AI at Ramp: Making Every Team Radically More Productive | Eric Glyman, CEO, Ramp

What is Ramp?

Matt Turck [1:25] Cool. All right, so the bulk of the conversation is going to be specifically about data and AI at Ramp, given the nature of this event and the podcast episode that eventually we'll create out of this.

Matt Turck [1:51] But first things first. Ramp, incredible company, incredible success for people who care about those things. $6 billion valuation. Congratulations on that. But for context, 30-second elevator pitch on Ramp.

Eric Glyman [2:22] You can think of Ramp as a command-and-control system for finance. So, from a single place, you can issue cards, make payments of all kinds—cards, ACH, that kind of thing—manage approvals, and even automate your accounting. And so, for people running finance in organizations, it's just easier to have a handle on what money is going out and get insights about it. And the upshot is, the average company using Ramp saves about 5% on their expenses. About 25,000 businesses use it, from early-stage startups to publicly traded companies like Shopify, to Boys & Girls Clubs of America as a nonprofit, to everything in between.

Matt Turck [2:37] And the company is sort of young, right? Like, you started in 2019?

Eric Glyman [2:42] Yeah, we are, as of today, 1,942 days old.

Matt Turck [2:45] Oh, yeah, that's right. I saw you on social media. I count the days.

Eric Glyman [2:45] Yes.

Matt Turck [2:49] Yes. So do you have, like, you cross the days every day?

Eric Glyman [2:51] Yeah. Just add another little line.

Matt Turck [3:02] It is actually a really interesting way to think about it. Is that a way to sort of maintain a sense of urgency? Is that how you think about it? Or is that—I'm curious why you track the days.

Eric Glyman [3:20] Yeah, originally it was a quirk, to be honest with you. So I remember it really distinctly. It was our first board meeting, and we had an outside director, Keith Rabois, who had just joined the board. And we wanted to say, hey, it's not just June 29th, but it's day 133. So, context: here's what we've done over that period of time. We mindlessly went to copy and paste and start the deck over, and it changed from 133 to 199, and it sort of just struck us.

Eric Glyman [3:47] We were like, wait a minute. It's been 66 days. Did we get the same amount done as we did the previous 66? Did we get more? Did we get less? And it set up these questions of, as far as I know, no one has more than 24 hours in a day. But looking back, there were some hours that really mattered, that ultimately drove our growth, were key developments, all that. Other hours took a lot of time, didn't have as much impact.

Eric Glyman [4:04] And so it's not so much a carpe diem or YOLO or let's get this bread kind of a thing. It's more like, as companies get larger, meetings tend to repeat. You tend to be spending time in where you are each day versus in the hours that mattered. And so, culturally, it helps us look back and say every 30 days, every week, am I spending time in the places that really matter, to give people at any level of the organization the ability to say, you know what, I actually should step out of this meeting, so I should focus on this part of our work that really mattered, or I want to spend more time on this priority than less.

How did the company start?

Eric Glyman [4:25] And it's been really useful.

Matt Turck [4:37] This is your second company, right? You did Paribus before. Maybe talk about that prior venture, and then how did you get to start Ramp? Why Ramp? And how did that all come about?

Eric Glyman [4:57] Yeah, yeah. So Paribus was about a decade ago. If it was 2024, we would have branded it as an AI agent, but it was a weird savings app. So basically, it was an app that lived in your Gmail or Yahoo. And the premise was, let's say you bought something at Amazon, Best Buy, Macy's, whatever. Let's say you bought a TV for $1,000; the next week it goes on sale for $900. Every store would guarantee that you could get the difference back if you asked.

Eric Glyman [5:27] We built an app that asked for you. It would detect receipts in your inbox, go and scrape and track the prices, ingest the policies, and if you were eligible for money back, it generated an email as you, sounded like you wrote to the store, chatted with their chatbots, and you would, as a user, wake up the next day to $100 or whatever back, and we charged a percentage. Really fun, sort of insane, weird business, but we launched it in 2015. Within a year, we had about a million customers, and originally we wanted to partner with Capital One, and they said, "We'd actually like to buy the company," and so that's how we ended up there.

Eric Glyman [5:58] In many ways, it was kind of the precursor that led to what Ramp is today. We learned a lot about turning data into savings to go trigger some action based off of it. We ended up in the credit card division at Capital One. They didn't quite know where to put us, but it was really lucky for us because we got to see what made the business profitable, how did it work. But, to short-circuit it, we found there were a lot of misalignments.

Eric Glyman [6:24] There were people thinking really hard about how do I get our customers to spend more money, earn more points, that kind of a thing. And if you were interested in the questions of what are they looking for, if you asked, was it points, cash back, something different? If you listened, they'd say, "Actually, I just want more. I want to live a better financial life." And if you really want to do that, the best way is to not spend that $100 in the first place on something you didn't need.

