although I think we’re early in actually seeing this play out, the AI companies could actually draw from labor budgets as opposed to software and technology budgets.
I’m sure as you’ve seen, these industrial companies often have limited willingness to pay for software. It’s not always the highest priority, not something they’ve budgeted for, but they spend a lot on labor. And so if there’s an ability to say, Hey, I’m augmenting your labor force, making them more productive, cutting costs, et cetera, and I can actually tap into that budget, that’s a much larger pool to pull from whether that actually plays out in practice. I think we’ve yet to see. And then I think what we touched on before, the other piece here is just like in these categories, you definitely need an enterprise sales motion. You need top down, you need to manage complex stakeholders, but I think this is one where AI can delight the end user. So the person who’s actually sitting with a tool every day, if you can take on that annoying work and just make their life better, then you no longer have that tension between, oh, the CEO sees ROI, but the process engineer is afraid they’re going to lose their job because they have a tool that they love and makes their life better every day. And so ideally this is a way that things can actually move faster in terms of a sales process too.
Shayle Kann: Yeah, I mean the only thing I’d add to that, I think this is really difficult to sell apriori, and so it’s a thing that companies are going to have to prove over time, but ultimately will become incredibly valuable and sticky, which is you mentioned tapping into the labor budget. I think actually ultimately some of these things can tap into the capital budget, into the CapEx budget because for example, for applications with regard to maintenance, maintenance is a huge cost. Or if you can end up making a case that you’re going to extend the lifetime of existing assets, and so the capital budget can decrease by 10%, you don’t need to replace some existing assets as often. Those are where the really, really big numbers come in, but you have a really hard time upfront convincing a customer that is going to be true, and so you have to come in with some other thing and then prove that one.
Sam Smith-Eppsteiner: Yeah. But yeah, I was going to say, I think the CapEx piece is a longer term play, right? Because that’s effectively a value-based pricing model and the value accrues over a long time versus labor you can prove kind of immediately at least. So it’s still some sort of value-based pricing related model, but if you can show that you can do the same work with 20% of the people, you can maybe take a week or two to do that as opposed to improving your CapEx or facility is going to last for 10 years longer, which you won’t know for 30 years.
Shayle Kann: Alright, Sam, this was awesome. Whirlwind tour of AI for physical industries. I’ve been loving the stuff that you’ve been putting together on this and it’s fun to talk through it with you here. So thank you.
Sam Smith-Eppsteiner: Yeah, thanks so much for having me. It was fun.
Shayle Kann: Sam Smith Epstein is a partner at Innovation Endeavors. This show is a production of Latitude Media. You can head over to latitude media.com for links to today’s topics. Latitude is supported by Prelude Ventures, prelude, Beck’s, visionary, accelerating climate innovation that will reshape the global economy for the betterment of people and planet. Learn more@preludeventures.com. This episode was produced by Daniel Woldorff. Mixing and theme song by Sean Marquand. Stephen Lacey is our executive editor. I’m Shayle Kann, and this is Catalyst.
