Seth Rosenberg:
I'm Seth Rosenberg, I'm a general partner at Greylock and host of Product-Led AI, which is a new series exploring the opportunities at the application layer of AI.
Today, we’re kicking off the series with Eric Glyman, who is the CEO and co-founder of Ramp.
Ramp is a fintech company that started with a simple concept: a corporate card designed to help you save time and money. It’s evolved into one of the fastest growing fintech companies in history, and is evolving into the leading AI software platform for finance teams with products including expense management, travel, bill pay, and procurement. Since launching, Ramp has saved clients some 10 million hours of human labor and more than $1 billion in spend.
Ramp is loved by both customers and investors, and recently raised a Series D-2 round at a valuation of over $7.5 billion. We’re lucky at Greylock to be investors in Ramp.
Ramp has become a leader in AI in both how they run their business internally, and how they use AI to further their mission of saving customers time and money.
Eric, thanks a lot for joining
Eric Glyman:
Seth, I really appreciate being here and hope today's session is interesting.
Seth Rosenberg:
Awesome. So first congratulations, you marked Ramp's five year anniversary a few weeks ago.
Eric Glyman:
We were 1,845 days old, so I think five years and two and a half weeks.
Seth Rosenberg:
I feel like that's part of your culture where you actually measure every single day. Tell us about that.
Eric Glyman:
Well, as far as I know, no one gets more than 24 hours in a day. If anyone does, I'd love to know about that. But there are some hours in some days that simply matter more.
Early on in the history of the company, we started making it a habit to ask, ‘How old are we? Over the past 30 days, did we get the same amount done as the prior 30? Less? More?’ And it forced us to really think about just the value and leverage of time, both to give ourselves and the team permission to say no to things that are less important and say yes to things that might give us a shot at being able to be a little bit more productive and a little bit higher leveraged in what we do than others. And so it's been a great thing. I think that counting the days and thinking about the passage of time as part of what's helped Ramp as a company move much faster I think than most organizations.
Seth Rosenberg:
Yeah, you've accomplished an amazing amount in 1,845 days. So maybe before we get into your AI strategy, just to rewind the clock: what is Ramp? You had this interesting independent idea five years ago, which is creating a spend card that's designed to actually spend less.
Eric Glyman:
Definitely. I would start with an initiative that it’s really expanded quite a lot from. At launch day in February of 2020, you could think of Ramp as the first corporate card that was designed to help companies spend less. Now it sounds really simple, but when you looked at most credit cards in the world, they were designed with incentive programs that would get people to spend more money, earn more points, that kind of thing. But any business owner I've ever met, any finance leader that I've spent time with, is often thinking about the longer term, and marginal points don't change the trajectory of the company. But being more profitable – having more time to find product-market-fit to sell, to focus on long-term outcomes – does.
And so first we inverted the nature of this program. We created software on top of the card and expense management that would show companies ways to constantly cut their expenses.
A dollar not spent is a hundred times better than getting a penny or 1% back on that dollar in the first place. And so a lot of aligned software, but very quickly we got obsessed with this notion of time. Often, historically, if you were a bank you could move money. And if you were a software provider, well you could build add-on software. And we sort of awoke to this crazy system where for most companies the reality was to buy one thing, you needed to manage at least two to four systems. You'd issue someone a credit card in one, you'd go to your expense management app like a Concur or Expensify to ask people to add receipts, you would then link it to accounting software and then if you were really on top of it, you would ask who reports to whom, how is that changing your link in HRS? And so to buy one thing, it was awfully non-productive to run and operate your company.
And so that very quickly expanded us into what Ramp is known for today, which is really a full stop and first class finance operations platform to run your business more efficiently. We're known for powering what today is still the fastest-growing corporate card (not just in the US but in the world), and the fastest-growing accounts payables platform. We have built-in procurement, accounting, automation. We're able to automatically tag books and records. We have ways to show companies how they can spend less, but the overall premise and the overall promise of the platform is we help the average customer reduce their expenses by, today, 5% per year on card expenses, which is much more than what any rewards program can do. We help companies close their books days to weeks faster and overall allow companies the time and the space to be selling to customers, be making their product better, focusing ultimately on their mission, not on the work of doing work. What else? We support 25,000 plus businesses, whether it's early stage startups, to healthcare companies, to leaders, some of the great companies over the past decade like Shopify.