Eric Glyman [6:45] It's 100 times better than getting 1% back. And yet no one was doing that. And so we got very obsessed with this idea of what if there was a card and software that was designed to help you spend less? And so that's sort of how we went from Paribus to Ramp today.

Matt Turck [7:03] And then jumping into the whole data and AI part of this conversation, it actually sounds like Paribus probably—I mean, you mentioned it was an AI company before AI was the thing that everybody was—but that sounds like a sort of model recognition of receipts and that kind of stuff, right?

Eric Glyman [7:03] Yes.

Matt Turck [7:21] Okay. So you're very much coming from that DNA and jumping into Ramp. I heard you say that externally people think of you as a fintech, as a corporate card company, but internally you guys tend to think of yourselves more as an automation and workflow company. Do you want to just talk to that?

Eric Glyman [7:45] I mean, so first, some of it just starts with what is the enduring mission of the company? We believe that we exist to help companies spend less money and spend less time. And we measure it, we report on it. And if you start going a layer deeper and you ask, okay, well, where are companies spending money? Where are they spending time? Could they be spending less time? If so, it leads you down this question of productivity. And I think when I think about the original and marquee product of Ramp as being a card you can tap and your expense report is done for you.

Eric Glyman [8:17] It texts you when the receipt isn't in, you snap a photo, it pulls the receipts from your email, automates all that. It sort of stemmed from following that through. And so the problems in the fintech world from our purview was not, how do I get a credit card? People have credit cards, they have too many of them. They get called too often by credit card companies. The problem is people don't turn their receipts in on time. The problem is finance team members are spending a lot of time manually tagging transactions.

Eric Glyman [8:44] And so a lot of what we're venturing to do is really simplify the process where you can functionally inject the expense policy into how a card actually behaves. You can pull the data directly from the merchant. You can append receipts from the merchant directly, or you can read it through OCR, pull it from email, that kind of a thing. You connect through to an ERP, and you start seeing what people do. You start seeing how companies keep their books and records.

Eric Glyman [9:10] And the next time they do it, you could say, "You did it last time." And you sort of have autocomplete for your expenses, and you can create learning from how other companies are also doing this. And so underneath it all, I think really relevant to the focus of today's session in this series is sort of thinking about data as a mechanism to save people time, to save them money, to show them better ways to act. And so it's, I think, really a productivity or a data-driven company kind of in disguise of it's your friendly neighborhood credit card.

Technical aspects of Ramp infrastructure

Matt Turck [9:34] Yeah, everything that you just mentioned sounds like a lot of gnarly data plumbing, like extracting data from different systems. So what do you connect to at a more granular level, and how does that work?

Eric Glyman [9:56] Yeah, I think in terms of bidirectional, fairly deep integrations, we're talking low hundreds. Probably when you start expanding into one-way, getting low thousands. And you can think of it in a few different classes. We underwrite our customers, and so we provide credit. People apply. Some of that can be connecting to credit bureaus. Some of that could be connecting to your bank account or to your accounting software to understand how much money do you have in a certain place, how much is coming in and going out.

Eric Glyman [10:29] So some are financial-based data sources. Some of this is HRIS, so we can connect to see who reports to whom, who's someone's manager, so when someone wants to purchase something, we can automatically route it. And if you hire or let someone go, we do that. So we do HRIS, we do Okta, Google single sign-on, we do ERPs. We even do productivity through Teams and Slack and approvals. And so a lot of what we're trying to do is sort of connect where companies have their base data models, using it to power some products directly or to automate whether it's approvals, management, how things are coded ultimately at the end.

Eric Glyman [11:09] And there's interesting sub-products or kind of magical experiences you can create. Part of what allows us to automate accounting is we know who reports to whom already. Part of what leads to low losses is we can start to ingest a lot of this data about what's typical, who's the identity, what kind of things do they tend to buy. And it can create new products too, because you don't have expense management as separate. We have a lot of unfair advantages over pure-play expense management software because we're powering the card transaction.

Eric Glyman [11:39] When someone turns in a receipt, it's not a blind job where you just have an image and need to guess what it relates to and figure out a bunch of things. It's actually a matching problem. You have some raw transaction data, and so we can be radically more accurate than any of our competitors. And so a lot of what we're thinking about is: how do we be privy to a wide variety of data to power first-party products? In certain cases, when you're connected to multiple other products, you can uniquely create data or capital advantages, like lower fraud losses, somewhat better underwriting.

Eric Glyman [12:06] And are there outcomes that you can drive where, if you only had one sliver of the model, you can't really automate accounting if you only know a little bit, just what's on the card and what was the spending, not who reports to whom, all that? It's very messy. We could go very deep on all this, but that's loosely—we're connected to who are these people, what's the nature of the transactions, and what systems in and predominantly outside of Ramp might it touch.

