How Companies Are Actually Spending Money on AI Now

0:00Your best bottling plant employs 3 ,300 people. How do you get 3 ,300 people working at peak efficiency? Your best store has reduced waste, water, and energy usage. How do you make every store like your best store? Your best property has every guest raving. How do you make every property like your best property? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. For enterprise organizations, managing all your food needs is a tall order. But with EasyCater, you get a single workplace food vendor with the tools and resources to make it easy.

0:40Giving teams across your organization an easy way to order from a huge variety of restaurants, all on one platform. All while consolidating your corporate food spend so you can control costs, streamline billing and payment, and simplify reporting. EasyCater, your business tool for food. To learn more, visit easycater.com slash podcast. Bloomberg Audio Studios. Podcasts. Radio. News.

1:20Hello and welcome to another episode of the Odd Lots podcast. I'm Jill Weisenthal. And I'm Tracy Alloway. Tracy, we're recording this January 8th. We ran a good piece in the newsletter this week from Skanda about AI spend as like a sort of like meaningful macro driver or getting close to where it like starts to move the dial. Yeah, that was a really good piece. And I have to say some of it is slightly worrying. But I think one of the big things that's happening now is, OK, AI has become such a big pillar of the market. Yes. Right. Like the entire S &P 500, basically. It's like an AI play. Yeah, it's an AI play.

1:59And so at some point, the hype has to be matched by reality, i.e. all that investment has to be matched by some sort of revenue, right? You have to get money out of making this investment. I think there's three ways things could go. I've been thinking about this. There's three ways this could play out. One is companies don't get a lot of productivity gains from these tools. They cut back spending. There's a bunch of other projects get delayed. Market goes down. Another possibility is there's this great productivity breakthrough. Companies are more efficient than ever. Incredible boom. That's great.

2:35And then the other possibility is that none of this matters and they build God at one of these labs. And then everything we know about economics doesn't even make any sense. So to even talk about productivity or the S &P 500 or earnings in that regime is just like, It's like a secondary concern to the way the world has changed when they achieve AGI. Joe, are you okay? No, I think those are the three. Yeah, you might as well go big with your scenario analysis. Yeah, but you're right. It could go either way. And obviously, there's a lot of talk and nervousness about a bubble in AI at the moment.

3:08So I think it would be a good idea to maybe try to get an understanding of how much companies are actually spending and benefiting from AI. That's right. And we have the perfect guest because he has a great view into what companies are spending money on, including AI, and also as the CEO of a company himself in the tech space, but not directly in the AI space, a sort of user of AI tools, and maybe could talk about what's being used, what's not being used, where productivity gains are being had. We're going to talk about all of this stuff. We're going to be speaking with Eric Gleiman. He is the founder and CEO of Ramp, which is a New York City-based company.

3:46It's a spending management platform for companies, help them deal with expenses. We might also talk a little bit about expense management platforms because I have headaches about dealing with expenses. I think a lot of people do. Yeah. I guess that's why there's a business, a new business to be made. Eric, thank you so much for coming on AdLoss. Joe, Tracy, thanks so much for having me. It's great to be here. Real quickly, why don't you describe a little bit? What's Ramp? What's your story here? You can think about Ramp as a financial operations platform. It's a single place where companies can issue cards, make payments of all kinds, and even automate both expense management, which we'll get to, and accounting.

4:24But the ethos of the company and why we exist is actually to help companies save time and money. We're the only company in our space that actually reports back to our customers how much money and time did we save you. Over the past four years, we've saved our customers over$2 billion, 20 million hours, and half of that has actually come in the past year. And so we serve 30 ,000 plus companies from small, medium to publicly traded. Got it. What was the gap in the market that you saw? Because on the one hand, as Joe just laid out, a lot of people hate expenses and they're clunky and very bureaucratic and they take forever to do.

