Karim Atiyeh - Building Ramp - [Invest Like the Best, EP.445]
What this covers
Karim Atiyeh, co-founder and CTO of Ramp, talks with the show's host about how the company passed $1 billion in revenue in just over five years. The center of the conversation is "self-driving finance." Ramp's policy agents replace deterministic rule code with a living text expense policy. An AI agent with access to calendar, email and other tools classifies transactions around the clock and refines the policy over time. Karim presents this as AI as Product Not Tool: the code becomes an LLM plus instructions in a loop. He pairs it with a business-model argument, the Value Capture Mismatch. Card spend does not scale with company complexity, so transaction fees plateau, and he expects payroll spend to become software spend as agents grow.
The range is wide. On marketing, he describes a two-week brief cycle wrapped around ten minutes of work, billboard arithmetic about rotating creative, and ad tactics that work only until platforms change. On engineering, Bounded Breakage Architecture means money movement and risk must never break, while the customer experience is iterated on the assumption it will. He also argues, more contentiously, that standard process drags teams toward average and that the payment rail matters less than consumer rewards in US card costs.
The later turns concern people and capital. Ramp practices Spikiness Hiring, interviewing deeply on one claimed strength. Karim describes a management posture of being Divinely Discontent, including skepticism toward celebrating quarterly results. On fundraising, he covers the difficult late-2022 round and treating Investors as Employees, with small starting allocations that grow. He also touches on incumbents' legal attacks as validation.
Atiyeh argues that Ramp wins by treating business software as consumer product, automating finance workflows with AI agents that improve continuously, and tolerating fast iteration and breakage rather than over-engineering for perfection.
- Business software built for decision makers creates daily friction for end users, so Ramp designs for every user to minimize time spent in the app
- Ramp's value capture shifted from card transaction fees to software pricing because card revenue does not scale with business complexity
- LLMs and agents let non-experts reach expert-level competence quickly, so technical builders can redesign entire domains such as marketing and finance
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Paid advertising patterns work for a short period and then stop as platforms change, so marketers must track platform changes to find alpha; Paribus's silent-video banana ads worked for three to four months after Facebook made video autoplay silent.
“that ad became very effective and it worked very, very well for three or four months, and then it stopped working.”
The underlying money rail matters little to people making payments, who care about price, shipping and quality, and US card costs are driven more by consumer rewards than by Visa's take rate, while agents may later choose rails by different optimization criteria.
“a big misconception is that the reason payments are maybe more expensive than some merchants would like them to be online is because Visa takes such a big cut. That's not true at all.”
Credit card economics are attractive because the product is placed in customers' hands with no contract and revenue grows the more it is used, so improving usefulness directly increases revenue.
“There's no contract that anyone needs to sign. It's the more they use it, the more revenue you make.”
Standard processes can help below-average disciplines but tend to pull teams toward average, while the most extreme performers often work in non-standard ways, so avoiding standard process is one way Ramp stays fast.
“I think processes are a good way to move you to average in a discipline if you feel like you're below average.”
Businesses spend on the order of $100 billion or more per year on the workflows Ramp targets, and as agents built by other companies grow, payroll spend will increasingly become software spend.
“It is on the order of a hundred plus billion dollars per year.”
Changing a bought billboard costs about $1,000 while adding a new billboard costs about $100,000, so a single billboard rotating creative seven times a week for about $107,000 is more effective than two static billboards for $200,000.
“So you can essentially have a billboard in Union Square that changes seven times in a week, and that's $107,000.”
Requiring precise up-front estimates of task durations incentivizes padded estimates and slows teams, so measuring speed in advance conflicts with moving fast, and avoiding standard processes is often better for outlier performance.
“You end up rewarding and punishing people who make the right estimates. So you incentivize estimates that are longer than they should take”
Although lagging indicators such as conversion above 50 percent looked strong, Karim saw that lead generation was slowing, and because conversion and monetization have upward limits while the corporate card TAM is still under 2 percent penetrated, he took over marketing to re-accelerate leads.
“We are converting at more than 50%. ... there's an upward limit to them. And if we didn't figure out how to re-accelerate our ability to generate leads, we would be in trouble today.”
Early-stage engineering should assume things will break and optimize for recovering quickly and predictably rather than building systems that never break, a pragmatism learned at Paribus and carried into Ramp.
“the name of the game at Paribus is build it very quickly and make sure that it breaks predictably.”
Ramp's policy agents replace deterministic rule codes with a living text expense policy that guides an agent with access to calendar, email and other tools, letting it classify transactions 24/7 with more context than human reviewers and improving the policy over time.
