Eric Glyman

The through-line

Almost everything Glyman says traces back to one idea: build a business whose economics improve as the customer's spending and wasted time go down. Ramp launched in 2020 with what he calls a "really weird and different idea," that it should be "a card that is designed to help you spend less" [4]. That came out of watching the credit card division at Capital One from the inside, where "a lot of people thinking about what was good for the business which was to get people to spend more money on their cards," while customers, when asked, said they just wanted more money in the bank account [4]. His conclusion was arithmetic: helping someone not spend $100 on something they don't need "is 100 times better than getting 1% back" [4]. The mission he states is deliberately not product-shaped: "we're not in the business of slinging cards... our mission is to help companies spend less money and spend less time," and he argues that asking that question rather than a product question "will lead you down a different product set" [4].

The second, quieter through-line is temperament. He is consistently dismissive of external markers of success and consistently attached to daily inputs, whether the subject is startups, athletes or his own psychology. Valuations and press are "lagging indicators" [1]; the only real question is "the only day that I live in is today and what am I going to do with it" [1]. The same instinct shows up in how he measures his own job at scale: "I do measure success by how much time we spend doing the deep thinking" [2].

On aligning incentives with customers

His diagnosis of why so few fintechs have displaced incumbents is that financial services companies "are just notoriously misaligned with their their customers" [4]. He describes the classic bank pattern of a high-yield savings account and a zero-yield checking account, engineered so money is hard to move, and agrees it is "very bad behavior" that "every bank does," including some of the new ones, and that it has persisted because the only competitors have been other banks [4]. He is blunt about the structural reason: "unless you were a bank you couldn't move money," so the same players have owned the industry for fifty years and "if you have 20% of a market 40% of the market what incentive do you have to actually help people spend less" [4].

Ramp's answer is a business model that "gets better the more money and time we save you" [4]. Someone once told him Ramp might be "a productivity company masquerading as a fintech," a description he found charming and close to right [4]. He extends the point to pricing: rather than selling seats, "we think we should get paid based on the outcomes we deliver to companies," and Ramp measures how much money and time it has saved each customer, from roughly 2% of expenses at the February 2020 launch to about 5% today counting both hard-dollar savings and labor no longer needed [5][3]. He expects the seat model to age badly in a world where "you actually don't need people to do a lot of jobs," so "you don't want to sell people more seats you want to drive more outcomes" [5].

On the company equation and why purchase volume is the only variable

Ramp's first board deck, written before there was any revenue, reasoned through the business model from first principles [5]. The conclusion was that card economics have a handful of inputs, purchase volume, take rate and cost of funding, and that "if you studied every single variable there's only one thing that mattered which is purchase volume" [5]. More volume means more revenue, more retained interchange, a lower cost of funding at the bank, and better fraud and loss prevention [5]. From that fell out a simple operating instruction: "if you just make spending with us just simply better than spending anywhere else... rationally people should want to spend more with you" [5]. Product teams are therefore pointed at one internal question, how much money and time are we saving customers and how do we radically increase that, on the theory that it drives the output variable directly, which he says has shown up in a bottom line growing far faster than revenue [5].

That model also made the downturn an accelerant rather than a threat. When rates rose and valuations fell, a free product that pays you to use it "wasn't a line item to be eliminated like so many SAS pieces of software," and Ramp kept triple-digit growth while most companies struggled [5]. He describes it as going "from a nice to have to a painkiller" [5].

On becoming a multi-product platform

Card is still the largest line, a classic interchange model, but bill payments and software is a business barely over two years old already running above $100 million a year, treasury is about a year old with several billion dollars of deposits across checking-like and money-ladder products, and procurement and travel round out the set [3]. On gross profit, card was 90-plus percent a few years ago; by the end of the year he expects the second through fifth lines to be the majority of the business [3]. The entry point is no longer only spend cards: thousands of businesses came in for bill payments in a single recent quarter, accounting firms are describing themselves as Ramp shops and running 100 or 200 clients through one platform, and procurement is probably the fastest growing line [3].