5:00But on the other hand, this is a space that is dominated by some very powerful legacy players, right? I'm thinking about American Express, for instance, and you're basically going up against them. These are companies that are great in their own right, but I think we're built for a bit of a different era. Even the company you mentioned, the founders quite literally wore top hats, you know, and are thinking not so much about, I think, really the needs of the 2020s, where I think luxury in the 2020s is actually having an hour to yourself at the end of a long work week versus you have expense reports to do at the end.

5:34And so we thought the gap was a fewfold. First, could you actually infuse technology not to make an easier to use expense report that only took an hour instead of two hours, but actually an expense report that does itself, books that keep themselves. And so the difference was we saw an opportunity to create a card where you can tap it, make a purchase. We pull the receipt from the merchant or your email automatically, so your expense report is done for you. Your books and records are done for you. And more for business owners, we found this strange distortion where they were trying to market products like spend more money, earn more points.

6:08But every business owner I ever met actually wanted to spend less and be more profitable. So we just try to keep it simple and build a company just on those principles. I mentioned in the intro, because you have this expense management platform, you have some insight into what companies are spending money on these days on AI. What can you see? What are you able to see about AI spend within large corporations? I mean, it's real. It is dramatically increasing, but in actually interesting ways. And to give you sense of the panel data that we're looking at to get these insights, we see over$50 billion a year in spend by companies.

6:43Some of these are publicly traded. Most of these are private. And often, these tend to be on the bleeding edge. So these can be from AI research labs themselves to farms to nonprofits, mom and pop shops. And this is across both credit card data as well as bill payment data. So it's a pretty good subset of it. And what we've seen is maybe twofold. First, just in terms of raw and aggregate numbers. An average customer on ramp from the start of 2023 to the end was spending about four times the number of raw dollars on AI-based products. And so there's real budget that's starting to go to this in increased ways.

7:21And next, the products themselves are starting to actually go from experimental to operational. How can you see that? What do you see in your data that backs that up? So the best way we think to know this is that if you looked at an average AI purchase, maybe you purchased some software seat in 2022. There was a 50 % chance that within the next month, a customer that bought it would no longer be a customer. They were experimenting with it. In 2023, that had jumped to a 70 % chance. In 2024, it's continuing to go higher and will release data. Wait, sorry, 50 % chance you continued at the next. That's right.

7:59Got it, got it. And so there was a radically higher chance that you were keeping this around. And so it went from tinkering to this is starting to become a real part of engineering processes, sales tools that teams were using to be more productive, to even back office tools to manage accounting, manage expenses. And so I think we're still on the trajectory best in class of products are going to be in the 90s of percent. But the jump was dramatic in 2023 and four. How granular does your data go? Can you see people spending on, I don't know, a basic LLM subscription versus something else? Very much so.

8:37So the interesting part about what we do is because we automate the expense report process, we can see not just that a company spent on OpenAI, but specifically, was it an API call? Was it a chat GPT license? And so even among products, you're seeing itemized and SKU level data. And so you can start to get really interesting insights of even in terms of submarkets, one of the emergent themes that people are talking about now was in 2023, there was only one name in AI that mattered, and that was OpenAI. In 2024, suddenly 20 % of developer market share was going to Anthropic, which was, I think, at 3 % in data in 2023.

9:17And so you can start to get very granular of how is this even being used across which models are being called. And so it's actually this interesting level of insight that hasn't quite been seen in these markets. And then, sorry to focus so much on the data, but how do you actually classify an AI use versus something else? Because I imagine there's a lot of software, for instance, out there now that incorporates some sort of AI component, right? It feels like the Venn diagram of AI and basic tech spending is kind of starting to come together. I think you're totally right. And there's a variety of, I would say it was an easier question in 2023.