“So it can run 24/7 as transactions come in and apply the logic of the expense policy against these transactions”
Ramp can infer many unwritten company policies by observing how customers use the product, and these inferred policies drive the next generation of agents.
“We can infer a lot of these policies through their behavior.”
For small startups with customers doubting their capabilities, fixing problems quickly and visibly builds more credibility and trust than having products that customers never notice.
“I often find it a much better way to build credibility with them and trust if things break and you fix them very quickly”
When a company does something correct or important, powerful incumbents will try to stop it multiple times, so legal attacks on a tiny startup can be scary but also validating.
“If you're doing anything that's correct or right, people are going to try to kill you multiple times.”
Because demand for Ramp equity exceeded supply at almost every round, Ramp gave new investors small starting allocations and expanded them in later rounds as relationships deepened, so most investors ended up participating in multiple rounds.
“a lot of investors get a small allocation in one round. And as we build a relationship and they get more excited about the business”
Card spend does not scale linearly with business complexity, because large companies shift to bill payments and procurement systems, so a transaction-fee model plateaus and cannot capture value from very large customers.
“the amount of money that businesses spend on card does not scale linearly with the complexity of that business.”
Technical builders are especially suited to lead in this era because they can see what AI makes possible, as customers asking for incremental features are like people asking for faster horses instead of cars.
“the technical folks are more likely to see the possibilities in terms of product than non-technical folks.”
Ramp's marketing workflow required a brief reviewed weekly by the brand team, so any asset took about two weeks even when the work required 10 to 20 minutes, which made experimentation impossible and was fixed by redesigning the system rather than the creative.
“it would at the very least take about two weeks, which is kind of crazy.”
Ramp hires for spiky talent with exceptional depth in one area rather than candidates who check many boxes, because in a competition for talent with large tech companies, spikes reveal exceptional people and checklists produce average ones.
“The more things you are looking to vet someone on, the more likely you are to get average people essentially.”
A company can gain hiring edge by looking for exceptional early-college students in hard courses at schools its founders know well, since a top grade in a difficult freshman class signals spiky talent that others overlook.
“we're looking for freshmen where other companies were trying to hire juniors.”
Ramp interviews candidates deeply on the single area they claim as their strength, learning enough beforehand to test them, which reveals both their actual ability and whether they are an accurate judge of their own talent.
“I'm generally looking for what they're telling me in their resume they're really good at, brushing up on that topic”
Better models and growing customer usage data let Ramp infer customer intent from actions such as form filling and approvals, so each customer using the product produces data that improves the product.
“This allows us to infer a lot of what their intent is from their actions”
Online retailers began employing data teams to dynamically price items to maximize revenue, creating price discrepancies that Paribus could exploit on behalf of consumers.
“online retailers were starting to employ data teams and dynamically pricing their items to try and maximize their revenue.”
The next phase of AI adoption is companies treating LLMs as part of their product rather than as a productivity tool for writing the same code faster; the code becomes the LLM plus instructions and an infinite loop.
“Your code is the LLM now. Your code is the LLM plus instructions and an infinite loop.”
Celebrating current quarterly results is misguided because they were determined by work done six months to a year earlier, and relaxing signals that problems are solved, which risks neglecting work that will shape the next six months and year.
“the current quarter was baked in a couple months ago. So I always find it a little bit weird to celebrate lagging indicators today.”
Ramp views investors much like employees, since both contribute capital or time to a company that cannot be purchased on the open market, so investors are selected for the value they bring and nurtured as long-term partners.
“We don't think of investors very differently than we think of our employees.”
At Ramp, components touching money movement and risk must be built to never break, while customer-facing experience and receipt matching are iterated rapidly on the assumption they will break, and the product and engineering teams were split along that boundary.
“anything that touched money, money movement, risk, falls in the category of you need to build it right and make sure it doesn't break”
Ramp's hardest raise came in late 2022 when valuations had fallen; the company viewed the round as price discovery that sets a checkpoint, because raising only reveals a valuation rather than changing it, and enterprise value matters more than maximizing each round's price.
“It's a price discovery mechanism.”
Online retailers offer price-drop guarantees to encourage purchases but bank on consumers not checking prices after buying, which Paribus exploited by automating price tracking and recovery.
“they're also banking on the fact that no one really checks the prices of the items they buy after they bought them.”
A business card is valuable early on because the transactions reveal which tools a company uses and how much it spends, which indicates how the business functions and where it wastes time and money.
“If you know what tools they use and how much money they're spending on them, you get a sense of how the business functions”
When a marketing channel such as podcast ads fails, the cause is usually that the team has not yet found the right partner or format, so the attitude should be that the channel works and the team must learn how to make it work.