"We can tell you if you're paying too much"

Matt Turck [12:42] Do you think—I presume the answer is yes—of Ramp as having a strong data moat, data network effects? Because it sounds like the more customers you have, the more data you have, the smarter you are about all of this. And then the more products you have, the next product makes more sense because you already have the data from the first product.

Eric Glyman [13:08] We do, but I'll give you an example, because building up to this was a core part of some of the strategies. So today, Ramp is the only place where you cannot just spend money on a vendor, but we can tell you if you're paying too much. Before you pay a vendor like Salesforce, we can actually tell you: the price on the invoice that you just uploaded, how does it relate to what other customers on Ramp are seeing? Is it higher? Is it lower?

Eric Glyman [13:35] How did we get to that? Well, when you're not only powering the transaction itself, but you're collecting the receipts, people are uploading invoices to automate the operational process, you can start to know things and create more value back. And we use it; we don't sell it. This is ultimately value that we provide to customers so they can better run their finance team. So there's things that allow you to create net-new products. There's some network effects that we have, like in our bill pay product.

Eric Glyman [14:05] Every time that a finance team pays a bill, it goes to another finance team somewhere else. You can use that to automate. As soon as one person has paid a vendor and you start seeing a lot of density and clustering of that, you can know this is a real and valid vendor. Maybe you can speed up payments if you know people on both sides of it. There's data network effects, like in the price intelligence example. There's direct network effects where it can lead to a data advantage or even just faster acquisition.

Data privacy at Ramp

Eric Glyman [14:20] And so we do try to think about when you have not just one piece, but a wider variety of data, what can you do with that? How can you automate more processes?

Matt Turck [14:33] Cueing the unavoidable question, or inevitable question, about data privacy: as you collect intelligence from your network of customers, how do you think about preserving customers' privacy?

Eric Glyman [14:56] Yeah. I mean, first, as a company that moves money and is really core to operations, I would say probably one of the core values of the company, of Ramp, is that we need to be not just incredibly reliable, but very trustworthy. There are aspects of companies' financial transactions and even the health of companies that we have access to. And so what I would say is the vast majority, and almost all things that relate to individual company performance, is basically off-limits.

Eric Glyman [15:30] There are certain types of things where sharing of aggregated data—these are on, like, a give-get model, where if you want to have access to price intelligence, you submit, and you can always say, like, I don't want to participate in seeing that—where that's really key. There's certain aspects in AI-based products where we can do autocomplete for your accounting, matching on OCR. And we have to make sure that in any cases where a third party is being leveraged, we need to make sure we have data privacy agreements.

Eric Glyman [16:05] We need to be able to audit and understand not just how we're doing things directly, but on through. And so what I'd say is very nuanced, but I think that the core of it is, we believe that data can be extremely valuable for customers, but we're not engaged in the sale of it. And given that framework, we need to be very, very thoughtful about making sure this is predominantly staying private. And if there is any kind of release to other customers, that it is aggregated and truly not de-anonymizable.

Data infrastructure tools used at Ramp

Matt Turck [16:25] Great. Can you talk about the sort of tools and platforms and data infrastructure? Are you guys a Snowflake shop? Are you a dbt shop? Like, I read something about Metaflow. What do you all use?

Eric Glyman [16:55] I would say we use, I think in some respects, probably almost all, is what I'd say the truth. So we do Snowflake heavily. AWS is a lot for a lot of our storage and core app. We use OpenAI and Azure as well for some of our more AI-oriented... There's certain aspects of our platform. We're kind of on vector databases. ClickHouse, I think, has been fantastic. Materialize and others where, again, you're probably wondering, why are we using so many different databases?

Eric Glyman [17:27] There's certain use cases. Materialize is great for detecting and preventing real-time fraud. Happy to go a lot deeper, though. And also, we overlay on all this analytics platforms like Looker or others that enable less engineering-oriented folks to go ask questions about the data, go and dive into this.

Matt Turck [17:51] All right, so that's sort of the data part. We started alluding to some of the machine learning and AI stuff, but let's go into that now. So, in the world of AI in general, before generative AI and after generative AI, what's really interesting, and a little bit to the arc of the conversation, is that you guys have been doing this pretty much from inception. And there's a number of use cases that sort of predate the big generative AI intelligence product that we'll talk about in a second.

Traditional AI use cases

Matt Turck [18:10] Do you want to talk about the spectrum of use cases that were addressed, I guess, before generative AI through regular, what is now known as traditional AI?

Eric Glyman [18:34] Yeah, for sure. Back then it was just called good old machine learning prior to all this. And so I think I'll go into a couple of things, just how loosely we think about AI at Ramp, because we do have an applied AI team, which is very horizontal. There's almost no part of how we operate that's off limits, and I'll go into the depth of it. But there's a certain category of just how Ramp operates, how we sell, how we create ads, how we understand the Gong calls that we listen to, all of it.

