Ramp cuts customer expenses 7-12% by selling outcomes, not products

1. Incentives learned at Capital One — ▶ 2:31

   Parabus, a "prehistoric AI agent", scanned Gmail receipts, found price drops and emailed stores for refunds, charging a percentage. It launched in May 2015 and was sold to Capital One a year later. Inside the credit card division, one team grew card spend, another grew net margins, and another devalued points the first team was selling. Customers on calls said they wanted more in their bank account. Cancelling a forgotten $100 subscription leaves $99 more than a dollar back. That became the premise: invert the incentives.

2. Selling time, not money — ▶ 5:40

   An expense report means swiping a card, then re-keying a digital transaction into another vendor's system. A manager reviews it a month later, then an offshore center, and a month or two passes. Paying an invoice can take 50 steps: contract check, fraud check, duplicate check, coding. "The payments are the easy part." Banks sold money; Ramp sells time. AI that reasons can redesign the process, not just digitize it, so a card tap completes the expense and an invoice is paid on its last due day.

3. Savings from 2% to 7-12% — ▶ 8:05

   At launch, Ramp's best effort cut the average business's expenses about 2% a year. Today a typical customer cuts 7-12%, against an 8% average profit margin. Glyman says there is "no one weird trick", only many small time-stealing steps removed. Ramp tracks dollars going out, blocked out-of-policy spend and vendor negotiation, and it measures clients' calendars. Unlike social media, it wants to get users out of the app and never have them spend time there again.

4. Sell outcomes — ▶ 11:16

   Ramp started on narrow problems like matching receipt images and autocompleting accounting codes such as 60,635. It was an early GPT-3 customer, when the model couldn't do arithmetic. Glyman thinks companies get stuck selling a card, a car or a computer when "at best companies should be selling outcomes". He says Ramp should be able to tell a customer their business will be better, and if another tool is better than Ramp's, "fantastic". Competing means understanding who the customer is.

5. Why Ramp beat Brex — ▶ 15:31

   Glyman credits Brex with good ideas but sees it as a continuation of how financial institutions were built: sign up for a card and get 4x, 3x or 2x points by category. That was best for the business, not the customer. Ramp wanted the same things its customers wanted and accepted making less money in a given month so they could build stronger balance sheets. Even so, it handles only about 3% of US corporate card transactions, with 97% to go, across 85,000 businesses.

6. AI spend is 2% of wages — ▶ 20:39

   A year ago the world spent about $30 billion a year on inference across the major labs. Estimates now run $150-170 billion, about 2% of the roughly $8 trillion in US white-collar wages, and up 5x in a year. Glyman wouldn't find 10% strange for a knowledge economy. 70-80% of spend is agentic: models spinning up agents that hire other agents. Unlike staff or contracted vendors, it grows exponentially, and most companies can't see or manage it.

7. Router pays the penny — ▶ 22:32

   Labs race to build the smartest AI at almost any cost. They are compute constrained, so if you ask the $1 model, they won't route you to the penny model that could answer your question. Router does that routing. It takes under five minutes and a few lines of code, and the average business that turns it on cuts inference cost 40%. Glyman says Ramp is trying to do for AI spend what it did for spend generally, and Veral leads the product.

8. Hiring and small teams — ▶ 24:19

   Ramp wanted every hire to be better than the last, starting in 2019 with people who had shown unusual early achievement. One early hire was a competitive-math MIT graduate who created Shortcuts at 14. Teams are split once they reach 8-12 people, each with clear ownership and an objective function. Three leaders, including Glyman and co-founder Kareem, review every profile before an offer. Glyman meets every new hire in week one, so raising a real problem with the CEO later doesn't feel strange.

9. Metrics inputs, not outputs — ▶ 28:47

   Most R&D works on unreleased products, with teams left alone for about a year to find product-market fit. After that comes one growth metric, such as 10% a week or 20% a month, paired with counter-metrics. If satisfaction drops or things slow down, no new building until it's fixed. Managers focus on inputs: shipments, customer conversations. Glyman's analogy: you don't win a Super Bowl by announcing it at practice, you lift weights today and run sprints tomorrow. Small teams need good telemetry.

10. Ramp as economic map — ▶ 33:13

    Ramp sees about 1% of all corporate payments, enough for statistical significance on who pays for OpenAI and Anthropic, how many licenses, and how deep usage goes. Census data showed about 5% of businesses paying for AI early last year and about 20% now, and Glyman says actual adoption is higher. The data is not sold. It is published to everyone at once on ramp.com/data to "arm the rebels a little bit" so firms can benchmark prices.