9:56There was only a few strange companies calling themselves AI. Now you see kind of AI washing of companies that are not, would love their stock to pop. But I would say that we tend to classify these based off of kind of self-identification of the companies. These tend to be large language models, labs. These tend to be companies that are pure play AI products, maybe in 11 labs, if you want to generate an AI digital voice, a cognition or devin, you can hire an AI developer and these type of tools. I think you're right, though. My sense, and if you talk to too many people in the Valley, they'll tell you there will be no company that sells technology in five years that isn't an AI company.

10:29And we'll see how the jury goes there. But it's this basis. Two things. First is a statement. I've never coded in my life. So I've made a goal for 2025 to use AI to build an app. And I actually built a really rudimentary app but it wasn't really doing what I wanted it to do. I'm not going to talk about what it is. It wasn't really doing what I wanted to do. And then I like fixed the code and I tried to re-upload it and I broke it. So I had an app for about five minutes and then it died. But I'm going to, this is like my goal. I really want to like learn to use technology. I want to talk more though about, you know, me when I use, quote, use AI, it's just me like going to chat.openai.com, just like the most rudimentary user interface.

11:10Talk to us more about what you can see on the gap between just someone subscribing to the website versus someone paying for API calls, which I imagine is sort of a deeper level of sophistication building these models into a workflow in some way. Oh, for sure. And I actually think this is the most interesting development that if this happens successfully, it unlocks what many in the Valley are talking about of historically technology was software as a service. You could sell seats to people and it would do it. And there's this growing idea of service as software, where suddenly there are workflows where AI is not just a window you chat into and get a response, but actually at every step of it becomes very, very interesting, where you have kind of end-to-end products, videos being created, books being done automatically for finance teams, even art kind of getting created.

12:02And so what I would say is the way you can see it mechanically is often the type of license. Usually these are consumer licenses if you're buying like a ChatGPT Pro subscription versus there are, let's say, enterprise or even kind of developer plans where at the end of a month, you get an invoice from one of these vendors and you see, okay, there were this many calls to this endpoint. These many tokens were ultimately used. And while the specific use we aggregate anonymize and don't report down to that level, you can start to see suddenly this is much more similar to how maybe a company would call Amazon Web Services or Microsoft Azure.

12:37And this is core compute for some service that is reliant. And the different things about these graphs is done right. And what all of these companies are betting on is that they're going to grow exponentially and that it's going to be deeply embedded. There are some distinctions. One thing that I think makes this market very different and in in some sense, more vicious than anything I've ever seen is usually there's this idea of lock-in. You host all your cloud services on one provider and you can't change from one cloud to another only for small use cases. But in AI, there's this practice of kind of multiplexing.

13:11And so what developers are often doing and sort of why Anthropik came out of nowhere, seemingly in a way that wouldn't be possible, was people would try some knowledge work or some response on multiple libraries, open source, open AI, see which one's the best. And the winning call starts getting more calls. And so things are just, the markets are moving faster.

13:45Your best restaurant location gets five-star reviews. How do you make every location like your best location? Your best paper mill has been operating at peak productivity. How do you make every mill like your best mill? Your best data center has optimized every drop of water. How do you make every data center like your best data center? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. For enterprise organizations, managing all your food needs is a tall order. But with Easy Cater, you get a single workplace food vendor with the tools and resources to make it easy, giving teams across your organization an easy way to order from a huge variety of restaurants, all on one platform, all while consolidating your corporate food spend so you can control costs, streamline billing and payment, and simplify reporting.

14:41Easy Cater, your business tool for food. To learn more, visit easycater.com slash podcast. So the specific companies are anonymized, but can you see stuff on a sectoral level? We both went to the same, but that's what I was thinking. Like, for instance, can you see which industries seem to be ahead in AI and which ones are perhaps lagging behind? We can. And I think that there's a few things that were obvious and others that have started to jump out in interesting and unexpected ways. The obvious ones of certainly the earliest adopters are technologists like technology. These are engineering offices, startups.