Eric Glyman [19:05] But there's a lot around just the pure-play operations. There's certain parts that power, whether it's underwriting, the nature of the product itself. So it has less to do with the productivity of a salesperson or marketer, but it has more to do with, can we underwrite you? Can we prevent fraud?

Matt Turck [19:05] Right.

Eric Glyman [19:24] Can we match your receipts? That's another class. And then there's two sets of productizable AI. Some are zero-touch, and others are, which is probably the most emergent case, this agentic AI, where it's not just powering behind-the-scenes experiences, but the things that you can call in order to drive an outcome. So I'll go into all of them. I would say some of the original use cases in terms of powering our product, probably the first one, just given we move money and underwrite, we take risk, underwriting was very heavy, where we're taking a large set of bank transaction data, credit bureau data, and making decisions on that.

Eric Glyman [19:52] I think it's been battle-tested and used for now multiple decades in the financial services industry, and fraud models are often used pretty heavily there. And receipt matching, even back in—

Matt Turck [20:23] So just to double-click on underwriting, it's interesting, right? Because there was this whole wave of companies that offered underwriting as a service. And the conclusion of that whole wave was a lot of, well, I mean, I guess it's slightly different for companies, but FICO score is actually pretty good. So have you found that there is juice there to squeeze in terms of just using machine learning and more datasets to do better underwriting?

Eric Glyman [20:50] There is. I think this was a big part of the New York tech story probably a decade ago, when there were a lot of fintech lenders. What I'd say is it's real. I think that a lot of those companies overstated how useful it is. And to give you a sense, for us, if a sales team is talking to a prospect, it's almost 100% of companies that we're able to approve. You might say, okay, this can go from maybe if you're just getting started and you're not using great heuristics, maybe that number is 60%, 70%, then it goes to 80%.

Eric Glyman [21:28] I think great ML-based models can allow you to go maybe from low 90s to the mid-90s, something like that. But it's not going to totally change the nature of the business. I think for us too, part of why we were so interested in the corporate space as opposed to consumers was the level of losses were so much lower. It reduced the complexity for us. In consumer underwriting, if you have a subprime-type card, you might lose 5% to 10% per year if you're really going aggressive.

Eric Glyman [22:05] 0.5% to 2% losses per year, and we're below that. I would argue, I think, below, given some of the embrace of modern techniques, but the juice, I think, just allows us to say yes to more people as opposed to prevent losses. But it's important because when you lose, and if margins are on the order of a percent, one loss can wipe out the earnings on 100 other companies.

Matt Turck [22:20] And great, I interrupted you. You were talking about OCR as the next use case.

Eric Glyman [22:20] Yeah, yeah.

Matt Turck [22:21] Receipt matching.

Eric Glyman [22:40] 100%. Well, one, part of when we got into this industry, it struck us that no one had asked the question of, isn't it just fundamentally weird that for most companies to buy one thing, you need two apps? Someone needs to say, all right, I'm going to issue you a credit card. I'm going to set a bunch of rules on one. You're going to go buy this thing. And then you're going to get a paper receipt of a digital transaction, and you're going to go log into some other app that gets your data the next day, and you're going to go and key in everything yourself, and you're going to do this for every transaction you're going to make across the entire company forever.

Eric Glyman [23:19] It's just nuts. It is totally crazy. And we said, we shouldn't integrate with expense management tools. We should just be an expense management tool, so when you buy one thing, you can go to one app. I think early on, if you wanted to do expense management, you had to verify and match receipts to the actual transaction. And so even back in 2019, as we were building this, it was originally for receipts. We use this for bill payments, procurement, all kinds of stuff today.

Matt Turck [23:24] Fraud?

Eric Glyman [24:04] Yes. Similar. I mean, a lot of that is running backtests, trying to understand in-cluster what's the nature of past transactions that a particular company makes. I think for those types of fraud prevention, it's particularly important to use a more real-time-oriented database that allows you—because often when fraud hits and people detect it, whether it's account takeovers or velocity limits, it becomes very, very important because once people figure out there's an open line, they can spend $200 on a transaction. They buy, buy, buy, buy, buy over and over.

Eric Glyman [24:23] And so if you're using databases that are slow at running analyses, the difference in stopping fraud seconds after it happens versus minutes can be millions of dollars, depending on it.

Matt Turck [24:25] Hence ClickHouse.

Eric Glyman [24:27] We do use that and Materialize as well.

Matt Turck [24:34] And Materialize, right, right, right. And how much of this is machine learning versus rules in the database?

Eric Glyman [24:50] For fraud, that is machine learning, very, very heavy. For underwriting, more and more, just given the size of data, it's possible to lead through machine learning. But a lot of that started just as roles, underwriters, all that.

GenAI use cases

Matt Turck [24:54] Okay. All right. Let's go into generative AI.

Eric Glyman [24:55] Yeah.

Matt Turck [24:57] Generative AI.