Seth Rosenberg:
I love how some of the most complex concepts and complex things to accomplish are actually solving very simple problems. The Jeff Bezos philosophy of, “In 10 years, customers are going to want things faster and they're going to want things cheaper.” And I think that's one thing I really respect about how you think about Ramp: people prefer to save a dollar than to earn two cents back on a dollar that they've spent
Eric Glyman:
Completely. I think the constraint in most organizations, no matter the scale, is focus, and time is really at the heart of it. When you’re 10 people, it's obvious maybe you only have 240 hours in a day, some of which you should be sleeping in. If there is any portion of it being put towards non-core activities of not finding new customers, it's not making your product better. It's not like listening and learning and focusing on asking where's value created. I think it's a huge distraction and cost. And I guess what I'd say is we're known for moving money. We know a couple things about that, but at heart I think we're a productivity company and I think when you make people able to do more with less to reduce the force needed to perform work and do more, I think that's where wealth is created.
Seth Rosenberg:
You're so focused on saving people time and money that AI seems like a natural accelerant to make that happen. That said, I feel like there's a lot of buzz and distraction in Silicon Valley. There's a new hot technology every couple of years. So how in your mind is AI different?
Eric Glyman:
So look, whether Ramp is the winner from this or someone listening to this podcast, I think – full stop – what we're going through now is the biggest shift to productivity, certainly in my life. And I think that all of us, especially folks who are leading and thinking through the strategies of their companies need to be thinking about what it means for their business and working backwards from there.
To come back to that for a second, it's a space that we've been thinking about for a long time. So which was the precursor business to Ramp, it was my last company that was really a proto AI agent. The product and the premise was it lived in your inbox, it was an app that you could link to your Gmail, Yahoo, Hotmail, whatever. It would then scan your inbox for receipts as you were buying things over time, check the prices on the goods and services you bought, check policies so that let's say you bought a TV for Best Buy for a thousand dollars the next week it went on sale for $900. It would effectively detect that, automatically write a note to sound like a person and email customer support of Best Buy to say, “I bought this for a thousand dollars, now it's $900, your pricing policy says that I'm owed a refund for a hundred dollars for the difference. Could you process it?” And the user would wake up the next day and they had a hundred dollars back in their account. And so it was a very narrow AI agent. He was using very simple natural language processing both to generate the message as well as programmatically detect the response, categorize it, and then our business was charged a cut on it and it worked within a year.
I think part of what's changed now in the context of 2024 is model reasoning is in most cases approaching – and in certain use cases surpassing – average human level reasoning functionally through a network call that anyone can access for fractions of a penny. And I think that has fairly profound ramifications on every business. And I think that for folks thinking through these problems, I think we owe it to ourselves to be thinking through and planning for a world where the capabilities of models are exceptionally high and increasing dramatically. I think we can be certain that it's going to dramatically change all of knowledge work.
And I think for Ramp, part of why we're taking this so seriously is, I mean you're already seeing it in software engineering. I would argue it’s probably the most digitized industry, where I would argue that finance is probably not only one of the largest areas of knowledge work and economic activity, but I believe the second-largest or second-most digitized industry on the planet. And so thinking through, very carefully, of how this will change just the work of doing finance, of underwriting, but also all the operational processes around the decision to move money actually become very interesting. And I think it becomes very relevant from the Ramp lens is at the heart of it. Yes, we issue cards, yes we move money, but what we specialize in is helping companies issue cards with the right controls to automate the expense management process to figure out where transactions should be categorized.
Historically, a lot of this work necessitated a finance professional looking, finding the ultimate patterns within this data tagging and what culminated in what most finance people would feel, which is most of their job is actually a really boring, it's a lot of operational work, a lot of tactical work, it's a lot of falling up for receipts, it's a lot of downloading Excel spreadsheets, re-uploading to do payment runs, operational process of closing books and very little is strategic work.
And I think that there's the promise to take what is often the heaviest – and in some folks' cases like the most low level and least interesting parts of the job – and augment people's ability to do that in much less time and give them the space to work on, I think what folks in finance live for, which is figuring out where you can make the investments that are going to inflect the trajectory of your business go forward a lot.
