Why Owning Your Data Is the New Moat | Primary Venture's Emily Man

0:00My framework is the roadkill, the cockroaches, and the cheetahs. The roadkill is old school software companies. Maybe they were point solutions. Maybe they were low value that like are ultimately just going to get run over by AI. The cockroaches is the second category of things that can survive a nuclear blast, right? And the nuclear blast being the AI wave that is happening. But they might not have specific AI related tailwinds, payroll technology, things that move money. The last category being cheetahs are folks who have like a real opportunity to become AI native and to embed the way AI works into each one of their work clothes and their offerings.

0:40I think the best example here is Intercom which started in 2013 as a company that, you know, pre-AI generation and did a sort of refounding moment. Some companies will make up the curve here but there's others that I think are at a real risk. Is this thing on? yesterday's price is not today's price

1:11welcome back to run the numbers my name is cj i'm a recovering tech cfo and i talk to the world's top cfos as well as the investors who fund them on today's show i have emily man of primary ventures on and she is kind enough to take us through the five layers of the ai stack and explain where value accrues. We then touch on the changing P &L and why it is perhaps the most exciting time ever to be a CFO as AI works its way into every line and department throughout the P &L. We talk about the race to capture context. And by context, I mean organizational alpha that is maybe sitting somewhere out there that needs to be read by AI and then put into workflows.

1:50And finally, we talk about things like MCP, going headless and harnesses, what those terms mean and what it means for your data. All this and much more coming up. Emily, thank you so much for joining me on the show today. CJ, I am so excited to be here. It's a different scenario than when we first met. I know, right? I think the last time we chatted, you were a CFO of Carts Tech and I was a growth investor. So lots has changed. And you're at primary now. What's most exciting that you're working on today? Gosh, what isn't exciting at this point in time, right? As you can imagine, like we are at the forefront of taking intelligence and bundling it into useful capabilities and the types of things that we're seeing on the ground on a day to day basis.

2:34I don't know. It's just been the most exciting in my entire career to be an investor. You have such a succinct way of taking these complicated things and packaging them like you just did into something that's easier to understand. And like you said, we have all this intelligence. Now we're trying to package it into stuff that's useful. And one of the things that you've taught me is that there are different layers of the AI stack. Now, I work with a lot of finance people, a lot of CFOs, and we pay all these different bills for stuff, but we don't always understand like the technical underpinnings of it.

3:07I'm hoping just like set the table for us here. What are the five different layers of the AI stack? Maybe just to start, like this is all stolen from Jensen, but I think he has a very distinct way of talking about the layer cake that is underpinning the AI era, right? And I think about it as like a supply chain from what do you need to get from sort of raw material all the way to finished product. And so on one end, you have energy, right? So that's power, that's electricity. Those are the inputs into the hardware itself, the second layer being chips. From there, you take the chips and you put them into infrastructure.

3:41And so that's where you wire everything all together and make it useful for a set of models that are built on top of them. Now, the models are probably something that we can all easily understand. They have dominated the news cycle for the last probably two years at this point. But at the very highest piece of that, it's the application layer. So the applications that connect what the models can do with the end human output or the human outcome. Got it. And where has attention and funding focused on so far? So for the first sort of era of this, the AI boom, we've been in this massive CapEx wave where everyone is focused on building the infrastructure.

4:21So call it like the layers two through four, right? The chips, the infrastructure and the models. And if you look at all the funding, that's where a lot of it is going. So in the first inning, we've had this massive build out and we've seen that in every single technology shift, right? Whether that was, you know, with the railroads way back when or with the broadband boom in the early Internet days, you sort of have to lay the pipes in order to build everything on top of it. And so if we think about where we have come so far, that's where a lot of the funding, the effort, the time has focused.

4:55And we here at Primary, through my partners, have made some incredible investments along each part of those stacks. But I myself focus a lot more on that last level, the applications. OK, cool. I've been trying to get smarter on the energy component as well. I'm going to be interviewing the CFO Craig of Okla at Brex Mode in Chicago in a couple of weeks, which I'm excited for. Like, I didn't realize as a layman, we just didn't have enough energy. It just seems something so foreign to me. We've looked at some really interesting things in that realm, up to and including sort of small scale nuclear reactors that are co-located with data centers.

5:30to help them power the underlying infrastructure. We all need electricity and we've all experienced sort of peaks and boom and bust. But the way that the power grid works for us today, it's just not going to be set up anywhere near enough for what the demand is going to be from all the data centers that are coming online. And I think both me and you probably spend most of our time on that fifth component or fifth pillar, the application layer. And I've heard you say that the next era of AI adoption will be defined by the diffusion of AI into the broader economy through apps, right? Can you expand upon that?

6:06We are just so, so early, right? Like it's like less than 10 % of the world has even heard of and or tried in LLM in some way, shape, or form, let alone become a power user. The analogy that I kind of like to use in where we are today is we've created models that have become incredibly intelligent. So think of them as a bright college grad. For them to then become useful in providing value to society or to doing jobs, there's still a level of like training and refinement that needs to happen. And in my mind, in our minds, like that's the job of the application layer. I was talking to Claire Bramley, the CFO of Xero a couple of weeks back, and she was saying something similar.