Eric Glyman [25:24] Well, I would, even before going—well, so let's do it. I actually—so one of the things that people talk a little bit less about, but people know, Ramp has been one of the fastest-growing fintech companies by both revenue and valuation, I think, in general, and also the fastest-growing New York company by revenue. And part of what's powered that has been a very early embrace of how we use AI to make members of our team more productive. And so the average SDR, sales development representative, at Ramp, we believe, books three to four times as many meetings as their next closest competitor.

Eric Glyman [26:00] Our last ad campaign was Midjourney-based. I don't think there's a function at Ramp that doesn't use AI in some part of their job. Even today, we can go through each, but a lot of what we're trying to do, even in that, is sort of look at: can you understand someone's workflow and decompose it to either build a data platform, in some cases more generative AI-based automation, to help? And so let's take, like, the sales development rep. Their job is to go book meetings.

Eric Glyman [26:18] If you follow what a lot of these people will do, even early on, it's a couple sets of activities. They're trying to figure out what are predictors that someone would be a good client. Maybe they just raised money. Maybe they just hired someone on their finance team. Who knows? Then they detect it, they build a list of it, then they go try to look up someone's email, then they go try to write some copy that sounds good, make it personalized enough, send it off, hope they get a response.

Eric Glyman [26:55] If instead of trying to go buy an AI SDR, you streamline parts of the process, you can make them radically more productive. And so you're thinking about: what data do we have about different companies? Are there certain signals that we're able to detect through systems? Can we run and measure the results of messages that we sent in order to improve kind of the outcomes? And slowly but surely, you can start to build radically more productive people doing roles like that.

Eric Glyman [27:29] Or in some cases, you can actually have fully automated aspects of that. And I think that this analogy extends to so many different aspects of how many types of companies run. And so I think sometimes, for people working on more data-oriented platforms, people tend to jump right to full automation, when really what is incredibly valuable in the short run is augmentation of people on aspects of what they do. And so that would be the operational part of generative AI. But we can talk about the product side of generative AI.

Matt Turck [27:38] Just to stay on this, because it's super interesting. So first of all, is that homegrown? Do you use vendors?

Eric Glyman [27:45] Predominantly homegrown. I mean, we've been doing this for almost as long as the company's been around, before there were vendors.

Matt Turck [27:45] Yeah, exactly.

AI/human interaction

Eric Glyman [27:47] There are vendors we're testing, but yeah.

Matt Turck [28:07] What have you learned in terms of people's ability to work with those tools? Is that very natural? I guess maybe you just hire people that can work with those tools in the first place. That whole relationship between human and machine, I find a fascinating topic. So I'm curious if there's any sort of lessons learned.

Eric Glyman [28:37] Yeah, I mean, a couple sets of things. One, often these are disparate systems and disparate teams that don't work together. Let's take the sales development rep example. It was effectively an outbound rep, and it was the growth team with growth engineering all working together, saying, "We have one goal: we need to book more meetings each month, and these goals are going to go way up." And so there was joint accountability to it. And there were different ways each of them worked.

Eric Glyman [29:10] Some were focusing more on the data platform itself, some were focused on the automation, whereas the sales representative sort of had the intuition. He knew how to get meetings, but didn't specialize. And so each tried to focus on what they were great at in their respective practice. But I think the core of it was actually having single-threaded teams with different subject matter expertise ultimately working together, aligned to it. On more of a minor note, I think sometimes funny things happen as businesses get larger, teams get more siloed.

Eric Glyman [29:44] Engineers, data scientists, folks who are incredibly smart and have deep subject matter expertise are like, "Oh no, you don't need to get involved in sales or the business stuff. That's for the business people or the salespeople." And it's like, how crazy is that? I think that one of the things that we try to do at Ramp is, like, everyone has access to financials. People can see what inputs we have, track that, and there are engineering and R&D teams that are accountable ultimately to business outcomes and drive it.

Eric Glyman [30:08] I wouldn't say it's a solve to all problems, but actually aligning teams from disparate functions along single metrics sometimes leads to— So, taking one, organizationally, how does that translate?

Matt Turck [30:21] Does that mean you have a decentralized data and AI team which is embedded in different functions? Or is that a central organization? And then if I want to build my automated SDR, I go and borrow a person. How does that work?

Eric Glyman [30:48] Yeah, so it's varied at different points in the company's history. Today we're about 850 people in terms of size. And so the data team itself reports into our CTO. And so on that side of the house, I'll give you the R&D org, has engineering, data, design, product, as well as risk and underwriting. There are some interesting things too in there, by the way. Customer support ultimately reports into product. We don't need to go too deep into that, but I think it's because if people have problems, we don't want to just solve the ticket, we want to fix the product.

Eric Glyman [31:25] And there are certain projects where actually it can be a cross-functional set of, like, data, growth, sales aspects actually working together on pods. And so on, let's say, automation in a sales capacity, you typically see teams where it's a subset of folks working against it. What I would say about data, I do think that it's important to have a centralized function. Otherwise, you start to have disparate models. I think one of the most important investments that we make and continue to invest heavily in is our customer data platform, or CDP, which is, there's engineering, but it's primarily owned by the data team.