And so I actually think it's quite profound. It's going to change both the products that companies can create and also too I think there's going to be a lot of companies that do very well and others that's moats are going to erode and change very quickly.
Seth Rosenberg:
Yeah, I totally agree. I think financial services is one of the most interesting areas for AI. A year ago you and I talked with Reid Hoffman about this. So fast forwarding a year later to where we are today, I feel like there's two broad buckets of how you've deployed AI at Ramp: One is how you actually build the company internally, and two is in Ramp's products. So maybe starting with the first in terms of how you just operate efficiently internally, what are the different areas of impact that you found with AI in operating Ramp as a company?
Eric Glyman:
Yeah, I don't think there's a function at Ramp that hasn't been affected by it. So maybe I'll talk through a few. We can talk through growth, how we grow more efficiently than companies we compete against. How do we provide a higher degree of service, how do we improve the quality of our marketing? And also just how do we just broadly create leverage so that every employee at Ramp can do more.
When I think about the first use case, one of the top ways that Ramp grows is through outbound. The average sales development rep at Ramp we believe has approximately three times the productivity of our next closest competitor. They're able to book vastly more meanings. And how we've been able to do this is to understand the operational workflows. Where are people querying for data, running statistical tests, to figuring out what notes are leading at scale to a higher degree of response and relevancy. So some of this is statistical methods, but there's even generative use cases to be running first response back and forth and making the work not be [a question of] ‘Do I find someone's information trove through the data, understand signals of intent and send things out?’ to [being at a place where] here's a response, there's someone who's here, we've aggregated all the information up to this point and you can dig in and be consultative, versus doing the operational work.
And so I think the combination of thinking through and pulling across all of the large data intent signals and generating and improving that through feedback loops I think has been a very large case that has made our team vastly more productive.
Support, I think it's a very similar thing. It's an incredibly hard job because you have customers who can have extremely nuanced pieces of information and they're on the phone with you right then and they're stressed out. And I think part of the promise of AI is the ability to query what's happening in the product and databases and the customer's historical information and process that very quickly to generate not only, I think now the majority of cases automate the response to deflect. And so actually, you don't necessarily need a human interaction, but when it's done, it's not someone starting from square zero and trying to dig through a dashboard, but actually the dashboards themselves assembled to surface some of the most relevant information. So it allows our teams functionally to have superpowers and just knowing much more context about the nature of the problem a customer may be having.
Seth Rosenberg:
Yeah, I think support is super interesting. I think historically it's been viewed as this outsourced cost center, but at the end of the day it's how you're interacting with your customers. And so I think AI could bend the cost curve where it actually becomes more of a strategic part of every business where it becomes an integrated sales channel.
Eric Glyman:
I completely agree. I mean, what I would say just thematically about that: There was a moment early on in my career that just totally changed the way that I thought about customer service. My last company, Paribus, went through Y Combinator and they do this thing called group office hours (which is basically this group therapy session), and they ask three questions beyond the ‘What are you working on? How quickly did you grow this week? What's your biggest problem and what are you going to do to solve it?’ And that week we said, “Well, we grew 20%, we want to add 100,000 users by the end of the summer. And our biggest problem is we're getting too many customer support tickets, and the way I'm going to solve it is we're going to hire a customer support person.”
And Jessica Livingston, who's one of the founders of Y Combinator, promptly ripped me a new one and said, “Look, here's what's going to happen. First of all, anyone who writes to you, it is a gift for every person who has this problem. Maybe only one out of 10, one out of 100 is actually going to take the time to tell you about it and to fix it. You should take it very seriously.
Next, if you actually are spending time and you're surfacing all this up, first is a chance to build a connection with customers. But more than that, it's going to give you the information to bring that back into your product and to improve it. On the other hand, if you just hire someone to answer and respond to the ticket, you will self ticket, but you're going to be nonstop growing the number of people at it.”
And so I think that the insight first was (and which I think is just profound for people building businesses), is that you actually want to use these signals and act and spend time in the channel. You don't want to just take yourself out of the process. But two, when I think about just the broader capabilities of AI over time to not just solve the ticket but to understand where are customers having problems, what are the desires that they have and how do you improve your product to meet 'em? I actually do think it's a very core part of how companies grow, how great companies grow, and how they can better inform what they build.