6:51Like we work in Silicon Valley. We work in New York. We talk to and work with people who are just terminally online and early adopters. And she said, think about a bakery in Auckland. Think about a plumber in Boston. These people make up the majority of the economy and GDP, actually. Tech is even a smaller component of that. The application layer has not diffused down to the majority of people yet. Even for the 10 % or so of the economy that has had some sort of an interface with an LLM, the vast majority of that is like asking queries to a chatbot, right? And so like we still have so far to go before this starts to really get harnessed and used in a way that applies to our day-to-day lives.

7:36So when you think about applications, Emily, is it ways to harness the intelligence of the underlying LLMs in a way that's more uniquely useful to whatever, I don't know, sector you work in or job to be done? Is that essentially what applications are doing, just reconfiguring the power of an LLM? It's a great question. So if I think about the job of the foundation models and what they're doing, we debate this constantly internally, right? It's like they are pushing a frontier of what frontier intelligence means. ChadGPT first came out in what the fall of 2020. It was kind of cool. Like it hallucinated a lot.

8:15It didn't give you the right answers. It couldn't tell you how many R's were in the word strawberry. Yeah, I remember that. Right. And I remember back then the conversation was like, well, alums can't do math, like no level be able to do math. And well, now we're at the point, fast forward, just like three years later, they're solving fields, metal level math question. And so whenever I think about like where Anthropic and OpenAI, Gemini, all of these major labs are focused is they're really kind of pushing the frontier on how smart these models can become. I think of them as like producing something that is a very smart generalist.

8:46If you think about the way that our society works today, that still doesn't get you all the way there. This is why I like the college grad analogy in my mind is because it's someone who's high horsepower, is tireless, has a ton of energy, but like has zero context on your business, your domain, the job to be done. Right. And so there is kind of this last mile piece that needs to happen of onboarding, quote unquote, this intelligence and bringing it into an organization. And so when I think about the application layer, it's about bridging that gap between what the models can do and what the models will be able to do as they continue to get better and better, driving that to real business outcomes and to real output.

9:28Emily, in five years, somebody may clip this and say this was the dumbest question CJ has ever asked. Emily, is there a point in which the LLMs are smart enough or we don't need increased intelligence? Because I'm thinking about a lot of jobs to be done today that I do. And honestly, they don't require a Fields Medal in physics for something. What's your take on that? Is there such a thing of diminishing returns at a certain point of how good these LMs are? It's not a dumb question at all. And I think that that is a topic that is being debated. The way that I personally think about it is these models are going to continue to get smarter.

10:03There is going to continue to be a frontier that they push down further and further. But at the same time, just using the human analogy, just because someone is a world famous physicist or a mathematician, that doesn't mean that you can drop them into any other domain and they will automatically be amazing. Some of it comes from expertise. Some of it comes from real world experience. Some of it comes just from like fine tuning our own models, which is our brains, around a certain area. It's funny that the seminal paper in this field was on the transformers was called attention is all you need.

10:37That is the soundbite that I keep coming back to when I think about where LLMs are going to play in society is like, yes, focused attention, craftsmanship and honing that is going to continue to be important. This may be apocryphal, but I heard this story of this gentleman who at the time had the highest tested IQ ever. I think he scored 220, but he wasn't great at applying that intellect and raw horsepower to problems in the real world. He could ace a test. He could solve these math problems, but it wasn't useful from like the way that society would reward someone. And he was a bouncer at a bar.

11:15This may be a silly analogy, but I think about some of that when it comes to the power of LLMs. They can be as useful as you can configure them to the jobs that you're working on. And a lot of that will fall to the application layer. So something like a Harvey or a Legora to making it more specific to the legal profession. I think that's totally correct, right? And it's a great analogy. We're continuing to push the frontier of what is frontier intelligence. But the next 10 years are going to be about how do we harness that and use it in the right way that is appropriate for the job to be done.

11:49If you work in finance, you know what time of year it is. I'll give you a hint. It's our Super Bowl. Budget season is here. Woohoo! So how many versions do you think you'll get up to this year? I once got to V147 in the good old Excel days. Beat that. Actually, please don't. This is when your planning software should earn its keep. You don't have to choose between either Excel or an inflexible ballooning platform that takes a consultant to change. Abacom brings you the dynamic capabilities you love from Excel while preserving the rigor and governance of an enterprise suite. I met Abacom's founder and CEO, Julio Martinez, more than three years ago.

12:26He and his co-founder, Jorge, came out of finance, and they built a planning platform that can run a complex business without becoming another system that requires a babysitter. Abacom puts your actuals, forecasts, models, and business context in one place, so you can update the plan, run scenarios, and answer, why did this change? In fact, you can even ask it in plain English. You might not even have to spell it right thanks to AI. That's why I'm so excited about Abacum. Go to abacum.ai slash mostlymetrics. That's A-B-A-C-U-M dot A-I slash mostlymetrics. Or just tell Julio that CJ sent you. That'll probably work too.

