Orply.

Ramp Uses Financial-Product Economics to Fund a Broader Software Business

Eric GlymanGarry TanY CombinatorWednesday, November 18, 202614 min read

Ramp co-founder and co-CEO Eric Glyman argues that the economics of financial products designed to help businesses spend less gave the company more money to invest in software. He sees AI as an extension of that approach: agents and shared company knowledge could help employees act on information across teams, reducing the coordination costs that he identifies as a cause of corporate failure.

Ramp’s financial-product economics paid for a broader software business

Eric Glyman said Ramp’s founding insight was not just to build software that identifies waste. It was to use the economics of financial products to fund better software. Glyman and his co-founder Karim Atiyeh believed that if a financial product were designed to help customers spend less, rather than simply to make the provider more money, it could compete on value. The business model, he said, gave Ramp more money to invest in software than competitors focused on a single software category.

The card business offered a specific example. Companies were pushed to grow purchase volume and preserve margins, while rewards programs were a major cost. That created incentives to persuade customers that points were valuable, while also limiting their cost through measures such as devaluing rewards. In customer research, people asked whether they wanted points or cash back often answered, in effect, that they wanted more money in their bank account. Glyman argued that avoiding a forgotten $100 subscription could be more valuable than getting 1% back on a purchase.

Ramp’s products could help companies spend less while also handling expenses, bill payments, and other work. The customer value, Glyman said, was not a larger rebate, but more dollars left available for the business. He said Ramp measures how many fewer dollars and hours its customers spend, and that the median customer saves 7.5% a year on expenses.

7.5%
annual expenses saved by the median Ramp customer, according to Glyman

That model grew out of the founders’ experience with Paribus, a savings app that automatically sought refunds when an online purchase fell in price. Glyman and Atiyeh had built the company before joining Capital One, which acquired it after it reached almost a million users every day within a year of graduating from Y Combinator.

At Capital One, they saw how a large financial institution operated—and what its incentives could produce. Glyman said the company had strong processes, but those processes could also kill good ideas. He saw products shaped by organizational incentives that did not always match what would make customers better off. That experience sharpened the question that connected Paribus to Ramp: what if financial products were designed to help customers spend less?

The founders’ Paribus idea began with a problem Glyman noticed after booking a flight with friends. They had used an incognito browser window to avoid personalized pricing. The next day, when another friend wanted to join, the fare had fallen by $100. Glyman read the airline’s policy, found that customers could claim a refund within 24 hours, and called. Within an hour, the group had recovered nearly $1,000.

The opportunity was not just that prices changed. Recovering the difference required noticing the change, understanding the policy, and asking for the refund. Glyman and Atiyeh spent a year testing ways to automate those steps. Credit-card data lacked the itemized information they needed; email could provide it. Paribus connected to a user’s inbox, found receipts, checked prices, and, when a purchase qualified, sent a refund request from the user’s account. Glyman described it as an early, rudimentary agent: software that could gather information and act on someone’s behalf.

The founders tested early versions on friends, even asking for their email logins. Initially, they wrote scripts in VBA. By the time they met Y Combinator, one in 20 users could save money within minutes of signing up. They applied to YC twice; within a year of graduating, Paribus had almost a million users every day and was acquired by Capital One.

The early growth came from making a complicated proposition easy to understand. Glyman and Atiyeh carried signs around New York, approached friends, and posted on Reddit’s beermoney subreddit, where users were looking for small ways to earn extra cash. The post brought in a few hundred users overnight. They focused on one measure—weekly user growth—and simplified the pitch to: stores owe you money; Paribus will get it for you.

But the offer also sounded suspicious. Paribus only charged when it saved someone money, yet it needed access to email. The founders listened to customers explain the product to friends and noticed that people naturally included the drawback: Paribus kept 25% of the savings. They changed the referral offer so both the existing user and the referred friend paid a lower share. Referrals took off. The lesson Glyman drew was practical: listen not only to what customers say about a product, but to how they describe it to someone else.

Ramp moved from detecting waste to controlling how money moves

Paribus could identify purchases and request refunds, but its ability to intervene was limited. Glyman said the company could scan hundreds of millions of emails a day, yet only a small fraction concerned retail purchases, flights, or hotels. He believed there were billions of dollars it could help recover, but its main tool was contacting customer service and hoping for a refund.

Ramp moved closer to the movement of money. As Glyman put it, it is one thing to see that a company is wasting money; it is another to stop an unwanted purchase before the money leaves. That distinction shaped Ramp’s move from financial intelligence toward an operational layer: software not only reporting what happened, but enabling the company to act.

The first customers were startups. Glyman said Ramp grew alongside companies such as Cursor and Cognition, which were small when they began using the product. If a company had to choose between building for the past and building for the future, he said, he would choose the future.

That choice also meant trying to replace the patchwork finance systems that companies commonly use: separate tools for cards, expenses, bills, procurement, travel, and accounting. Employees have to learn multiple systems, finance teams get an incomplete view, and the fragmentation creates wasted time and money. Ramp’s first chapter, Glyman said, was to assemble those functions into a more complete system.