The last thing that I'd emphasize too on that point, I'd even flag one of the more popular members of the Ramp team is not a real person. We have a channel called Product Ask Toby, and Toby is a digital agent functionally. I don't have the time to listen to 50,000 sales calls and interactions and back and forth, but we're recording and pulling the data.
Originally, it was for use cases of helping our team better support and learn and get feedback to better sell. But one of the use cases we've stumbled on is in the case of product marketers and sales reps and even people building products. One of the first things you want to ask is, ‘Who are our customers, and how do they think about product X? Why did people pick Ramp over customer B when they think about expense management? What are their worries and fears?’ And it turns out over a weekend project, you can actually build the capability to query all of the transcripts. And anytime you have a question about what customers think, you can dig into it. We can pull up the data specifically of who the customer was, even the clip and dive into it. And so it's actually enabled us to do better product positioning, to do better product marketing, to build better products and better find and surface up and be closer to our users.
I think sort of thinking about that attribute of the ability to query vast swaths of data to process it and to reason through whereas it's most relevant are all capabilities that years ago were not a thing. And again, that's an internal use case, but I think it sort of speaks to the broad applicability of what this technology can do.
Seth Rosenberg:
I'm curious about the go-to-market side. There are a lot of products in the market around fully autonomous AI SDRs for example. How far away are we from fully automated end-to-end outreach and where do you think are the biggest areas of leverage today versus how it's going to evolve in 12 months?
Eric Glyman:
It's a great question. So some of it's already here, right? There is a meaningful part of even how we grow is fully autonomous and it is not a future state. I'm somewhat skeptical about the ability to (off the shelf) have a fully autonomous sales agent that can just drop into any company, but I think that companies that are building and using these technologies apply on historical data workflows that's specific to them. I do think that you can automate aspects of the sales journey if you're methodical about breaking that down. Maybe it's the outreach, but maybe it's not the response back and forth, maybe it's the pre qualifications and using lead scoring and modeling to figure out should you just go into self-serve or someone. On the person and sort of more efficiently using people's time [aspect] I think is already here. But when I think about how – just the same way you can have very good salespeople who in some companies are very successful and in others they're missing a quota and it doesn't work out – I think one of the most important aspects is both understanding your own product and data that your business has about customers about itself and using that training data to improve it over time.
And so I guess the short version is I think the capabilities of these models are going to continue to improve dramatically, but I think this notion that you could just buy someone off in the same way you need to train members of your own team in order to reach high degrees of effectiveness – I think that's a framework that I would apply even to just the capabilities of these agents. It's easier and more narrow ag agentic use cases versus broad, but I think over time it will move closer to maybe lower level aspects of the sales process will be fully automated, but still some of the higher degree, the complexity deep under or standing in relationship, I still think you're going to want people very much involved all the way through.
Seth Rosenberg:
Yep. And on that note, how are you structuring your team internally to build AI products? Is it centralized? Is it embedded within product teams? How do you think about that? I know you have Yunyu and the Cohere team leading the way.
Eric Glyman:
They're fantastic. I mean even to speak about them, Yunyu was one of the early first 10 members of Ramp who understood how we built philosophy for a lot of years, left and started a company Cohere, which were one of the first companies that scaled to use large language models in customer support prior to LLMs really becoming something used in production. I think they were using GPT-2 or 3, it was prior to the breakout around 3.5, and I think just generally they were brilliant at thinking about how you think about where model capabilities are going and how do you apply it into production use cases. Yunyu and Rahul lead the applied engineering team today at Ramp, which is first order like a horizontal function. They're really endeavoring to understand what are the processes and whether it's sales and how do we more efficiently reach out to risk and underwriting, what are the processes where we can give leverage to human underwriters to software engineering, to even in our own sales development process.
There's great companies like Tome, which it's fantastic to have an off the shelf Ramp deck. It's better to think through how can this deck be personalized for the customer on the other end and how can we connect that to data pipelines? And so I would generally think that it's high leverage for most companies to be thinking about applied AI. Where in your business can you improve the throughput and capabilities of members of your team and how can you automate processes through using novel capabilities of large language models?
I would extend that to infrastructure. AI is very, very data hungry. It's better the more sets of data can reach, but if these are existing in different databases, different pipelines, you're not going to see the real benefit. And so some of it is an application layer, some of it is on an infrastructure layer and there's also more vertical in product -oriented teams.