13:04Remember when pricing was simple? One product, one subscription, one invoice every month. Those days are over. The AI economy is changing how everything gets bought and sold. We're seeing usage based pricing, hybrid contracts and something called a credit. The finance systems built for vanilla subscriptions weren't designed for this. This isn't just a rev rec problem anymore because outdated revenue architecture can kill a great pricing idea before it ever reaches the market. Right Rev lets you recognize revenue in whatever shape it comes in. Whether you're launching consumption credits or testing entirely new commercial models, RightRev gives finance the flexibility to support whatever pricing model comes next.

13:44So product teams can keep on innovating and finance teams can keep pace. If you're architecting revenue for the AI economy, learn how RightRev can help. Visit rightrev.com slash CJ. That's me. That's rightrev.com slash CJ. A wise man once told me that every startup is really just a loosely functioning disaster. Ain't that the truth? If your startup needs a little extra support, and let's be honest, who doesn't need a little help now and then? You're in for a pleasant surprise with Fidelity. Fidelity Private Shares helps early and growth stage companies stay investor ready with cap table, data room, and scenario modeling all in one place.

14:21A messy or missing cap table might not just slow you down, it could cost you your next fundraising round. VCs are flooded with pitches, and if your equity is confusing or missing, they'll move on fast. Fidelity Private Shares gives founders the structure and simplicity to focus on what actually matters, building your company. If you stay investor ready, you don't have to get investor ready. Check out fidelityprivateshares.com to learn more. That's fidelityprivateshares.com. I want to talk about how this all changes the P &L. I think it's the most exciting time ever to be a CFO because every line in the P &L is changing to accommodate AI.

14:58And I'm starting to see all of this as a budget envelope where you have a total amount you can spend. And that can be on people. It can be on tokens, tools, travel, coffee. Does the CFO's role become finding the optimal Pareto mix between human capital and token capital? I think that's a thousand percent correct. For the benefit of the audience, I have spent most of my career investing in CFO technology along the curve. And over the last sort of 15 or 20 years, you saw this major shift from the operational CFO to the strategic CFO. Right. And I think that the next 10 years are going to be about taking the strategic CFO and continuing that to the I don't know what you call it, the super strategic CFO or capital allocator CFO effectively.

15:40Right. Because at the end of the day, you're starting to to think about, right, like, OK, I have a set of business problems that I need to solve. every single line item on my PNL is changing and in flux. Just as like a very concrete example or a thought exercise, there's a lot of talk happening right now about model routing and lower cost model, whether that's the open source Chinese models versus, you know, using Fable or Mythos or whatever the most expensive model that you can get from one of these core labs is. But I actually think that we're just trying to break humans back into the conversation as well, right?

16:17It might, in some ways, be faster for CJ to pick up the phone and call someone and be like, hey, this is the way that we should solve this problem versus having an army of agents in a set of loops and an orchestration across a bunch of different models to achieve the same task. And so the job within an organization, I think every department head should be thinking about this. Ultimately, it should all roll up to the CFO and then the CEO is how are we thinking about where and when we should be using AI and what kind of AI and where are humans still going to bring their unique capabilities? I don't believe in a world in which if AGI comes and everyone gets replaced, that's a whole different set of problems and questions that we're all grappling with.

17:01But in this interim period where we're going to see a hybrid world of agents and humans coexisting and co-working, there are a ton of second order problems that need to be solved from there. What's the ratio of payroll to compute you're seeing in these early innings? There's a very widespread, right? I think we're still so early on the adoption curve for most companies. But if we look at like the top 1%, Anthropic reportedly spends around 2.3x their payroll on compute. That is Anthropic and compute is a core part of their business and what they're producing. But if you look at the data, the top 1 % of enterprises today are starting to see compute costs as somewhere close to like 40%.

17:46That is a very significant number. And so you play out the tape. I see very easily a world in which inference or compute costs are as important or as large of a P &L line as call it G &A. And what's fascinating about this is where we decide to, this is the art versus the science of allocation within a company. What part of that compute ends up in your cost of goods, which therefore impacts your gross margin? And what part of it is looked at as getting more efficient within your org and ends up in OpEx? Because I do think it's a nuanced discussion of how you count that burden from the compute and then how you measure the ROI.

18:28How are you, one, budgeting for it? Where is this being accounted for on your line items? Two, how are you thinking about it vis-a-vis the employee base as well, right? So, for example, if I'm a star seller or an engineer, if I'm the CEO or the CFO in that seat, I'm like, use as much as you want. An engineer having access to the latest and greatest models means that their productivity difference can be potentially tax, right? I think right now we're still in this phase and we see this across all of our portfolio companies too, where everyone's kind of wised up to say we're not token maxing, quote unquote, but we're trying to get everyone within our organizations to adopt and use these tools.

19:10And so there haven't been any wholesale restrictions around what you can or cannot use. But you can imagine a world over time where it's like you are an ops analyst and you're running a pretty basic standard report every single week. You don't really need the most expensive or frontier intelligence for that. There's plenty of models, whether that's older version and or an open source model that can achieve very much the same result. And so I think what we're going to start to do is see very much like what you did in the corporate card days. there to be certain types of like cost controls and or budget controls around model and inference usage at an employee level over time.

