Summary

Karim Atiyeh’s core AI claim: most companies are stuck in phase one — using LLMs to do the same work a bit faster — while Ramp is entering phase two, where “your code is the LLM plus instructions and an infinite loop.” The deployed proof is a policy agent with calendar/email integration that enforces expense policy 24/7 with “more context about the transaction than most people reviewing those transactions today,” and Ramp is now inferring customers’ undocumented finance policies from behavior to drive the next generation of agents.
The TPV-to-SaaS pivot was forced by math, not fashion: card spend plateaus with business complexity, so “our mechanism for capturing value… breaks for large businesses.” Internally feared, the shift to charging for software accelerated growth — paid products made sales pitch them and customers demand more. He’s still unsure how to price agents, and explicitly distrusts per-time/per-token models that “incentivize the engineering team to not be as efficient as possible.”
The runway argument for the bull case: past $1B revenue in five-six years, conversion above 50% in some segments — yet still “sub 2% of corporate card alone,” with the ambition expanded from the card to every workflow before and after money moves, target state “self-driving” finances where customers never log in.
The anti-MX wedge is structural: incumbent business software is built for the decision maker and gives “everyone else at the company just a small paper cut every day”; Ramp inverts the model — and inverts card economics too, helping customers spend less rather than dangling rewards, betting share of wallet follows.
Engineering doctrine worth stealing: split the product at the boundary of “build it so it never breaks” (money movement) vs. “break predictably, fix fast” (everything else). “What’s one great way to make sure that you have no bugs? Don’t ship anything… if you are solving for great outcome and great impact, you want things to be breaking.”
On rails and stablecoins, he’s deliberately agnostic — card costs are substantially driven by consumer rewards, not simply Visa’s cut, and stablecoins’ consumer value prop is unclear; but if agents become the payment decision-makers, rails could shift, and “I couldn’t care less whether that runs on likely ACH rails or the card rails or the stable coin” — Ramp wins either way.
Company-building signal for investors: the best time to buy may be right after someone tries to kill a company — “if you’re doing anything that’s correct or right, people are going to try to kill you multiple times” — and the 2022/23 down-round was reframed as pure price discovery: “the act of raising a round only just makes it known.”

Deep dive

1. Past $1B, still an upstart — and AI’s real phase change is “your code is the LLM”

Patrick’s frame: Ramp is the rare specimen — no longer a startup, one of the fastest companies ever to $1B+ in revenue in five-six years, yet run like day one against incumbents like MX. Karim’s honest reaction: “I still haven’t internalized this” — the mechanism for staying fast is breaking every problem into small teams with “full autonomy over the problem they’re going after.”
His AI adoption taxonomy: phase one is using LLMs to do the old work faster — copy-pasting code from ChatGPT, or going further with agents like “a Cognition or a Cursor.” The phase Ramp is entering: “your code is the LLM now. Your code is the LLM plus instructions and an infinite loop” — the LLM is programmed as the product itself, not an assistant to writing it.
He acknowledges the mixed public evidence — articles every couple of weeks alternating between “no benefit” and “immense impact” — but is categorical that “the impacts of adopting the technology are transformative”; most companies are simply stuck in the early phase.

2. The policy agent: the concrete specimen of an LLM-as-product

The example he leads with: every company has a travel & expense policy — “a document you write to drive behaviors” — enforced manually, after the fact, with endless back-and-forth. Ramp’s policy agent is integrated with calendar and email, knows the policy, and runs “24/7 live enforcement” with “more context about the transaction than most people reviewing those transactions today.”
The compounding part: over time the agent advises how to make the policy clearer — “a living, breathing text document that can evolve over time that’s guiding an agent.”
The extension is the bigger business: most of what companies do around an invoice — fraud check, matching to what was ordered and received, price verification — isn’t written down anywhere. Ramp is inferring these implicit policies from customer behavior at scale, “and a lot of those policies that we’re inferring are driving the next generation of agents that we’re building.”

3. Why Ramp takes customers: paper cuts vs. consumer-grade design

Patrick relays intel from a Visa conference: “everyone there was bitching about Ramp taking all their customers.” Karim’s attribution goes back to the original wedge — previous-generation business software was built for the decision maker: “you solve one person’s problem, but you give everyone else at the company just a small paper cut every day.” Ramp wanted “the user experience of an Instagram but applied to business software.”
The repeatable practice, for anyone building consumer-grade B2B: interrogate every customer interaction — “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?” Retry the failed payment yourself; put the form inside the email. Design obsession matured into “an obsession over minimizing the amount of time people spend in our app.”

4. Divinely discontent: why he was furious on the $13B day

David’s story via Patrick: on the day of the $13B valuation announcement, Karim spent it screaming about product problems. His defense is analytical, not temperamental: “current quarter was baked in a couple months ago… I always find it a little bit weird to celebrate lagging indicators” — celebration risks convincing people today’s problems don’t exist and “we might mess up the next six months.”
The deeper motivation: the charts of US healthcare and education spending rising for decades with outcomes probably better, but not as much as expected given the spending — “it’s just being wasted on administrative BS… not being spent on more training for doctors… It’s being spent on a lot of bureaucracy that doesn’t move the needle.” Ramp’s mission, in his telling, is to attack exactly that waste.
The second reason, disarming: “if things were good and we didn’t really have problems, I wouldn’t know what to do with myself… the good thing is the job’s never done.”