Seth Rosenberg:
Now let's move on to kind of the second big bucket, which is how you are building products for customers using AI.
Eric Glyman:
Yeah, for sure. So I think if you go under and look broadly at our Ramp intelligence suite, these are a few of the more obvious areas where you can see it.
Some are very in the background, like you'll notice if you submit an expense on Ramp, it is auto suggesting the right accounting category. For most companies, Ramp's algorithm is more accurate and certainly faster than an accounting team's ability to predict [things like] how this expense matches towards your general ledger. There's expense intelligence, which is another vertical product where we can tell the difference and say, ‘All these receipts are fine, but I would review this receipt. It has an old fashioned, which is alcohol, which is out of your expense policy.’ So rather than having finance teams auditing all transactions, they can audit the most efficient things like price intelligence and vendor management where upon renewal you can see how you rate versus data.
Some of it you see is actually sort of this digital EA sort of digital expense admin and supporter for folks.
Next is sort of operational capabilities. It's pulling data both from within your business broadly what's happening in the market and crowdsourcing to know how you can structure future negotiations and even automate aspects of that to probably the largest use case, which is accounting. Not only is it pulling data from who reports to whom HRIS, how you categorize expenses before but is more effectively predicting that for you. Some of the other use cases are how you interact with the UI. I think a lot of the history of SaaS has been, okay, learn our interface, click this button, go here and there, and it sort of forces people to do work.
You can use things like command+K today where if you can just ask things in natural language, we can generate exports, data queries and even do work on your behalf. And so some of this work culminated in a launch where Microsoft Satya Nadella announced on stage at Microsoft Ignite last November Ramp is one of the first copilot integrations where people can request whether it's spend, ask questions about their expense policies and it allows folks to both read from and write to the product itself. And so we can issue cards, we can route approvals within your company, do all this without ever needing to log into Ramp all within your team instance.
And so it's affecting both vertical products that one can use how you interact with it. And I think some of where this is going for us is more agentic like behavior. I think when you sort of think back to Paribus one or even Ramp, one of the core premises is we exist to save your company time and money. Can you actually use Ramp in order to in specific use cases, save you time directly. And that can be automating processes today that people are doing to save you money to manage negotiations on your behalf for you or if you just want to just review it and jump in critical points. Those are all types of products to purchase, to move funds from place A to B on your behalf to purchase things, tickets, softwares, and run negotiations. Those are all things that I think are very much possible and heading into.
It’s often that there are specific product use cases that are more verticalized. There's horizontal use cases, which is probably where I'd start for most companies and I actually think it's where I'd spend the majority of time for more application layer companies, but that's broadly how we think about it, but we can go a lot deeper.
Seth Rosenberg:
Yeah, that's super interesting. I feel like it's like many waves in technology. I remember at Meta in 2012 there was a mobile team. Now there's no mobile team, it's just an obvious part of how you build every product and that seems to be the trajectory with AI where you start where you need the internal expertise horizontally, but then it evolves to being embedded into every aspect of the company.
Eric Glyman:
I totally agree, and it's just one of these things, too, where it's so…I mean one of the joys I've seen of people using interacting with large language models is you ask a question in a simple way, but then you learn to ask a better question or you go and you ask GPT-4, not just would you run this analysis, but you are a very careful and methodical analyst. Please be thorough in your approach and it increases the quality of the answer, which is very alien and bizarre for some people.
Seth Rosenberg:
Yeah, it's fascinating.
I'm curious, how do you think about the advantages that you guys have built up on the data side? What's the advantage of actually owning the card and owning the underlying spend data and how that impacts your AI strategy?
Eric Glyman:
I mean, we spent a lot of years at Paribus and we had immense data scale. We were processing over a hundred million emails today, but our only ability to affect the transaction was to send an email to write to customer support for a narrow set of transactions and say, ‘Can I get a refund?’ It was very limited. And so when we were thinking about Ramp itself, we were thinking through how do we not just have an ability to read vast amounts data, but how can we write, how can we do things on customer's behalf? And so part of the conclusion was you really needed to be deep in the workflow in the transaction layer to be not just powering the transaction itself, and that's part of how we grow and monetize as a business, but ultimately it positions you to do things like when companies say we want to cut expenses, we can see duplicative vendors and you can automatically turn off cards for specific vendors for specific styles of spend. And so it allows people to not just get insight but to operate your business.