15:20A lot of folks doing training are trying and adopting very, very quickly and are nimble to use this. The interesting thing in the areas that have actually most surprised us maybe correlate to what you might see and experience. A lot of newsrooms are actually using kind of recording tools. So if you're having a call, notes are taken for you. A lot of sales teams are using this tools that listen to sales calls, take down the notes, suggest the next steps and start to go into this. But you also see, I would say, historically, if companies wanted to cut costs and sort of focus on efficiency, their choice was higher, lower cost labor.

15:54Often in spaces where some of this can be, if you can get the rules right, things like healthcare, which tends to be late adopters, the rates of use of increase is actually starting to be much faster because when costs and margins are very low, if you can start doing network calls to do more work when there's few two people, these are some interesting emerging industries, but it's still quite early. I would say the big question that people have out there about the valuations of these companies is still very present. The relative increase is dramatic. Forex year over year in raw dollar spent per company is very, very large.

16:29But to start to pay back some of these large capital cycles, it's going to be early. And so the bet that needs to happen from where we sit is AI can't just be a product skew that people buy. But I think to pay this back, you probably start to need everyone to use it. Actually, this was going to be my other question. So we're talking about AI spend. Can you see the other side of it in the form of, I guess, savings, either by being more efficient or perhaps cutting jobs? I would say it's very interesting questions. And I would say generally as AI has taken place, unemployment has started to come down.

17:05And so I think these are very real questions for the long term. And in fact, I actually think one of the biggest misses in 2024 for those promising AI was it was going to be the year of the AI agent was whatever. pitch. You'd have the AI CFO, the AI engineer, the full kind of jobs. And I think that what you start to see is you see slices of things happening. For example, I hope it's not anyone's job in 2024 to do just expense reports, but actually an AI can do your expense reports. It can kind of look at your invoices. It can do kind of the lowest value tasks where that's starting to become a present thing in these tools.

17:45And so I think in the short run, I actually find generally we hear from customers work is getting more interesting in some sense when AI is being adopted. I think long-term there are real questions of might it actually be able to take a workflow end-to-end. I think practically speaking, AI often has very limited context. It gets a question, it can prompt out a response, but doesn't see the rest of the knowledge work. But as it starts being everywhere, it might be possible. I need that end-to-end agentic AI for this app that I'm building because what's really annoying is it'll be, because I'll write some code in Google Colab, and then it'll be like, I'll try to push it to GitHub, and then I have to go find my token.

18:24And I don't want to do it. And then I'm like, wait, where do I find the token in GitHub so that I can put it in here? And that's the stuff that I don't want to, I mean, I guess I have to do it, but I need the AI to just go find. Does that count as coding if the AI is doing the end-to-end process? If you're literally just typing in design an app. Taking the wheel. I have some ideas. I'm trying, but it's still a little bit annoying, like all these different windows and everything. Okay, but speaking of all that, let's talk about what you're seeing in your company. I imagine every engineer that I talk to is like, yes, in 2024 or 2025, they have a window open with their AI and they have their coding window and it's improved their productivity.

19:05So like I assume that's happening in your company that a lot of code is being written either directly or indirectly or with assistance from AI. Where else besides engineering, what actually are you spending money on in terms of AI resources? So the three big places, and you nailed it. Number one is in engineering. It is one of the most digital jobs. All code is digital. And so in a strange way, that is actually arguably maybe the first industry that is closest to AI. Second is sales and growth. Ramp, beyond having a large amount of spend data, is one of the fastest growing companies or startups in history.

19:42And part of what's allowed us to do this is the average salesperson at Ramp is about four times as productive as the next closest competitor. A lot of that is a heavy use of AI to automate aspects. Specifically, let's talk specifically. What do they do? Yeah. Someone who makes a sales caller, so what are they using AI for? So one of the most important functions and the first rule for anyone going into sales is this role called a sales development rep. And the job of that person is book meetings. Find people out there who maybe are doing expense reports the old way, and let's bring them in. And if you kind of decompose what that task actually is, it's first asking, who are these businesses?