19:48And that's going to be something that the CFO and the CEO are going to have to manage. It's almost as if we're going to need this orchestration layer. To your point, I remember the first time that we had virtual corporate cards and if something was out of policy, it would stop you. You couldn't spend it anymore. I don't think we have the same visibility or ability to toggle up or down the volume within our companies and all the way down to the role level yet? We really don't. And the thing is, we just quite frankly haven't needed it yet because everyone is still establishing the baseline, right?

20:17I've had many CEOs tell me that it's like, we're at the point where we've gotten in front of the company and told people like, hey, if you're not AI native in the next six months, there is not a job for you here, right? And so that sort of top-down pressure, I think, is what you're seeing a lot in the consumption numbers these days. At some point that that is going to normalize and rationalize towards cost. I was speaking to the CEO of a growth stage venture-backed company who told me last week that their AI spend went from$200 ,000 last year to a run rate of$4 million. Think about that as the order of magnitude difference with no end in sight.

21:00That is something that I think we're just continuing to see. And the pricing on the models is also continuously in flux with the way that compute pricing is changing and where the bottlenecks are in that sort of layer cake that I laid out in the beginning. So we're just like in early innings in a way that I don't quite know where everything is going to land. Many of these companies, the compute costs are going up like that, but it's not like they're taking other costs out of the budget at the same time. Even though you hear about some companies downsizing, the majority of what I hear are just certain companies who aren't backfilling.

21:32So when somebody leaves on their own accord, they're just not hiring somebody net new. So there's not nearly as much cost coming out of the P &L as what's going on. Yeah, I think if you look at Morgan Stanley ran a survey on this of where is the budget for AI, there is a good portion of it is coming from net new budget, maybe something like 30 on the board of magnitude of 30%. But the vast majority is being taken out of existing budget pools, whether that's for software applications, or broader IT projects. When it comes to labor, what you're seeing is almost counterintuitive, where the top percent of spenders of AI are also hiring at the highest rates.

22:09And now I think that that might come down to this being a very specific type of hyper growth company. I don't think that there's anything to support in the data just yet that we're seeing sort of a wholesale labor replacement starting to happen. I don't want to scare people, but is there the risk that we are using all these tokens in a subsidized fashion and that they could actually be three times more expensive once the dust settles? I think it's going to come down to the level of competition that exists between various frontier model providers. I mean, you're already seeing the spread widen between what open weights models can do and the frontier labs.

22:45And the cost difference is anywhere from like from 100x. My belief is that like if cost becomes a really strong driver here, there's going to be pressures that bring that into control. And there's ways to mitigate. I want to follow up on that because Alex Karp went viral for talking about how orgs should control their weights. Can you explain this to me? What was he really saying? Alex Karp has been going viral and, of course, in a sort of self-interested way for what they do at Palantir. But I don't think it's incorrect. So he's been saying about how if you're controlling your weights is control of your faith.

23:21And at the end of the day, the question is sort of like, why are you letting these labs in the door who are data hungry, context hungry, giving them access to everything within your organization to allow them to train and approve their models? You're effectively like leaking your organizational IP. If I understand what Palantir's CEO is saying, you should be able to own your data, your IP, and your information, and not just give it up to these LLMs. And in many ways, it reminds me of the on-prem versus cloud argument all over again. I remember working to back up in a recovery software company, Emily, and there were these banks who said, we will never host data in the cloud.

24:06Does this have any parallels to what's going on today? There are absolutely parallels, but I think that this is at a totally different level. Here's the way that I think about it, right? The example that I always use is the Toyota's Kaizen process, right? And so the way that I think about it or talk about it to folks internally is the way that car auto manufacturing used to work in the United States is you had all the pieces together, you piled an entire car, right? You built everything to the ground and then you tested it. What that meant was you had a very high sort of failure rate and or problem rate with a number of vehicles, which led to a ton of wastage.

24:44Toyota piloted this process called Kaizen, where they made each individual worker on the line responsible for their component part and responsible for everything that had come before it. And so if there was any issue on the line, you were expected to pull the on-down cord, stop the process, and report that to make sure that none of the other pieces that were coming down the line had the same issue. This has been widely studied as an organizational process and how this elapsed. But ultimately, the business results were it allowed them to gain 8x market share in the U.S. car market, automobile market, over the next sort of 30 years of being in the space, right?

25:20Because they had rapidly lower costs and they had rapidly higher quality of the cars that came off the line. And so this is what I call organizational alpha, right? It is your personal way of doing things as an organization or as an enterprise in how you show up with your customers. It's how you process information. It's how you make decisions when presented with the same set of inputs. That is what has allowed for competitive advantages of one company over another. The easiest way to understand this is through consumer brands. And each one of us has an airline of some sort, right? Because they've all made slightly different decisions around like, well, you know, Delta has better customer service and if they're more pleasant airline to fly with, maybe United has a better loyalty program.

26:05And each one of us as consumers at the end of the day is making a decision around what we prioritize. Like those are organizational level decisions that are being made. And that's exactly what these models are now codifying. If you were to build a customer support agent, I think that's one of the easiest things to understand. The way that a Delta agent responds to you or handles a situation might be very different from the way that a United agent responds. And those come from very, very different business values. And so when Alex Karp is talking about like, hey, you're letting all of your IP soak into these models, what he's effectively saying is like you have developed over the years of making profits and building a great company, all of these ways of doing things as an organization.