So what I would say to ladder it up, what's different about Ramp is we're actually how companies power a lot of their own operations and are the workflow layer for companies. We're not just how you spend on a card but how you gather all data to process an expense report, how you close your books, how you determine which purchase is approved or not. And when you're deeply embedded in the workflow, not only do companies want you to pull in more data to affect more workflows (and we've seen that as we've expanded through the office of the CFO), but it also allows you to provide a higher level of value. There are things that we know about how companies do their card expenses that allow them to spend less on accounts payable, on accounts payable that allow them to close their books faster. And so some of this is both what data do you have and are you getting both the inputs of that? Can you affect the nature of the output and are you seeing the output So it allows you to better run workflows and personalized.
Next, I think distribution is fairly important for companies (which Ramp is powering tens of billions of dollars of transactions per year) and that broader scale allows us to both see friction points and to understand where money is wasted as time is wasted more efficiently and increase the value prop.
And the last – which is I think more emergent but probably important for any company thinking about in the application layer of their AI strategy – we're trying to structure around certain network effects. There's weaker network effects like data network effects. Do we understand where prices are better? I think it helps a little bit to product-based network effects like accounts payable and bill payments.
Every payment goes to another finance team, which in some sense can allow Ramp to grow more efficiently, but I think more broadly it allows more automated systems to do things like think, increase the level of value, reduce the operational work for both the payer and payee. And so I think all these have feedback loops within it. More distribution can drive more data. More data allows us to do higher level work. The more companies we serve, the more efficient, the more we can ring out costs for both parties and create efficiency. But we try to think structurally about what our products do and how they compound on top of each other in this notion.
Seth Rosenberg:
What other kind of system of record or transaction data do you feel like you want to own in the future, whether it's accounting or banking, or do you feel like spend data plus workflow is actually where you want to sit?
Eric Glyman:
It's interesting. I think there's something just powerful about where so much waste happens, it's the intersection of time and money. What AI is doing is radically increasing people's productivity and I think that what finance has sort of done to cards, API platforms, all this kind of stuff, never really thought about time and so it's this interesting thing where you can actually just make companies more operationally effective than they could have been before if they're managing these systems separately. I think that if you are in the position of helping companies run more efficiently, I think over time you actually have a more authoritative set of data about the records themselves, about how decisions are made and the unwritten rules of what is from permissions, who is allowed, who is allowed to spend what on what above some level, who permission is needed as elements around the profitability, the growth curve kind of change, what expenses lead to higher levels of ROI, what tends to be lower.
So, functionally being in the flow and not just being, for lack of a better comparison, I think companies that use Ramp have access to a steering wheel in an accelerator or a brake and they can go faster on certain expenses or not. Whereas companies that have just records but are not the operations platforms, you're in the car but you can't turn it or you need to ask someone else to turn the wheel and just the feedback loops are slower.
I think from a customer’s [perspective] none of them want to be spending their time chasing people for receipts. None of them want to be obsessing over and wondering, ‘Is this transaction categorized to the wrong place?’ I think that they want to focus on their mission and what matters and so they both have increased control over their business should they need to make changes or be informed of what's happening. But I actually think a lot of people do want an autopilot and a copilot for a lot of things that must happen, but less important day-to-day parts of their job. And I think that's not just for business owners, but it's for finance teams too.
Seth Rosenberg:
Definitely. Definitely I think people over rotate on the fear of change with automation, but I think it really has the opportunity to break us out to do more creative work and to focus less on the mundane.
With these generative models, evaluation and accuracy is still a somewhat unsolved problem. How do you think about that Ramp?
Eric Glyman:
Yeah, for sure. I mean I think that's part of why the copilot model versus the autonomous agent model is so much in vogue. I think that in the same way you would see very early Waymo's always have a person who could take control, to now, where if you get into one of those cars, people kind of have the comfort and actually can see in some cases a fully autonomous system is safer. There's this process of reviewing QA testing, when higher error bars are there at the beginning you want to be checking the work.