“It's not that the thing doesn't work. It's you haven't figured out how to make it work.”
Ramp's move to charge for software did not hurt conversion or growth; it accelerated growth by motivating sales to promote the products and customers to give more demanding feedback that guided the roadmap.
“in many ways it accelerated that growth because it incentivized our sales team to mention those products”
Ramp's conversion rate exceeded 50 percent from sales conversations to customers within 30 to 60 days, but conversion and monetization have upward limits, so the real risk was slowing lead generation, with corporate card penetration below 2 percent of the market.
“We are converting at more than 50%.”
Within five years, Ramp's success would mean customers never need to log into the product, as finance becomes self-driving with decisions and payments handled automatically in a loop, analogous to the shift from manual cars to Waymo.
“my hope is that people don't have to log into Ramp at all.”
Charging for agent time or token usage may incentivize engineering teams to waste compute and be less efficient with AI, which is why Atiyeh does not like that pricing model.
“I worry a little bit about that model sometimes incentivizing the engineering team to not be as efficient as possible”
The method for building consumer-grade experiences is to examine every customer interaction and ask how the company can answer the question itself or do the step for the customer, such as retrying failed payments or embedding forms in emails instead of sending customers to a website.
“how can I figure out the answer to that question myself without asking the customer? Or if I'm telling the customer to do X, why can't I do it for them?”
Ramp built a culture of mutual accountability among peers and cross-functional teams rather than accountability only to managers, so sales cares about product, product cares about sales, and every team advises product builders with customer feedback.
“we tried to create a culture where people have mutual accountability to each other as opposed to this top-down culture”
The only interview bar Atiyeh uses is whether he would join that person's startup or start a company with them if circumstances differed, prioritizing demonstrated perseverance and alignment on values and mission.
“if this person was starting a company and I was looking to join a company, would I join them? Or would I start a company with that person”
Companies usually die because they spend too much time discussing options and not enough time acting on them, rather than because they fail to see the options, so perseverance means moving quickly into action.
“the companies that don't do that well tend to die because they spend a lot of time thinking and talking and not enough time doing”
Ramp's first investor was reached through a gaming network, and the founders approached Keith Rabois and Founders Fund more for advice and operator knowledge than for capital.
“We were interested to talk to him more for the advice than the money, which is, I guess, the best way to meet investors.”
Ramp's core growth algorithm is that a better, easier product leads customers to spend more through Ramp, which increases Ramp's revenue, and the company estimates total wasted time across finance teams as its addressable market.
“we think a lot about the value that we create for customers in terms of minimizing time waste.”
Ramp's land-and-expand motion starts with a customer using only the card, then adds receipts, procurement and AP automation as trust grows, and by two and a half to three years into the company it was building accounting automation software.
“So that Land and Expand Motion really started to become real.”
The gap between lacking subject-matter expertise and having it is smaller than ever because LLMs let motivated learners ask better questions and become expert-like quickly, so engineers can redesign domains like procurement, accounting and marketing.
“The only thing stopping you from getting that knowledge is your ability to learn and to ask the right questions of an LLM”
Ramp chose yellow as its brand color because finance apps cluster in blue and green, so differentiation for its own sake helps build recognizable brand identity over time.
“the primary reason why we chose the yellow was because it was different. That's it.”
The most important tactical lever for speed is shortening the cycle between an idea and getting it in front of customers, including time to production and test and deploy speed.
“you want to shorten the cycle as much as possible between idea and putting the thing in front of a customer.”
Traditional business software was built for a single decision maker, solving one person's problem while giving every other employee a daily paper cut, so Ramp was built to serve all users.
“You solve one person's problem, but you give everyone else at the company just a small paper cut every day.”
Karim says Ramp operates with startup-style autonomy by splitting problems across small teams with full ownership, even as the company is past a billion dollars in revenue and competes with incumbents like Amex.
“we are trying to maintain that as much as we can. And we do this by essentially breaking up all the different problems”
Previous generations of business software were built for a single decision maker, solving that person's problem while giving everyone else at the company small daily frustrations, so employees come to dislike the tools.
“You solve one person's problem, but you give everyone else at the company just a small paper cut every day.”
The moment a company survives an attempt to kill it may be a more informative investment signal than product-market fit, because doing anything correct provokes repeated attempts to kill it.
“If you're doing anything that's correct or right, people are going to try to kill you multiple times.”
Paribus started when Eric noticed that a flight price he paid differed from another buyer's price, revealing online retailers' dynamic pricing, and the founders built a consumer tool that tracks prices and recovers price drops from retailers' guarantees.