26:52Why would you let that go and be trained into not just like the base models itself, but potentially into your competitors' hands?

27:06underlying model. Okay, I'm sold. I want that. I want to continue to compound value over time. But now I'm worried, Emily, that if I do this and I control or own my own weights, am I going to be left with a stale model because I'm not using these large language providers like an open AI or anthropic who are at the bleeding edge? The misnomer is a little bit of like the own your weights, right? I think it doesn't come down ultimately to the actual weight, But it's like, what are the systems that you have in place to manage context? I'm a big believer that we're at the brink of the context force.

27:39Every player is in the process of trying to codify and utilize the context that we'll have. And the thing is, like, these LLMs, they were trained on the information that's available on the internet. That is a lot of information. But that is not barely scratching the surface of what gets done in a day-to-day basis, right? Like 99 % of society and the economy is not yet codified into data or information that these LMs can process. And that is sort of the next frontier. And that's why you're seeing them all buy data at rapid clips. Well, it's also a really messy problem because it's not like we've been going about our jobs or lives saying, let me do this in a way that can easily be digestible by an LM.

28:21I was talking to the CFO of Zora, and he was saying that implementations are going to get faster using AI to get you into new software faster because it can digest a lot of data. But there does have to be this investment of whether it's through forward deployed engineers or through some sort of data ingestion protocol to get people up and running. I think that that is the next frontier that all of these organizations are tackling. And that's why you've seen the number of job openings for, you know, four deployed engineers, 12x over the last couple of years. We've both worked in large organizations.

28:59There is a standard operating procedure for that things are done. That doesn't always capture the full picture. There's a ton of nuance in the way that things ultimately get done and decisions are made. And it's that gap between the standard operating procedure and the actual operations that is not codified or captured. And what we're doing in the meantime is throwing humans at it to try to make it talk to each other. I think that as we think about where we're investing or where we're spending time, there's a handful of different categories. On one end, there's what we call new sensors. And so these are things that, whether it's a hardware component or a piece of software, like Granola is a great example.

29:41It is starting to capture information that previously would have just lived in maybe a summary at best or sort of floated away in the ether. And so that is information now that can be captured, codified, and inputted into an LL. in a way that could have been before. And so a lot of the work that is being done by these forward deployed engineers is one, figuring out where the sources of data are. Two, what is the unified ontology or the data structure as a capture of what you're trying to achieve? And then it's just really getting like all of the messy files in your organization into the right boxes so that you can start to do work on top of it.

30:20You touched on some examples of organizational alpha. So if every company can buy inferences Is that by definition make inference beta? Is the model the product at the end of the day? Right. Or is the product model plus context plus workflows? If it's the former, then you're going to get that out of the box. It's going to continue to get smarter. We're going to continue to see improvements in what these base models are doing. That is just inherently beta. If you're starting to capture something that is proprietary, has additional components that the base models are not going to understand, you can build areas of alpha.

30:52You've made the distinction between models and loops. Can you explain the difference? So the way that I think about it is the model is something that is provided to you through one of these frontier labs in OpenAI or in Anthropic, through an open source provider. There's plenty out there that you can fine tune on your own. The way that I think about loops is how are you then taking that and structuring, breaking down the jobs to be done into discrete component parts, aligning those to business metrics, business KPIs over time, and making sure that the system that you've built for yourself continues to get better and better.

31:29So this is a lot less about like the intelligence that is coming through the API spigot and more about like, OK, how do I then take that and harness it into something that's useful? Maybe to make that like super concrete. Right. And I keep coming back to customer service because I think it's like the easiest thing to understand. Let's say that you have a system that you're like, all right, I want a set of autonomous agents that go and answer support tickets. But you tell it its goal is to close the highest number of support tickets it can, right? Believing that that's going to be what ultimately drives business value for you.

32:05And I've seen this happen with my portfolio companies. And I've seen this happen in instances where we've tested things out where the model given that goal or the agent given that goal then decides, hey, actually, like the best way to close as many tickets as I can is to just refund everyone. Or to tell them, like, sorry, there's nothing I can do. Being very precise about tying the goals of the system with the business outcomes is that messy middle that I think the application layer is going to solve. I've heard that software is losing its head. It's going headless. Can you explain that to me?

32:38The UI and UX layer of the past is no longer the UI UX layer of the future, right? I think when we think about software, we think about whether it's opening an app on your computer or your phone or it's typing a website or like logging into Salesforce, right? Like there is some interface by which the human puts information into the system. A lot of software today increasingly is living and meeting people at the point of need. We have companies in the portfolio that are building MCP servers to service the data that is needed by an end user in a cloud interface, if that's how they prefer to live, or running autonomously in the background and ultimately sending an email, right?

33:22And so effectively what I think is that increasingly while there's a lot of deterministic structure that happens underneath the hood, the way that we interface with software is changing rapidly. And it might not be a landing page with some cute dashboards and some reports anymore. You're saying I don't have to re-authenticate when I log into Salesforce five times a day? Well, at least you're logging in. Think about the number of sales reps that are not putting in the information into their CRM. They're just getting bunt about it on a weekly meeting from their VP of sales. And I actually think that that's part of what's been the most valuable thing about AI is that the data capture can happen autonomously in a way that is updated constantly.