The pitch for aggregation is that customers stop paying for five or six point solutions and stop wasting time, and he ties it to an outcome number: the average Ramp customer grew revenue about 16% last year against roughly 5% for the average US business [3]. The ambition he states is "a better, almost kind of digital brain for organizations to allocate resources and make sure spend is not wasted" [3]. Turning transaction data into savings is the same move he first made at Capital One [4]. Ramp's Price Intelligence aggregates de-identified data from receipts, invoices and MSAs on top SaaS vendors so a customer heading into a renewal can see prices by product line, by monthly versus annual, and by seat count, which he frames as doing for B2B SaaS what Zillow did for homes and TrueCar for cars [4]. Seat intelligence is the companion: connect single sign-on data and Ramp can show that you are paying for ten seats and six people logged in [4].

On expense policy as a moral code

Asked what the right employee spend policy is, he rejects both ends of the spectrum. High-growth companies tend to resemble the Netflix "no rules rules" approach of trust but verify, and permissive spend generally works [3]. But rigid rules produce bad outcomes: staying at Motel 6 and taking a two-hour taxi before signing a two million dollar contract "does not make sense" [3]. What you actually want is a moral code, the feeling that "if you were the owner of the business it is your own money," and the hard problem is preserving that instinct as a company goes from the founder and the founder's brother to tens of thousands of people [3]. He notes the behavioral option of publishing what people spend, with more verification for whoever tops the leaderboard, but sees it as fragile [3]. His resolution is contextual rather than deterministic, reaching for "12 Angry Men": you need more evidence before you can reach a moral conclusion, including outcomes that only appear a month or two after the expensive hotel stay closed the deal [3]. Hence his argument "in favor of tools that have more context, that do more jobs of work," because breadth is what lets you answer the real question of where capital should be allocated [3]. He accepts the framing that expenses are cultural, "a shared belief system" [3].

On agents doing the work no one should be doing

The breakthrough of the past year, in his telling, is that a plain-English expense policy can now be run through an LLM instead of encoded as if-this-then-that rules and verified by managers [3]. Ramp is processing over 100,000 expenses a day reviewed agentically, a very fast-growing subset of the business, with the agent seeing transaction metadata, receipt data, timing and policy, ruling in real time, and keeping a full audit trail of its reasoning [3]. Accuracy is above 99%, "which turns out it's much more accurate than people are," since people generally do not know the expense policy [3]. He is sharp about the underlying absurdity: "I don't think it should be anyone's job to go and review people's expenses, but somehow, while it's in no one's JD, the law kind of requires everyone to waste an hour of their time every month" [3]. Sarbanes-Oxley compliance is preserved because the point is separation of duties rather than human labor, so a system can execute the procedure with a human signing off, and no one self-certifies [3]. He cites Susan Li's observation that reviewing expenses is not a good use of an expensive machine learning algorithm's time as instructive on exactly this [3].

This connects to how he judges his own effectiveness. The risk in organizations that grow bigger and more complex is "you are spending less time than you should actually thinking about the substantive issue, that thing, which is actually the key differentiator" [2]. Automation's real place is "tedious data gathering, putting together things into a presentable form," which gets teams to the thinking faster [2].

On adopting AI inside the company before shipping it

He calls the shift the biggest productivity change in his lifetime, driven by human and in some cases superhuman reasoning available through an API call that any business can access [5]. His prescription is to develop taste before productizing. The first post-ChatGPT wave he judges bizarre in hindsight: "I've never met a single customer who said... I I just would love to chat with my bank account" [5]. So instead of launching Ramp AI, Ramp spent roughly the first year building an applied AI team, a horizontal group, and using the technology internally to reach companies more efficiently, win more deals and learn faster from customers [5].

The concrete examples are a largely programmatic sales motion where lead scoring, outreach and copy iteration are automated so human time goes to interested customers [5], and Toby, a Slackbot built in a weekend on top of tens of thousands of recorded sales calls: "no person has a time to listen to 50,000 sales calls but a large language model does," so a rep can ask why customers pick Ramp over a competitor and get summarized insights across thousands of transcripts in seconds [5]. He notes this generalizes, "you could be selling widgets" [5].