“he notices that there's a discrepancy in the way items are priced online”
Avoiding all bugs by not shipping is the trap in product development, so when pursuing great outcomes, teams should accept things breaking in parts of the business where downside is minimal and upside is high.
“Don't ship anything, don't write any code. You will have no bugs. And that's the problem with that approach.”
The marketing system is built on fast experimentation with scientific feedback loops, accountability for the person who created an idea rather than committee decisions, and tools such as an internal on-brand image generator that remove dependency bottlenecks.
“It's fast iteration cycles. It's scientific, where there's experimentation and feedback loops.”
Companies typically move through three AI phases: first using LLMs to do the same work slightly faster, then using agents that write and review much more code from English descriptions, and finally embedding LLMs as part of the product being built.
“I would describe the first very early phase as you are using those LLMs to do the same work that you were doing before, maybe a little bit faster or more efficiently.”
Paribus's first version was an Excel spreadsheet with VBA macros checking daily Amazon prices for a set of SKUs, and within the first month the frequency of price changes was already rising, showing a viable business.
“Even in the first month that we were tracking prices, we not only noticed that, yes, prices were changing, but the rate at which those changes were happening in that one month in 2014 was increasing”
Ramp does not try to maximize its valuation at each round; the goal is the ultimate enterprise value created, which comes from the sum of value delivered to customers.
“We're never really trying to maximize what the valuation is at every single round. That doesn't really matter. It's about what the ultimate enterprise value that we can create is.”
Ramp was conceived as Paribus for businesses, aimed first at helping companies save money and then time, because businesses waste time and money in ways that are roughly interchangeable.
“businesses, as we both tend to waste a lot more time than they do money, and those are generally interchangeable.”
Rising US healthcare and education spending has produced only modest outcome gains because most of the extra money goes to administrative bureaucracy rather than training doctors or teachers.
“It's being spent on a lot of bureaucracy. And yes, that doesn't move the needle.”
At the Virage startup in Boston, Atiyeh's research on video news transcripts used Markov chain classification, techniques that are ancestors of the methods underlying LLMs.
“a lot of techniques that we're using are essentially the ancestors of what LLMs rely on.”
Speech-to-text and Markov chain classification techniques used by early NLP startups are conceptual ancestors of the methods that LLMs rely on today.
“a lot of techniques that we're using are essentially the ancestors of what LLMs rely on.”
Invoice processing requires checks for fraud, receipt of ordered goods, and price matching against negotiated terms, and these steps are usually undocumented.
“you make sure that the price matches what you had negotiated.”
Amazon stopped emailing full itemized receipts and replaced them with links requiring login, illustrating how platforms increasingly make it harder for third parties to access transaction data over time.
“You now get a link that says, Thanks for shopping at Amazon. If you would like to see your receipt, please click here.”
Customers ask for incremental features, like a faster horse, and cannot articulate the transformative product they want, so companies should build the car that gets customers from A to B.
“if you had asked people before cars what they wanted, they would have said faster horses.”
When war closed Lebanon's airport while Atiyeh was at RSI, his friend Zach invited him to stay with his family in New York, and the family prepared meals, pocket money and emergency contacts, an act that became lasting family ties.
“War breaks out in Lebanon, the airport's closed.”
Solving small scary problems earns bigger scary problems, so the process of company building is a continuous escalation rather than a path to safety.
“If you solve really scary small problems, the reward is scary bigger problems, and it never stops.”
When Atiyeh broke his arm while Paribus was a two-person company with him as its only engineer, he was forced to delegate to junior engineers, learning to focus on the team's total output and explain concepts clearly.
“100% of the engineering capability of the company was just me.”
Atiyeh now sees that the reward for solving challenges is more complicated challenges, so enjoying the journey with talented people matters more than expecting a satisfying endpoint.
“the reward is just a journey, to be honest.”
Growing up in Beirut during fragile peace after the civil war, with parents stressing escape and education, gave Atiyeh a sense that things are ephemeral and an ease with external risk.
“a sense that things are very ephemeral and could disappear very quickly was there in the air.”
Atiyeh's proudest achievement is the caliber of talent Ramp attracts and the people who leave to start companies, and he sees alumni who want to approach each other as a sign Ramp hires well, rather than hiring only for loyalty.
“I don't think that's a good idea. I see it as a sign that we're hiring the right people if we keep doing that.”
Atiyeh, growing up in Beirut, became fascinated with MIT by reading science magazines and used contacts at his high school to pursue the Research Science Institute, which he attended in 2006.