34:05And the methods by which that is happening runs in the background in a way that you're not constantly prompting humans to have to make sure that the software works for that. So if software becomes headless, what do you think this means for companies pre-2022, SaaS companies? That's a big debate right now, right? Some of these companies are going to be innovative enough and thoughtful enough that they make it across the curve. We've built something really valuable. And there are some companies that are at risk here. Internally, like we talk about sort of three major categories. My framework is the roadkill, the cockroaches, and the cheetahs.

34:43The roadkill is pretty self-explanatory, right? These are old school software companies. Maybe they were point solutions. Maybe they were low value that like are ultimately just going to get run over by AI. And I'm happy to kind of double click on the kinds of companies that I think that are at risk there. The cockroaches is the second category of things that can survive a nuclear blast, right? And the nuclear blast being the AI wave that is happening and still remain important. But they might not have specific AI related tailwinds. And so examples that I think about here are payroll technology, things that move money at the end of the day.

35:23You're not vibe coding your payroll system. And as much as ADP and Paylocity are legacy, I still believe that they have a role to play in the AI, sort of in an AI driven future. and that these will become tools that AI agents or software calls rather than replaces. The last category being the cheetahs are folks who have like a real opportunity to become AI native and to embed the way AI works into each one of their workflows and their offerings as companies, right? I think the best example here is Intercom, which started in 2013 as a company that, you know, pre-AI generation and did a sort of refounding moment when Owen came back on board in, I believe it was late 22, early 23, and sort of full reset the business to a point where they actually meaningfully re-accelerated the new Fin product over time to a great exit to Salesforce.

36:20What was it, a couple months ago? Some companies will make up the curve here, but there's others that I think are at a real risk. Not to be morbid, but what's an example of a category that could be in the roadkill bucket? Within the roadkill bucket, I have largely stayed away from what I'm calling kind of like the fast fashion. If you think about small business tech or a lot of personal technology, there is not a lot of complexity into what needs to be built into a solution. And those are very easily biocoded yourself. If I was a bakery or an auto dealership, it may make more sense for me to build software myself whether that's through a small agency or just by coding it myself as these tools have gotten better.

37:05Where you start to get into like real complexity and the reason why enterprise software has always been enterprise software and a mainstay is once you have to do a deal with auth and permissions and audit trails, all of the infrastructure around making something work within a large organization, like that's where it gets really tricky. And it's not to say that like people couldn't these things internally and that was never the barrier, right? People could have built their own version of Salesforce or their own version of Workday, it's become exponentially easier to do so. But it becomes like a competitive advantage thing.

37:37It's like, is being the best in the world at HR processes going to really move the needle on the business that you're trying to build? But probably not. And like that still rings true. I got news for you. The ERP category is finally getting disrupted. And if you haven't heard of it yet, please pay attention. It's the AI native ERP built specifically to replace NetSuite, and it's already won over hundreds of finance teams. Their mission is to make these zero-day close a reality, and they're actually doing it. We're talking teams closing the books at 1.35 p.m. on the first day of the month. Companies like Windsurf, Mercore, and hundreds of others run their entire finance stack on really.

38:15Revenue recognition, close management, multi-entity, native Stripe, and Salesforce integrations, Woo! Everything a scaling company needs. They've got 5.0 stars. 5.0, wow. 5.0 stars on G2. They're backed by A16Z and Sequoia. Heard of them in CPA-led implementations that get you live in 45 days. That is simply unheard of in the ERP space. If your books aren't running as fast as your business, check out Rillet. Book a demo at Rillet.com slash CJ. That is R-I-L-L-E-T dot com slash CJ. That's me. You are what your gross margin says you are. Sorry, my kids are tired of hearing me yell that at the kitchen table.

38:56Dad, I'm two. I don't know what cloud costs are. Well, I do. And cloud infrastructure is the second largest line on most tech P &Ls. And yet, most finance leaders couldn't tell you how it's priced. And as a CFO, I understand there's a non-linear impact to your company's valuation when you improve your gross margin. And the path of least resistance is through your AWS or GCP bill. Because like anything you buy, cloud gets cheaper when you buy it in bulk. In fact, you can get the rate down by 20%, 30%, even 60 % in some cases. The flip side is getting saddled with a bunch of capacity you never touch.

39:31Happened to me once, I think I still got some megawatts in the kitchen cupboard somewhere. That's exactly what cloud capital solves. Cloud capital underwrites the risk so you can realize the same savings of a deep, long-term commitment. Sands the downside. Oh, and it takes no engineering work. Bingo! I've known Edward since back when he was just starting the biz, and it's incredible how much they've saved customers on their cloud bills. That could be you too. So head over to cloudcapital.co, that's cloudcapital.co, and find out how much you could save. Every decision a company makes is inherently a financial decision.