On where AI moats actually come from

He expects classic software defensibility to erode: what took months will take days or hours, "you're not defended based on the features that you have," and vendor lock-in will not hold [5]. His counter-position is that value comes from being deeply integrated down the stack rather than narrow. A narrow AI SDR is "going to be a tough sell," since you can send more email and create spam, while the interesting signal comes from customers you lose and from what happens six or nine months in [5]. That means the real question is upstream and downstream: whether you have a good customer data platform, firmographic and intent signals, fresh rather than years-old lead lists, and a feedback loop from long-run customer value [5]. He frames the whole thing through an analogy of great intelligence severed from a body, invoking Stephen Hawking: profound reasoning with limited ability to sense and act is limited in output, so what matters is "intelligence and action," enough input signals to process and enough surface to act on and get feedback from [5]. This is his answer to investor doubt about whether AI favours startups or the incumbents sitting on large, poorly integrated data sets [5].

On how software gets built now

Roles at Ramp are blurring: a designer ships code, customer support agents ship to production, and marketing reports into his co-founder and CTO Karim, who he says is doing some of the best marketing he has seen [3]. What is shrinking is "the half-life of you see a problem to how long it takes for you to go fix it," to the point where you can ask Ramp Inspect to change the color of a button and it spins up, verifies and validates the change within minutes [3]. On whether that accumulates tech debt, he offers a different shape for code. Tech debt comes from discrete deterministic trade-offs, one code path per condition. In a world of improving models, code might instead be written as a statement of the outcome you want under given conditions, and today's models may deliver that as a Rube Goldberg machine that nonetheless works, with the base implementation rewritten periodically as models improve and accuracy climbs from 90% toward 100% [3].

On why B2B finance software is ugly and bill pay stayed antiquated

The UI gap has two causes in his account. The first is that in B2B software the buyer is not the user, made worse in the office of the CFO because card rules, expense apps and bill payment were separate point solutions that did not connect [4]. Worse, the buyer often "wants control... wants to make things harder," and if you sell on being cheaper you get cheaper, cruder software [4]. He says Ramp had to turn down businesses early on that wanted particular access controls or integrations [4]. The design consequence is stark: in a 500-person company maybe five people have seen the accounting software, two the bank account, eight bill.com, whereas 300 to 400 might use Ramp; those systems of record are "not really designed for use," they are designed for record keeping [4]. Ramp's answer is a mesh where CFOs and controllers keep control at the level of each individual card while employees who have the context get a workable experience [4]. The second cause is institutional: payments companies carry asymmetric downside risk and have built internal processes with legal, regulatory and finance sign-off, so shipping is slow, whereas a bank branch's beautiful design carries no regulatory risk because it is form rather than function [4].

Bill payment is his sharpest case of a network that resisted upgrade. He points out that plenty of comparably old systems were upgraded in place: credit cards began with no real-time authorization at all, then phone authorization, then modems, then today's systems; checks once had to be physically flown around the country until the Check 21 Act allowed a scan to suffice and later phone photos [3]. His explanation for why invoices did not follow is partly complexity that resists if-this-then-that encoding, including accounting treatment that differs wildly between a five-year depreciating asset and a pay-as-you-go SaaS app, and partly incentives, since AR and AP are "almost an adversarial process" where every controller wants to pay late and collect early, which is exactly why paper checks survive when rates are high enough [3]. He distinguishes that from pure dead weight, like mistyped bank details that benefit no one but phishers [3]. The fix he entertains is more connected data: a small supplier owed money by Google borrows at 20% while Google borrows near government rates for a hundred years, and only linking the receivable to the payer's credit lets that mismatch be priced fairly [3].