“I ended up looking into that program and applying and coming a year and a half before college in 2006 to MIT for the Research Science Institute”
Paribus ran one of the largest US scraping operations, processing billions of emails and tens of millions of receipts per day, because Amazon removed itemized emailed receipts and forced users to log in to view them.
“Pretty sure that we were running one of the largest scraping operations in the US at the time”
Ramp's early pitch to design partners was modest, claiming the card was no worse than Amex and asking them to trust the team to improve it, because trust came from the founders rather than the product.
“it was, this is not worse than Amex and you might as well give it a try.”
Ramp's first real differentiation was parsing business email to match receipts to card transactions while preserving privacy, followed by cleaning merchant names, which saved time in expense submission.
“that step alone actually saves a lot of time from the workflow of submitting expenses.”
Eric and Atiyeh initially did not want to raise capital for Ramp because as second-time founders they had personal capital and credibility, though building a receivables-funded business still required capital.
“Eric and I did not really want to raise at the beginning.”
When Amazon Retail sent a security-compromise complaint, a call with AWS leadership including Andy Jassy or his team showed AWS was excited about Paribus's architecture, indicating the conflict was with Amazon Retail rather than AWS.
“our problem was with Amazon Retail. It ended up getting resolved.”
Paribus was built on retailer websites that did not want it, changed constantly, and were incentivized to make scraping harder, making it fundamentally harder than building on a stable platform.
“you are building a product on top of a very unpredictable foundation.”
Atiyeh's discontent also comes from the fact that if things were perfect he would not know what to do, while the job is never finished because there is always a way to improve each user interaction.
“if things were good and we didn't really have problems, I wouldn't know what to do with myself.”
Ramp set out to give business software the user experience of Instagram, polishing each interaction, asking only necessary questions, and pre-filling forms, with the goal of minimizing time users spend in the app.
“we wanted the user experience of an Instagram, but applied to business software”
Ramp evaluates podcast sponsorship by whether the audience matches its target customers rather than by click or view counts, because targeting the right audience is the better metric for hard-to-measure channels.
“It's more about targeting the right audience than it is a higher-level metric”
Articles on AI adoption are mixed, with some reporting no benefit and others reporting transformative impact, which Karim interprets as evidence that most companies remain in very early phases.
“But a lot of companies are still stuck in the very early phases.”
Paribus's sale to Capital One began as partnership talks about identifying recent online shoppers, but legal challenges from retailers made the team need more capital and firepower, turning talks into an acquisition.
“It's clear that we needed more firepower, whether in the form of partnerships or in the form of capital, just to withstand the storm.”
There is no reason payments cannot settle live 24/7 and cheaply, and the current inability to settle on weekends or outside business hours is an anomaly that will likely change.
“It's kind of crazy that today payments cannot settle on a weekend or outside of business hours.”
A large aerospace customer that wanted only the card and API adopted AP automation and more of Ramp over about six months once it saw value, demonstrating the land-and-expand motion.
“They can plug into the card that you've built and we can map it to our internal ERP”
Ramp aims to let founders run companies focused on their mission, such as a restaurant or a soccer team, without needing finance operations expertise, with finance running on autopilot, similar to how AWS abstracts infrastructure for developers.
“make it a lot easier to build and run companies without having to be finance operations experts”
Ramp keeps investors informed about challenges and evolution over a long period so they see the company's trajectory, rather than contacting them only when raising money.
“I think it's a lot more powerful to do that over a long period of time so that they see the evolution”
Karim's teenage exposure to MIT through the Research Science Institute, where he was assigned computer science research without being taught the subject, forced him to teach himself quickly under pressure.
“the expectation was that I would do computer science research and not learn about computer science.”
Ramp's first customers and capital came from friends, family, and former Y Combinator peers who trusted the founders' ability to improve the product, and they were design partners more than customers.
“They were partners more than they were customers, frankly. They were design partners.”
Karim is most proud of the caliber of people Ramp attracts and the alumni who go on to build companies, and he hopes Ramp will be the last job people ever need to apply for.
“I still tell people who join us that I would love for Ramp to be the last job that they ever have to apply for.”
Ramp is a finance automation platform for finance teams from startups to companies with over 100,000 employees, covering expenses, procurement, accounts payable, accounting and reporting, with the aim of automating busywork across fragmented systems that take weeks to close books.
“It's for finance teams of companies of all sizes, from startups to large hundred thousand plus employee companies”
Karim challenges anyone to find someone who enjoys using Concur, arguing that it exemplifies business software built with no care for the user experience.
“I would challenge you to find anyone who enjoys using Concur, for example.”