Eric Glyman [31:57] There's analysis, there's analytics tools, there's automation tools, there's all that that's built on top of it. But having a centralized function to make sure it's standardized in some way and we're thinking about the infrastructure and how the model is organized is very important. AI, I think almost everyone uses it, but there is a dedicated applied AI team that's making sure, first, to the infrastructure point, our database structure isn't too disjointed to not be able to run analyses and pull insights from different databases.

Eric Glyman [32:32] And so ClickHouse was ultimately implemented, and a lot of things were sped up by the applied AI team, is what brought them in. But they also have an unusual remit in the Ramp context, which is that they can go into any team to sort of look at how we can apply AI to make things better. And so they saw that underwriting times for manual cases that we couldn't use machine learning to underwrite, it could take an underwriter two days to do that.

Eric Glyman [33:06] They sat side by side with the underwriter, and I think within a couple—I think, like, a two-week sprint—the average underwriting time fell to about, I think, a quarter of a day, half a day. And so often they're trying to look at where are their manual tasks, and are there places we can use data signals, we can automate more tasks, and we can give you a decisioning or propensity. This applies to now we're using it in how do we generate marketing materials.

Eric Glyman [33:29] We're using it in sales capacities to even products that we go and build directly. And there are some AI-driven products directly, like the Ramp Intelligence suite. So this is price intelligence, expense intelligence, all that, to some of the more experimental, like the agentic use cases where—

Ramp Intelligence Suite

Matt Turck [33:47] Yeah, that's fascinating. Let's go into all of that. So, Ramp Intelligence, which is the suite of GPT-powered generative AI capabilities that you all launched in 2023. So, yeah, let's go into that. What does it do? How do you think about it?

Eric Glyman [34:09] Yeah, well, first, even just going—I mean, it feels like ancient history, but back in 2023, there was this weird, funny phase where people wanted to chat with everything and just put a chatbot everywhere. And it's like, we never met a single person who's like, "I just wish I could chat with my bank account." And so it felt really off to us that that was going to be the way it was going to go. And we tried to figure out, are there certain data-oriented products where you need AI in order to get a better result?

Eric Glyman [34:41] And so in the Intelligence Suite are things like price intelligence. When you go onto your vendor page, we're cleansing the merchants. It's not just "SFDC star invoice," but that's Salesforce. We're showing you how you compare on prices to others and giving you that insight. There's expense intelligence. It's what allows finance teams to say, "Hey, this is a shrimp cocktail." You don't need to ask them, "Why did you have alcohol at this dinner? This is an Old Fashioned."

Eric Glyman [35:01] And allowing compliance managers to focus on the right things, to accounting intelligence and automation. Can we think of it as autocomplete for your accounting categories? And so there's a cluster of products around that that's sort of in the Intelligence Suite. And so usually that's trying to take data and simplify an operational process.

Matt Turck [35:06] How experimental is all of this versus you're 100% happy with it?

Eric Glyman [35:33] So what we have out in the product, like, very good. That stuff is very standard. The things that are very experimental, and we try to have a heavy portion of that, generally fall under what we'd call more agentic use cases. And in some sense, sort of circling back to Paribus, you can think about it as this highly limited agent. It would sort of assess vast amounts of data, over 100 million emails a day, and send price adjustment requests. And we restricted the surface area that it could send emails to directly to a very small amount.

Eric Glyman [36:01] But even 10 years ago, agents worked for over a million people. These were in production. And so I don't think the idea that a digital agent can do things for you is a 2024 idea. I think that's been around for a while. What we're working on now and really interested in is particularly around two sets of things. One, some of the multimodal models that are coming out. When the GPT-4o model was released, we were interested in the question of, look, about half of support requests that aren't fully automatable now fall under the question of user confusion.

Eric Glyman [36:31] I want to do this thing. I don't know how to do it. Can you show me how? And so we said, let's allow this model to see what the user sees. And so this is in our alpha group, our Ramp Labs group. People have access to it, but you can hit Ctrl+B and say, "I'd like to issue a card for $50 that can only be used at Starbucks. Go." From that point, the model can effectively see what's on your screen, and it acts as a tour guide.

Eric Glyman [37:00] It shows you how to find things on the site, and it takes action. It clicks on the different tabs, it enters the input, and it can even issue the card or go book a flight or hotel for you. Super interesting when you kind of think about where things can go. I think it's kind of unusual, actually, historically, that people force you to learn how their SaaS app or product works. The way most people work is you hire someone and you say, "Figure this thing out and help me do this thing."