He described the next chapter as going beyond reducing outflows. Ramp had done less, he said, to help customers grow revenue or increase the return on money they chose to spend. Products such as accounts receivable, Glass, and Inspect were meant to make the dollars and hours a company did spend go further.

The product screens shown during the discussion give a sense of that scope. One Ramp dashboard groups transactions, cards, people, reimbursements, vendors, and accounting alongside spending insights. Other on-screen examples show a receipt being matched to a transaction, procurement approvals, accounting automation, and cash management. These are not just card features: they illustrate the wider set of finance workflows Ramp is trying to bring together.

AI adoption depends on redesigning the work around the tools

Glyman argued that many people have not absorbed the pace of AI progress, and that fewer organizations have asked what their structure should look like when this kind of intelligence is available. He compared the difficulty of adopting AI with the difficulty of asking a finance professional who has spent years using Excel on Windows to switch to a Mac. Even a familiar tool can become hard to use in a different environment; asking someone to relearn how they work is a much larger change.

At Ramp, AI use was uneven. Glyman said 1% of the company’s users were accounting for 80% of its token spend, and that some people had become dramatically more productive while others lagged. The gaps between functions created problems: engineers could move much faster than designers or product managers, and existing staffing ratios no longer fit. Capacity planning became harder, too. Teams using AI effectively often needed fewer people, while teams using it little could request more staff.

Ramp stopped hiring for a quarter, despite growing more than 100% a year, in part to make AI adoption a core competency rather than a small side project. Glyman said hiring itself takes time and makes teams less productive before new employees become productive. More importantly, he wanted people to reach the point where they understood what the tools made possible and did not want to return to their previous way of working.

That required more than telling employees to install an AI tool. In a large organization, a salesperson who downloads a coding agent may still need to open a terminal, install packages, and get permission to connect it to company systems. The result can be a powerful model trapped behind a chat box, able to exchange messages but not to work with the company’s information and tools.

Ramp built internal systems to remove some of that friction. Glyman described Inspect as a coding agent with access to the same code, data, browser, and virtual machines available to an engineer. It can investigate a customer bug report alongside recent code changes, locate the issue, make a fix, and respond to the customer. Glyman said Inspect accounts for more than 85% of Ramp’s pull requests. Separately, he said the number of pull requests per engineer was growing by roughly 40% each month; he presented that as a measure of the pace of engineering work, not as a measure of Inspect’s share.

Glyman said Ramp did not mandate one tool; employees could experiment with different products, including early versions of Claude Code, Devin, and Cursor. But he argued that organizations should assess AI spending by outcomes, not by a flat cap on dollars or tokens. Spending $1,000 to generate $10,000 in value is different from spending the same amount on work that produces none. Ramp’s approach, he said, was to measure what tools cost and connect that spending to results—for example, whether a landing page improved conversion enough to generate additional revenue.

Glass is Ramp’s attempt to provide shared company context for that work. Glyman said 1,925 of Ramp’s 1,930 employees had used it in the previous 24 hours. Its design combines access to company information with controls over who can see what: it connects to data warehouses, analytics tools, and customer-support systems, while preserving individual access permissions. The company-level view can make connections across datasets; an individual still has only the access they should have, which can vary by company.

Glass also includes shared “skills.” If someone in sales finds a more effective way to do research, that method can be made available to colleagues and become a default. Employees can select different models, including open models and systems from Anthropic or OpenAI. Glyman said he believed some work now done with frontier models could be handled by simpler ones. The system can make trade-offs between speed and cost: for example, Ramp may let a background coding agent take 10% longer in exchange for a 50% price reduction.

For Glyman, this is not only about making each employee faster. Shared skills and company context can preserve how the organization works: its standard procedures, reasoning, and decisions. A method developed by one employee can become a capability others use, rather than something the company must relearn when that person leaves. The shared view can also help leaders see where teams are using different processes.

He illustrated this with expense review. Companies have policies and audit requirements that often lead someone to review every purchase. At large organizations, that can mean highly paid employees spending time approving routine expenses—sometimes approving everything, or applying policies inconsistently. Ramp’s policy agent can approve or decline expenses, or flag an incomplete policy. Glyman said the agent operated at 99% accuracy and learned from approvals, changes, and overrides, adjusting the policy as the company changed.

The product Ramp uses internally is distinct from what it was beginning to offer customers. Glyman said that, as of the week of the interview, Ramp was onboarding customers onto Glass for their own companies. He also said customers could use Ramp in more agentic ways, including adjusting policies and conducting custom research, and that use of Ramp’s MCP was growing about 10% a week. He expected it to become a major way customers used Ramp.

A shared company brain makes transparency a design choice

Glyman agreed with Tan that departments can become information silos and executives may have reasons to preserve control over them. Startups, he said, are often open by necessity: before they have revenue, assets, or code, their central task is creating something, so they hire smart people and rely on their judgment. As companies accumulate processes and things to protect, that openness can erode.