40:07Folks, I've been trying to tell you that for 300 episodes. So it's wild that most finance teams have to spend a week on what already happened. While my favorite movie is Back to the Future, we'd be living in the past. Speaking of movies, here's a terrible plot. Tired accountant chases down a receipt from March and then polices a$20 Anthropics subscription, starring the accountant Ben Affleck. So stop living it. You don't have to play defense on what already happened if you get your system sorted. And Brex's intelligent finance platform clears the backward-looking work. Policy enforced at the point of sale, receipts matched instantly, closed short into days, not weeks.

40:44What's left is the forward-looking part. Where to deploy capital? How fast you can move? What next quarter actually needs? Because the defense is handled by Brex. It's time for finance to go play offense. Throw the ball deep. Thousands of companies, including Anthropic, Coinbase, and DoorDash, already run on Brex. Head to brex.com slash metrics to join them. That's brex.com slash metrics. My friend Brian Morrissey, he writes a newsletter and has a podcast called The Rebooting, and he explores the intersection of media and capitalism. So it's kind of tangential to tech, but he made this analogy that I can technically do a lot of the stuff that a plumber will do, but I just don't want to do it.

41:26And I think that applies well to some of this stuff where, yeah, you could technically go and build a CRM and the barriers to entry are a lot lower and you could do it faster. But is that even just the best use of your time? There's a difference between can you do something, should you do something? I 100 % agree with that. And I think you're starting to see the pendulum swing back. In December, with the release of Cloud Code, everyone went on vacation and they all lost their minds as to like all the things that you could build yourself, right? And I think we saw this across conversations in the portfolio and otherwise, where folks were like, why am I paying X amount of dollars for a piece of software that I could build, right?

42:04And I characterize like the first three months or so of the year as everyone was sort of grappling with the capabilities that existed today and figuring out what made sense for them to do or not to do. Over time, over the last, like, call it two months or so, that conversation has come back in a different way, where I think increasingly folks are realizing that some of these vibe-coded systems that they had stood up overnight are great, but we're going to require a lot more work to maintain. And it's taking away like half to an entire engineer's time. And so when you start to hit that trade-off of like, well, am I going to have that engineer working on our internal CRM or are I going to have them work on building like the next new features that are going to make my product even better?

42:52It's kind of a no-brainer. Anthropic earlier this year was hiring for a Salesforce admin and it went viral on X. It's like, well, what does that tell you? They're big users. How should we think about where value accrues if people don't spend nearly as much time in the classic UI layer? I think that there are sort of two questions that we're all grappling with, right? One is, is there value to an application layer? And I am a firm believer that, yes, there is going to continue to be value to application layer for all the reasons that we've talked about. The models are not going to continue to push up and up where these folks sit today.

43:25And two, the question then for me to figure out and for all of us investors to figure out is, OK, are there durable venture scale outcomes to be had? And for us, that question becomes around durability. What are the compounding advantages that are allowing companies to become more and more important with every single deployment? And so we have a heuristic internally, a concept that we call the continual learning. Are these systems seeing full end-to-end decision-making and able to close the loop on whether or not that was a good set of decisions that then help them improve what they do over time?

44:02And what we've largely stayed away from, there are a lot of recommendation-type systems, right, where it might connect to the context, might take some input and, you know, give you a recommendation on what to do and then ultimately have a human action. without any sort of trace of why that decision was made, how that decision was made, and what the ultimate outcome of that decision was. And I think without that full loop start to finish, there's just no way for that system to become better. And so without that sort of compounding advantage, I think you very much become at risk of being displaced by a lower cost provider or by a coded tool.

44:44Yeah, it's almost like what gets commoditized away as stuff gets cheaper to do and skills don't have to be as specific anymore and what will be durable over time. 100%. And I think one of the important things to also think about is objectivity versus subjectivity. What's the difference there? Sure. So I think where you've seen AI do the best so far from a use case perspective is places where there is an objective right or wrong, right? And I think that on one end, like coding is the best example and arguably the most successful application of LLM technology that we've seen today in terms of creating durable enterprise value.

45:20To put it really bluntly, the code either runs or it doesn't run. And so it's easily verifiable. We were building tests historically to test cases for human-generated code. Now we're just doing for AI-generated code. It was not that different, right? On the other end is subjectivity, where the decisions that you make may not have a right or wrong answer. And there is no way to ultimately measure whether or not that was the correct thing to do. It's interesting because legal has actually been probably the second hottest adopter. Which was not on my bingo card. Pre-LLM, like Eagle Tech was almost a taboo.

45:57It was CS Disco and that's about it. Yeah, it's like a legal tech, HR tech. But the way that I think about it is the way that a lawyer at Kirkland and Ellis marks up an M &A deal might look very different from the way a Gunderson lawyer does or an Orrick lawyer or like a Kravath. lawyer. And there's no way for anyone to say like, what was the right outcome? By the way, I think I saw the future today, Emily. I'm going back and forth for a sponsorship contract right now. I'm using my lawyer, Claude. I've detected from the notes that procurement is leaving that they are also using their lawyer, Claude.