Eric Glyman [37:29] And I think the promise of generalized agents is you may be able to have software do that for you. And so we're interested in testing the surface area of where that's possible. What I would say is it works often. I think right now, an agent probably works for, depending on the nature of the task, 60 to 90% of the tasks when it's controlled and inside of the Ramp experience. Obviously, when you're dealing with financial products, that's not high enough to go into general release.

Eric Glyman [38:02] But we're testing, trying to use it, and also saying, "Hey, it's experimental. It's not going to go do things without you watching," but we should be doing that. And so there's agentic-type use cases like that, but there's also fun stuff too, where we've now recorded, for training, to help salespeople improve, give feedback from managers, over 100,000 Gong calls. No one has time to listen to 100,000 Gong calls, but a large language model can. We've used it to create things like an AI podcast host that can listen to each week: who were the happiest customers, what were the most interesting moments? You apply sentiment analysis to who were the most disgruntled customers, which is a really useful analysis of actually figuring out and hearing bad news when people don't want to tell it to you, and understanding and being closer.

Eric Glyman [38:50] You can pick that up too. There's a member of our team, a digital person, not a real person, named Toby, who you can ask Toby any questions and hear kind of the voice of the customer and say, "Why do people—why does Ramp win over Concur?" And you can query all those calls. It pulls out the most relevant transcript. And if you're an SDR, you can write better copy when someone says, "I'm on Concur. Why should I switch?" You can write better marketing copy.

Eric Glyman [39:11] You can be a product manager and do research in that way. And so there's interesting—there's no job there that it's displacing, but actually new use cases that are unlocked when you sort of think about what's the nature of the data you have and how can you query it in interesting ways. And so I think we're early. Those would be, like, the more, whether it's internal or experimental, use cases, but I think it has a lot of interesting promise of where, if there are continued order-of-magnitude-level model improvements, there's actual work.

Eric Glyman [39:35] I think that products, instead of helping organize what you're doing, might just be able to do certain aspects of work for you to help you do higher-level work.

How Ramp keeps high product release and product velocity

Matt Turck [40:01] Circling back to the beginning of the conversation, one clear thing from this entire discussion we've just had is just the sheer pace of release, and before that, experimentation and willingness to try new things, especially as you grow and you keep launching products. How do you keep that high throughput of product release and product velocity?

Eric Glyman [40:22] Yeah, I mean, this could be like an hour long. It's a really interesting question. Two things. One, I think we sort of fell into it, but counting the days, I think, really helped. Having a real-time counter that's culturally important that people can look at and say, what have we done? What did we do today? What did we ship? It's been a new day. Was there something new created? I think it's just a really important mantra where it's not just some random executive can go and ask that whenever they feel like it, but anyone can ask that.

Eric Glyman [40:49] I think it's a really important thing. I think it's celebrating when you have new launches. It's putting a lot of focus on the launch process. There was an example yesterday when someone asked a question or had a feature request on Twitter a week ago, and a few days later, a member of the team ultimately built it, and they responded with—

Matt Turck [40:52] Over the Fourth of July, right? The Fourth weekend? Was it that tweet?

Eric Glyman [40:53] Yeah, exactly.

Matt Turck [40:54] It was amazing.

Eric Glyman [41:18] It was a lot of fun. And so, sharing it and celebrating, I think, is pretty important too. So there's a lot of cultural and soft things that you can do. There's also organizing principles about how you build different teams. I think one of the worst things you can do to people building products is go from, hey, it's you and four people trying to decide what a well-run product, a great experience, could look like. They go do a good job, and you say, great, now get 30 other people to agree with you.

Eric Glyman [41:50] But that's often what teams do. You make teams a lot larger, and you take these really small and nimble teams, and you saddle it with people. And so there's some trade-offs with it, but predominantly the way that we organized in terms of building product and having releases is you have lots of single-threaded, decentralized teams where Ramp is primarily known for spend management. The core spend management team is, I think, now the largest it's ever been. I think it's 13 people, 14 people, something like that.

Eric Glyman [42:20] There's lots of systems that it plugs into, builds on top of, reasons that we can do all that we do. But keeping teams small, having clear ownership, and letting people drive, and both being able to experience the joy and the upside in hiring decisions, but also when there's problems. It's not just tickets that go somewhere, you're getting the tickets. And if SLAs are broken, you can't ship new product. You need to fix the existing one. And so there's a lot of sub-organizing principles.

How did Ramp get to product-market fit?

Eric Glyman [42:37] But I think when you do that, when you keep teams small, when you talk about days, when you celebrate releases, and you make it a point to amp it up and launch new things, it becomes normal. And you ask, can we go even faster?

Matt Turck [42:42] How did you get to product-market fit initially? And how did you know you had it?

Eric Glyman [43:05] It was probably almost a year after launch until we had it. I actually think when we incorporated in March of 2019, we launched in February of 2020. And we had this idea of this card that wants you to spend less. And that was just weird. No one had ever heard something like that before. If you really got down to it, our value proposition was similar to other people: 5% cash back, they're kind of nice people, maybe I'll use it.