The earlier use cases that you see this popping up in is in the case of generations of recommending how a transaction should be mapped to a general ledger. Well, the average accuracy may be higher than any one person could do, and accountants make mistakes. That's why there's outside accounting firms, that's why there's audit firms. That's why these entire industries exist. I think that kind of behavior but digitized, you'll see more and more where the work is not, ‘Hey, it's a blank Excel spreadsheet, tag this, tag this, try to build some automations and every month go through it by hand.’ It's actually for 95% of the transactions, [a scenario where] ‘This is 99.9 degree certainty, you can check it (but at a pretty fast rate),ok, great – move on to it for 4% it's at 90% confidence and for the last remaining transactions, we're going to need to go by hand and as feedback loops come through,’ you'll see that you start to go.
So I think that's going to be the patterns and behaviors that you're going to see quite a bit. But I do think that for more narrow use cases where you have not just human supervision but ultimately when there's actually not a lot of degrees of creativity and complexity, it's just about process automation, I think that you will see full autonomy be deployed a little bit faster.
I guess it's all to say part of how we tried to think about it is how to both model and assess your own accuracy and speed of doing that, and even build your own product deployment process and QA process to be measuring that very, very actively. And once you hit certain thresholds, you offer the ability to move into more fully autonomous types of use cases. But earlier on a lot of the interfaces actually reviewing and checking is what it prioritizes. It speeds up the workflow versus fully moving there. But that's a loose framework about it. I think there's other types of products where you're generating and people are taking that to edit and design and map policies to book transactions on your behalf, but I think you're going to see that copilot model turning into agentic models as the ability to ring out errors and to be confident about the recommended approach increases, is at least our take.
Seth Rosenberg:
Yeah, it is an elegant way of ramping up as the models improve.
So let's take a second to kind of brainstorm the future. So let's say it's 1,845 days from now and maybe we're doing another podcast and what does the world look like, right? I'm a customer using Ramp and how am I operating my business? How does that change both with Ramp and more broadly as these AI agents and AI workflows start to really continue to advance at the rate they have?
Eric Glyman:
I mean, first, broadly we think about what our mission is. The simple way is we want to help companies spend less money, but it's very inspired. The name Ramp itself comes from one of the simple machines in physics is the lever in pulley, there's a screw, there's the incline plane – the ramp, which is a simple machine. And all simple machines, what they do is they reduce the force needed to perform work to allow companies to get more done with less is what we're about. And what we're working towards is a world where now maybe I think in many, many years (we can’t say all), but maybe in 1,845 days from now where the vast majority of work is actually purposeful, where finance isn't tedious or monotonous but is strategic and insightful.
And I think when you sort of zoom out over the next five years to how things are going to be different, I think it's going to be shocking if you see people doing their own expense reports by hand and yet you can go to an airport today and look over at the shoulder on the person on a plane next to you and you'll see a high number of them spending an hour on their expense reports every month. I think that's going to be gone and I think that's going to be great. People will be more relaxed.
Seth Rosenberg:
More movies on planes.
Eric Glyman:
Exactly. More movies on planes.
I think on accounting, I think it's going to be increasingly checking heavier automations. Your books are done in real time more accurately and with less effort so that your team has time to think about not just how do I tag transactions, but what transactions were valuable, where is leverage coming from in my own business to when you're buying a new set of software, you can know other companies liking it, love it, the price is fair, it's been verified. And even that process of procurement both internally, to figure out if we have the requisite approvals, can be done and routed automatically. But if there was a back and forth in negotiation, you're armed with data, parts of that process are being done for you and your business is more efficient. And in the same way we went from saving the average company about 2% per year on their expenses four years ago. I like to think that number is going to be much closer to 10 or beyond.
When you think about the hard dollar cost and the soft dollar cost for companies, we want to rewire companies to be more profitable, [those] are the things that we're thinking about. I think that when we do that, I actually think you can use human intelligence and I think people's creativity and desire and what makes us human actually (not just on tagging transactions, booking things), but I think the desire to build, the desire to create, the ability to create progress in the world versus just a record of what we did, which I think is going to be profound. And so we're really excited about it.
Seth Rosenberg:
Awesome. Well let's leave it there. Eric, thank you for building this future. Feel very lucky to be a small part of the Ramp journey and appreciate your leadership in this area.
Eric Glyman:
Seth, thank you for your support and for everyone listening, I hope this was interesting and useful.