Glyman called organizational “indigestion” a primary cause of corporate failure: coordination costs rise, information becomes siloed, and leaders avoid hard decisions. AI, he argued, could let some large companies move like smaller ones again. Ramp reported more than $2 billion in annual revenue and growth more than 25% faster than a year earlier. On-screen funding and news material during the interview also identified Ramp’s $60 billion valuation and a $1.85 billion funding round.

That kind of system makes organizational transparency a question of design, not just access to information. Glass connects company systems while retaining individual permissions, and its shared skills can make a team’s methods available beyond the person who first developed them. Glyman said the system can help leaders understand standard operating procedures and spot discrepancies between how different teams do the same work.

The expense-policy example shows what that visibility could mean in practice. Rather than treating a company as a fixed set of rules, Glyman described policies that can be adjusted as the company changes. When an agent’s decision is overridden or a policy proves incomplete, that information can inform the next decision. The goal is to make the reasoning behind the work legible, rather than leaving it buried in separate systems or dependent on particular employees.

AI could make both large companies and small specialists more viable

Garry Tan put two competing possibilities to Glyman. In one, a lab or a small number of firms might capture most future value. In the other, agents could support much more specialization, allowing many companies to serve narrower groups of customers. Glyman said he strongly favored the second possibility, while agreeing that it would have to be fought for.

His case turned on the difference between capability and what a task actually requires. Glyman said that as models move from below-average human performance to average and then beyond, the gains can be substantial for jobs that require a high level of ability. But further increases may matter less for tasks that do not need them. Accounting, he suggested, might be solvable with a model of a certain level of capability; a far more capable model would not necessarily change the work in proportion.

He also pointed to what he described as a recent break in the relationship between falling prices and rising usage of intelligence. In his account, cheaper intelligence had previously led to more consumption by an amount that more than offset the price reduction—a pattern he compared to the way lower costs can increase use of a network. Over the prior two months, he said, that pattern had stopped. He cautioned that it might be temporary, or a local maximum, and that new model families could change it again. Tan suggested that the limit might instead be society’s capacity to absorb the technology: the people ready to adopt it had already done so, while others were held back by organizational politics, habit, or resistance.

Glyman said the possibility still gave him hope for startups. If the cost of intelligence continues to fall, he argued, some problems that were previously impossible to solve could become tractable, making new kinds of businesses possible. He described the broader value of inventions as often exceeding the value captured by their creators.

For Ramp, that means not defining the company only by its original product. Glyman said people sometimes think of Ramp as a fintech because it began with cards, but he sees it as a productivity company: one that helps businesses waste fewer dollars and make their time and money go further. He agreed with Tan’s formulation that a system-of-record company may need to become an AI harness, and described Glass as a harness. Ramp was beginning to onboard customers to use it inside their own organizations.

If you believe that AI is this radical tool for productivity, how could you not try to make everyone 100x engineers?
Eric Glyman

Glyman also described a change in who could build software. In one recent quarter, non-engineers created 7,000 pull requests; about two years earlier, that number had effectively been zero. Most of those requests came from the customer experience team.

He contrasted that with the familiar path from customer complaint to product change: a bug is reported, someone writes a user story, a product manager prioritizes it, a designer works on it, and an engineer may eventually pick it up. The request can take months—or disappear into a bug database. With AI-assisted development, Glyman said, some changes that once took three months can take minutes. Employees who speak directly with customers can make a fix themselves rather than handing the problem through a chain of teams.

Tan compared the change to continuous improvement on a factory floor: the person closest to the problem can improve the process. Glyman agreed, adding that the shift raises expectations. If a change can be made today, waiting until Monday becomes harder to justify. He said Ramp had released a major product update every month for the previous five months, compared with one or two major releases a year in the past. Sales quotas had risen too, though he noted that salespeople were earning more.

Glyman said the median Ramp customer had been able to cut expenses by about 5% a year the year before, compared with 7.5% now. He contrasted that with an average U.S. company profit margin of 8%, arguing that reducing waste could matter to a customer’s ability to grow.

Look for evidence that someone builds

On hiring, Glyman offered a simple rule: if you want to find people who build things, look where people are building things. A teenager who runs a Minecraft server that others use and pay for has demonstrated an ability to make something people want. Someone who has excelled at Mathcounts, the International Mathematical Olympiad, or the International Olympiad in Informatics may have shown strong problem-solving ability.

Glyman said he was less interested in how a candidate presented themselves on a résumé than in evidence of work they had done independently. The point was not that conventional hiring methods never work, but that observable proof can be more informative than polished claims.

Asked what he would tell his younger self, Glyman returned to a different kind of evidence: being willing to admit when you do not know what to do. He said founders often avoid asking for help because people want to appear as if they have everything under control. At YC, he remembered office hours organized around three questions: How much did you grow this week? What is your biggest problem? What are you going to do about it? Being able to say “I need your help,” explain the problem, and show your reasoning, he said, can make it easier to confront obstacles rather than wait for them to resolve themselves.

The frontier, in your inbox tomorrow at 08:00.

Sign up free. Pick the industry Briefs you want. Tomorrow morning, they land. No credit card.

Sign up free