20:19What is a relevant moment? Have they just raised funds? Have they just hired someone on their finance team? Maybe they're posting an open role in some needs. There are these signals out there in the world. Then you notice, okay, you've got your lead list assembled, you have to go write them. Maybe you have a junior person just out of college going and trying to guess, what's this person's email? What's this person's phone number? How do I get a mailer in front or how do I knock on their door? Then they write kind of the message, what's going to be compelling. There's all these little steps. What makes a human seller great and what makes sales interesting is the genuine human connection, someone who can go deep, understand all the context and actually close that great sale.

20:59And yet, if you looked at the task in 2022, how most people were spending their time was things that algorithms are great at, finding people's email address, testing which copy will ultimately work better, detecting across vast swaths of data, what's the signal and the noise. And so effectively, part of what's powering this level of growth is a broad set of AI tools, which do exactly that, where AI is finding the person's email, AI is detecting these signals for intent, sending the message in the job of an entry-level salesperson now is the majority to respond to interest, to close people on the call.

21:38And so that's one example. We can go through several kind of throughout the sales cycle, but it's changed the role to be much more interesting. I'm getting flashbacks to Glenn Gary, Glenn Ross. Didn't companies used to buy lead lists as well? They still sell them. They still do. Yeah, right. So I imagine if you can basically build your own lead generator, you would save some money as well. That's exactly right. But the interesting thing is just the speed moves faster. There's suddenly signals of intent of maybe there's an IP that you can back into. This is a company has gone to your site five times today.

22:13Maybe that's a low value list. It's the C list. It's not it's not the A grade list. But with kind of a modern stack that's sifting across these signals, you can get more interesting. And so I think there's things like that. The other big thing that has just transformed the job is there's just so much more noise than there is signal out there in the world. So if you're a manager and you're trying to help a 22-year-old new in their career get better at sales, it's just too many hours of calls to listen to across your whole team to do that. The large language model has no problem listening to a thousand years of calls in a single day, more than any human can.

22:50And so suddenly, I actually think one of the more interesting stories and lessons I learned about this actually came from a 100-plus-year-old credit card company where I was first skeptical when I heard the story. The executive was explaining how they stopped checking net promoter scores. And I'm like, wow, they must have gotten that bad. But the truth was something more interesting. What's a net promoter score? So a net promoter score is this process of often you sample. So let's say there's 10 ,000 people who call in with customer support. and you ask them at the end of the call from zero to 10, how happy are you with the service?

23:22And often in these surveys, you ask a small sample and you can see an aggregate, how good are your agents? The response was much more interesting. What they started doing was applying a large language model to listen to all calls simultaneously. And they could apply sentiment analysis from the tone of a voice, how happy were customers or not. And what the algorithm started doing was routing calls from customers increasingly to people who did a great job that made customers happy. This is always the case. If you're good at your job, you get more. You get more. That's the reward. But it's a real way.

23:56Suddenly, you can actually tell from every customer how are they doing and actually get more output per unit of input. It's a real use case. I don't know how you do a few years ago. And I'm sorry. Right. You were talking about where you're deploying AI and you had engineering and sales, and we didn't get to the third one. Oh, sure. Yeah. We'll see if it works. Maybe most avant-garde would actually be kind of in growth and marketing. Of course, there's customer support and other areas that people have talked about. But I actually think there's a real case that marketing and growth is becoming a technical function.

24:29People thought that art would be one of the last use cases and suddenly you can create images, videos, interesting things. And now Now, in a world where you can go and start to do this, you can see based on intent on conversion. Can you start to combine the mathematical function of what really works with people to creating beautiful, striking, interesting images on demand? And so the jury is still most out with that one. But it's been real. I think early tests are promising. And I think that the early example of this that to me in some ways kind of inspired this was maybe twofold. One, Amazon.