On compounding, and what Y Combinator actually taught him

He deflates the mystique around the accelerator he went through in 2015, noting it now carries prestige it did not have and that people are "buying and seeking the brand which like wasn't what it was" [1]. The substance was a ten to fifteen minute weekly meeting with the same three questions: how quickly did you grow this week, what is your biggest problem, and how are you going to solve it, where the problem had to be the real thing blocking next week's 10% rather than a theoretical one [1]. His summary: "you don't need to solve the problems of 10 years from now... it's just how little by little how I to get there" [1]. He reaches for athletes to make the same point, that nobody shows up to practice saying they have to train today to win the Super Bowl, they "focus on the inputs the things that were close that were in their control" [1]. People overestimate six months and underestimate ten years, and 1% or 2% a day compounds [1]. He describes Ramp's own history as roughly five years of daily obsession over whether it grew this week [1].

He also resists the overnight-success reading of Ramp. At launch it was not consensus, competitors existed, early coverage framed it as looking like some derivative of X, Y or Z, and the value proposition to early customers amounted to "some nice people you kind of know some cash back maybe they'll save me money" [1].

On what counts as validation

The moment he names as the first sign something had changed was not a round or a valuation. A few weeks after launch he overheard a stranger at the next table pitching and explaining Ramp to someone else, "a sign that like it landed like people could hear it it made sense and it was something you wanted to tell your friends" [1]. He sets that against the grind of early-stage work, where "the biggest battle I think for companies is people not giving a [ __ ]" [1]. Press and valuation marks are pleasant but lagging, and he warns against building a self on other people's approval, because "what comes goes people up to tear people down" [1].

He roots this in his own history: a father who told him as a third-grader that there would always be someone smarter and faster, which he reads as a lesson to stop finding happiness in rankings and put in effort [1]; and an older brother with learning disabilities and an extraordinary memory who was nonetheless chewed apart by teachers for not sitting still, which convinced him that a scorecard "is like just one framework there's a lot of different intelligences," and that standardized schooling "is great for the average but it's not going to capture all the people's spikes" [1].

On money, work and meaning

The clearest illustration he gives of hedonic adaptation is his own: a life-changing cheque after the acquisition, a student checking account he had to move money out of, feeling on top of the world, and then two or three weeks later waking up to the realization that he now worked for a 50,000-person company, that the joy of setting direction, hiring people and watching them succeed had become "a little boat tethered to a giant cruise ship" [1]. His rule is that "highs are never as high as you think they are the lows are never as as is low" [1]. Beyond a threshold, money is not really consumable, "you just can't eat that much"; it can be invested, given to causes or used to build companies, but pure consumption is a waste [1]. What he treats as the actual luxury is scarce time spent in ways that bring joy [1]. Which is why he finds futures where work has disappeared and everyone gets an allocation terrifying: much of the joy is building things with people, and it is the hero's journey of growth and friendship, not the last chapter, that makes stories worth reading [1].

Takeaways

• Ramp's founding thesis inverts card economics: helping a customer avoid a $100 purchase they don't need "is 100 times better than getting 1% back," and the business model is designed to improve as customers spend less money and time [4].
• Only one variable in the card business compounds everything else, purchase volume, so the operating instruction is to make spending on Ramp better than spending anywhere else [5].
• Ramp is no longer a card company by profit mix: card was 90-plus percent of gross profit a few years ago, and bill pay and software, treasury, procurement and travel are expected to be the majority by year end [3].
• Over 100,000 expenses a day are reviewed agentically against plain-English policies at above 99% accuracy, higher than human reviewers, with a full audit trail that satisfies separation-of-duties requirements [3].
• Adopt AI internally to build taste before shipping it; the first post-ChatGPT wave was misconceived because no customer ever asked to chat with their bank account [5].
• Narrow AI tools have limited value; defensibility comes from breadth of input signals and ability to act, "intelligence and action," not from features, which will be copyable in days [5].
• The whole of the Y Combinator method was three weekly questions about growth rate, the real blocking problem and the fix, which is why he treats long-term outcomes as the residue of controllable daily inputs [1].
• Valuations, press and prestige are lagging indicators; the signal he trusted was overhearing a stranger explain Ramp to a friend weeks after launch [1].