11. The trillion-dollar back office — ▶ 39:51

    Thousands of cities, states and agencies use Ramp. Glyman says waste and fraud start from lack of visibility: separate systems for cards, expenses, audit, reconciliation and payment. Defense spends roughly $10-12 billion a year on cards alone and has failed five or six audits, which no private business could do. He sees about $1 trillion of work beneath people's potential, with finance-trained staff reduced to receipt bean counters. He says AI should repurpose them to field work and capital allocation.

A company earns trust by wanting what its customer wants. Eric Glyman builds Ramp on that rule, and he treats it as strategy, not ethics. "At best companies should be selling outcomes." The outcome Ramp sells is less spend. He says Ramp should be happy to make less money in a given month if customers end up with healthier balance sheets, because Ramp grows with them.

The rule came from watching the opposite. Glyman's first company, Parabus, was bought by Capital One a year after its May 2015 launch, and he spent time inside the credit card division. One team worked to grow card spending. Another worked to grow net margins. A third devalued the points the first team had been selling. Customers on the phone said they simply wanted more money in their bank account. Cancelling a forgotten $100 subscription leaves a person $99 better off than a dollar of points does. An institution built the old way could not say that out loud. Banks with a high-yield savings account often don't mention it. What an organization says and what its people work toward drift apart once incentives drift apart.

The evidence that inversion pays is the savings number. At launch, Ramp's best effort cut the average business's expenses about 2% a year. Today a typical customer cuts 7-12%. The average American business earns an 8% profit margin, so that is a large share of profit. Glyman insists there is no single trick. The gain comes from deleting steps. An expense means swiping a card, re-keying a transaction that was already digital, waiting a month for a manager, then waiting for another review. An invoice can take 50 steps. Banks sold money and ignored those steps. Ramp sells time, and AI that can reason lets it redesign the process instead of digitizing it. A tap completes the expense. An invoice is paid on the last day it is owed. Ramp even tracks how long clients spend in the app, and it wants that number to fall. A social media company wants the opposite.

The same logic now applies to AI spend, which is where the incentives are most visibly crooked. Inference spend went from about $30 billion a year to $150-170 billion in twelve months. That is about 2% of the roughly $8 trillion Americans earn in white-collar wages, and it rose 5x in a year. Glyman sees nothing absurd about it climbing further. A knowledge economy spending 10% on thinking might even be low. The trouble is that this spend breaks the tools finance teams have. Companies can model what an employee costs, and they can enforce a vendor contract. Agentic workloads, which are already 70-80% of spend, have models spinning up agents that hire other agents. Costs grow exponentially, and few companies can see them.

The labs will not fix this. Glyman is careful to say their people have good intentions. But they are compute constrained and racing to build the smartest model at almost any cost. When one model could answer a question for a penny and another for a dollar, nobody at the lab is paid to route you to the penny. Router does the routing. It takes under five minutes and a few lines of code, and the average business that turns it on cuts inference cost 40%. It is the Ramp thesis applied to a new line item. It also fits Ramp's customers, most of whom have no software engineer. They don't want to be sold AI. They want the expense to file itself.

Running this at 2,000 people takes the same discipline turned inward. Teams are broken up at 8-12 people. Each owns a problem and an objective function, and each can live or die by its decisions. Three leaders, including Glyman and co-founder Kareem, review every candidate before an offer goes out. Glyman meets every hire in their first week, so the first real problem doesn't feel too big to bring to him. Most of R&D is working on products that have not shipped, and those teams get a year and some goals. After that comes one growth metric, such as 10% a week or 20% a month, with counter-metrics that freeze new building if satisfaction slips. Managers are measured on inputs. You don't win a Super Bowl by announcing it at practice. You lift weights today and run sprints tomorrow. All of it depends on good telemetry, because the leader is not in the room.

Ramp's position also gives it a view few others have. It sees about 1% of corporate payments, enough to show who is actually paying for OpenAI and Anthropic. Census surveys put paid AI adoption at about 5% early last year and about 20% now, and Glyman says the real figure is higher. He says the data is never sold. It is published to everyone at the same moment. The aim is to "arm the rebels a little bit" so small firms can see what a peer paid for the same software.

The obvious objection lands on the biggest claim. Glyman talks about a "trillion dollar back office" of work beneath people's potential. He means finance-trained people matching receipts to transactions. He points to the Department of Defense, which spends roughly $10-12 billion a year on cards and has failed five or six audits. His answer is that machines take the clerical work and people move to field work and capital allocation. That is a hope, not a mechanism. The same conversation says AI spend could multiply again and that, if the labs succeed, intelligence will be free. Free intelligence does not obviously need a human reviewing receipts, or even a human who is repurposed. He never says how many of those roles survive or what replaces them.

The quieter question is about the product's own incentives. Ramp says it will take less in a month so customers do better. But the same customers' savings are now 7-12% a year, growth is above 100%, and the company has 97% of corporate cards left to win. Whether outcome-selling stays cheap at that scale is not something he addresses.