46:34So we just we just have two Claude's duking it out right now. Which is the future? Yeah, I don't know how to feel about that. It should lead us to like the most optimal results at the end of the day, right? Because if your claw knows what you're optimizing for, like, hey, CJ wants to maximize, you know, dollars per hour, but he doesn't really care about data retention rights. So we went through this whole loop where I was prioritizing what I cared about. So I said, I want to make sure we get paid in advance and that there's no termination clause here because it's a media buy. And once we book that inventory, we can't rebook it for somebody else.

47:06And then it said, well, what do you think about like all these warranties around like insurance coverage and all that? I'm like, I do not care. Like that that is not top of mind for me. And so it did know. And I don't know if this is where the organization alpha comes in, but it probably knows from all the stuff I discuss about my business, what is important to me. And then I directly tell it to do not prioritize that point. I think that's 100 percent a great example of what I'm talking about when it comes to organizational alpha. It's like at the end of the day, like everything is a prioritization matrix.

47:36Everything is a tradeoff. And if we have two very smart systems, rather than like spending, I don't know, thousands of dollars in legal fees, letting the lawyer's lawyer go back and forth, it should be able to find like that clearing price and save us all the hassle. My friend, Michael Stanz, he's a pricing consultant. And he was telling me this story where there was an AI assistant working on behalf of this person to book travel. And then there was a travel company that had an AI chatbot. One of them called the other about a hotel reservation and it realized it was talking to another AI agent.

48:10And it said, hey, it may be easier if we just talk in our own language to sort this out. And it started like beep bop booping instead of just using human language because it could get to the point faster. I think there are going to be more of these weird now real world experiences occurring between agents. We didn't think about how these interactions would go down and what the efficient frontier of an outcome would be. So like for me with my contract or for this person trying to book travel of how these things get resolved. I think it's not limited to just AI to AI interactions. Two, we're actually seeing some really interesting data points around AI to human.

48:47We have a portfolio company called OneMind, which you may be familiar with, that does digital avatars for sort of high volume mid-market sales. And what they found is that buyers talking to the AI agent are actually more forthcoming with information around what they're paying today, what their pain points are, what they're optimizing for than they are with the human sales rep. Really? There are sort of these social constructs that I feel like we're all held to when we speak to another human that somehow melt away. I have a friend who runs a company that does sort of AI user interviews on behalf of product teams and investors.

49:31And they found sort of similar outcomes where people are just like much more forthcoming in some ways when they know they're talking to a bot with information than they probably would be with a human. We've talked a lot about where the value accrues for AI companies, but most people listening do not work at a frontier lab or even at a company that is post-AI. And many are not yet selling an AI product. What are ways that organizations who are not AI native can create at least some medium term alpha from leveraging AI? I think it's very basic, right? Most companies that we speak to are not even at the earliest days of figuring out how to get their data into the right systems where you can leverage AI for operational outcomes.

50:19And so that means starting to have the conversations internally. How should we be thinking about where we're centralizing all of our information? How are we reporting this information? How are we structuring the context that exists on a day-to-day basis and the decisions that are being made to unlock some of that value? Yes, there is value that is going to be built that we're investing behind in terms of net new products, companies that are being created. There's a lot of internal friction that I think deserve to go away over time. In order for an agent to do good work on your behalf, it needs to be armed with the right information.

50:57And so just getting that into one place is the first step. For our next segment, I want to ask some better questions brought to you by our friends at EY. And let me tell you, they think a lot about great questions across enterprise risk, audit, diligence. They're a smart bunch. So I'm going to ask some smarter questions right now. Me and you spend far too much time on the office of the CFO and the finance tech stack. What's exciting you? Any cool companies out there? Any areas of the tech stack that you're honing in on? I'm so fascinated by what the future of the CFO tech stack looks like because we're still just scratching the surface, right?

51:35We've got a portfolio company, Tabs, that is doing exceptionally well, tackling one of the messiest problem areas that CFOs face today, which is account for CFO. How do I take, go from all the way from contract through to getting dollars in the door? What the right structure is for that. And they've built a lot of incredible technology out of the box, helps their customers be able to do that end to end. And so I think things like that are going to be incredibly important. And that what I'm looking for on the frontier is going back to that problem that we talked to at the beginning. I think the CFO's challenge and role is going to only become increasingly important as everything moves from fixed costs through to variable costs through to fluctuating variable costs.

52:23Right. And as you can see, like the volatility in compute pricing and inference pricing in markets and commodities is just been astronomically high over the last couple of years. What I'm excited about is the next generation of companies that are serving both the technology teams and the CFO teams into helping them figure out what the right way to navigate this resource allocation problem across like when does it make sense to hire someone? When does it make sense to build an AI agent? When does it make sense to procure a piece of software or a jet-to-tool that's going to help us along that process?

52:58Emily, it's always so great to see you. I appreciate you coming on the pod. This has been really fun. Thanks so much for having me. Run the Numbers is a mostly media production. Yelling an intro by Fat Joe. Artwork by Meg D 'Alessandro. Nothing said on this podcast is intended to be business or investment advice. It's the sole opinion of me, a guy who feeds his dog way too much ice cream and has a history of net operating losses, lol. If you like this podcast, hit subscribe and give us five stars. It will take like two seconds and our algorithm overlords love it. Drink water, call your mom and have a great day.

53:29Peace.