Eric Glyman [43:26] And we really had to fight and muscle it in order to grow. Every month of growth was a challenge. But where I think we really cracked it had to do with driving home and closing out what we talked about earlier in the conversation of, wow, you shouldn't need two apps to buy one thing.

Matt Turck [43:28] Mm-hmm.

Eric Glyman [43:51] You should see one. And that was controversial at the beginning. We lost a lot of business. A lot of controllers just said, why don't you integrate with Concur or Expensify? Every credit card on the planet does. There must be something wrong with you, and if you did, I would just use your card. But we held steady because we believed that we could only serve smaller companies at the beginning, but it's just a much better design of a system. And once we started to have that, there were companies that would say, I don't need to worry about being the bad cop policing people to get receipts in.

Eric Glyman [44:10] People turn in receipts in 30 seconds instead of 30 days. I'm just going to use it. And at that point, that's when we really started feeling pull and led up to the big 2021 year that we had.

Matt Turck [44:18] How did you sell initially? Maybe a second or a minute or two on go-to-market, and how does that work?

Eric Glyman [44:40] Yeah. Initially, there's a very funny saying in our world, or in the venture world, which goes something like, if you go and you ask someone for money, you're gonna get advice. And if you ask someone for advice, maybe, maybe, maybe you'll get money. And that was kind of similar to how we approached the initial sale. We tried to get introduced, or we knew somebody. We'd go to different companies: hey, we have this idea for a credit card that's designed to help you spend less.

Eric Glyman [45:10] We know some things about savings. We spent a few years in that. I have some ideas about ways you could spend less, but are there places in your business you feel like you're spending too much? How do you run this? And we tried metaphorically to sit on the same side of the table as people, try to solve a problem together. And some portion of people would say, actually, this is really interesting. If you build this, these are real problems I have.

Eric Glyman [45:33] If you solve it for me, I'd be interested, and I'll use it. Maybe I'd invest or whatever, but we kind of went from there, and that helped us in the early phase. People kind of agreed on what the problem space was, were interested, and in some way invested in the mission, and allowed us to start doing work. And it's evolved significantly. I mean, now we've organized it: how do you book meetings, how do you close, how do you activate, how do you sell the right product to other people? And there's lots of infrastructure around it. But at the start of it, trying to discover where there's pain, it was not to sell, but to get advice and try to problem-solve together.

Eric's perspective on building a company in NYC

Matt Turck [46:09] This is Data Driven NYC, and you're the poster child of success in New York. Any thoughts on building and scaling in New York versus, I don't know, other parts of the country or other options you may have thought about?

Eric Glyman [46:36] Yeah. I mean, look, I think I owe the city a lot. I think this is a city whose time has come in the startup ecosystem, and I think it's been a huge part of why we've been able to succeed. Look, I think over the past decade, some really fascinating things have happened. You've had truly homegrown, amazing successes built here all the way through: Datadog, MongoDB, extraordinary—you know, one has data in the name—Cockroach Labs, companies built right here.

Eric Glyman [47:05] I think part of what's hard about building a startup is they grow really quickly. It's unusual for companies to grow that quickly, and having people who know what it's like to be on that journey from every part of it is really important. I think you had a lot of that. You had the direct-to-consumer community. You had a lot of great companies built here for the first time. You had extraordinary engineering-led companies open incredible offices here: Google, AWS, Meta, Stripe, you name it, all of that.

Eric Glyman [47:33] And so I think it's a place where you have the know-how, you have a raw density of talent. And last, for builders, it's not like being in the Valley where people are mercenaries. I think average career tenure is 12, 14 months. I think it takes time to really understand a company and to do great, great work. And I think for people who are seeking to start and build a great company, I think there's a real untapped opportunity. There are sadly not that many companies yet like Ramp growing in this way, like Datadog, like earlier-stage companies that are on the come.

Eric Glyman [48:08] And so I think for people who are going out and venturing, I think you can get—maybe it's a smaller market than the Bay Area—but your ability to compete and find extraordinary talent is immensely high. I think there's a lot of people who are currently in other parts of the world who are there because that's the only place that builds these rapid-growth-type companies who actually would want to be in New York. Something about the city calls to them. And last, it's just amazing—it's the financial capital of the world.

Eric Glyman [48:28] It's one of the centers of taste. A lot of the design industry is here. Media is here. I think there's lots of advantages. People want to be in the city. And so it's been a great experience for us. But look, I'm very bullish on the city and people building here.

Matt Turck [48:31] Eric, this was fantastic. Thank you so much. Really appreciate it.

Eric Glyman [48:32] Matt, thank you. This was awesome.

Matt Turck [48:53] Hi, it's Matt Turck again. Thanks for listening to this episode of The MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already, or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build the podcast and get great guests. Thanks, and see you at the next episode.