5. Culture as mutual accountability — and the one interview question

The tell Patrick offers: a founder selling to Ramp was told the deal wouldn’t close “until you’ve confirmed that you are a Ramp customer for your business.” Karim’s explanation: Ramp runs on “mutual accountability to each other as opposed to top-down” — sales cares about product, marketing about finance — so everyone sells, everyone advises the builders.
His entire interview, if he had to summarize it: “if this person was starting a company… would I join them, or would I start a company with that person?” Because “a company is just a collection of people solving problems together, one after the next, and they keep getting more difficult and bigger — the question is how much can you endure, for how long.”

6. Beirut to MIT: growing up on ephemerality

The prologue starts in Beirut, born “‘89, ‘90” at the end of the civil war, raised by a generation “scarred from the war… everyone was kind of living on edge.” The formative residue: “a sense that things are very ephemeral and could disappear very quickly was there in the air” — do your best, get a visa, get out.
At ~16 he found his way to MIT’s Research Science Institute (RSI) in 2006 after obsessively noting that every discovery in the science magazines came from MIT. He’d chosen computer science thinking he’d learn it — then discovered he was expected to do research, and taught himself under pressure.
His RSI project at a startup near the Kendall subway stop called Virage: converting local TV news feeds to text and classifying it with Markoff chains — “essentially the ancestors of what LLMs rely on” — and training those models more efficiently. The lasting output, by his own account: the friends, some of his best to this day.

7. Paribus: arming the rebels with an Excel macro

Genesis: Eric (same co-founder as Ramp) noticed he’d paid a different price for a flight than someone else, right as retailers were hiring data teams to dynamically price. Karim, a discontented consultant already sneaking software into manual projects (consulting “pricing for number of people spending time on projects” killed any incentive to automate), built v1 as “VBA macros in Excel” checking Amazon prices on a set of SKUs daily.
The insight: prices were changing fast — and accelerating even within the first month of tracking in 2014 — while every retailer offered price-match guarantees they were “banking on the fact that no one really checks.” Paribus automated the check: retailers had “armies of data scientists” to price discriminate, “and we’re going to arm the rebels.”
Scale and fragility: ~1M users within a year, built atop retailers who “one, don’t want us there; two, are changing their pages all the time; three, are heavily incentivized to make it harder” — Amazon stopped sending itemized receipts entirely. At one point: “one of the largest scraping operations in the US… billions of emails per day, tens of millions of receipts per day.”

8. The engineering doctrine: break predictably, and split the product on that line

The Paribus lesson: “you’re never going to build the perfect system. You’re better off building something quickly that will break in very predictable ways and that you are able to recover from incredibly quickly” — the opposite of big-company engineering that optimizes for never breaking.
At Ramp this became an org-design decision: anything touching money movement and risk gets the never-break treatment (“assume you’re not going to have to innovate that much on it”); everything else — receipt matching, form optimization — iterates fast and breaks. “There are little parts of Ramp that are probably breaking 10 times a day and we’re fixing them 20 times a day and no one’s really noticing.”
The counterintuitive trust claim: for a small startup, “it’s a much better way to build credibility… if things break and you fix them very quickly.” And the maxim: “what’s one great way to make sure that you have no bugs? Don’t ship anything. Don’t write any code… if you are solving for great outcome and great impact, you want things to be breaking.”

9. Cease-and-desists, the Amazon call, and why kill attempts are a buy signal

The crazy-moment story: a team of 12-13 fresh graduates Googling the multi-thousand-person law firms sending them letters. The best one — Amazon accused Paribus of compromising account security, while the product was built on AWS with AWS architects’ help. The response: “we’d be happy to get on a call with your team that is helping us build this” — a call with “either likely Andy Jassy or someone very senior on his team,” which ended with AWS seemingly excited and the problem clearly Amazon retail’s.
Patrick’s investing hypothesis — better than product-market fit is the moment someone first tries to kill you and you survive. Karim: “Oh, 100%. If you’re doing anything that’s correct or right, people are going to try to kill you multiple times.” His survival mechanism: “I never see the possibility of us being killed as an option” — companies die when they “spend a lot of time thinking and talking and not enough time doing.”

10. Capital One: selling by accident, and discovering the card business model

They weren’t selling — they were hunting distribution: “who better than the credit card companies” to know who just shopped at Amazon or Walmart. Partnership talks with Capital One (interested in product differentiation because “consumer card products are still different versions of the same product just marketed differently”) turned into acquisition talks just as the legal storms demanded “more firepower.”
The pivotal education inside the deal: the card model itself — “there’s no contract anyone needs to sign… the more they use it, the more revenue you make, and you can just focus on making the product as good as possible.” Paribus became Capital One Shopping; two years later, the itch returned.
Ramp’s founding logic followed directly: “Paribus for businesses” — save money — which became save time and money, “because businesses tend to waste a lot more time than they do money.” The card, like Paribus, was primarily a sensing mechanism: know what businesses spend on and you learn where they waste. Since then “the scope of our ambition” has grown from the transaction to everything before it (procurement) and after it (accounting and reporting).

11. “This is not worse than MX” — the launch, Fortnite, and the investor playbook

The candid origin pitch, verbatim: “it wasn’t ’this is amazing and going to change your life.’ It was ’this is not worse than MX and you might as well give it a try’… it was a terrible pitch.” Early customers were YC-batch friends and his brother — “partners more than they were customers.” The first real differentiation: privacy-preserving email parsing that matched receipts to card transactions, plus tech that turned “MCDX” merchant gibberish into “McDonald’s.”
They didn’t want to raise — second-time founders with capital, and receivables were a funding need. Then the Fortnite story: playing with his brother’s friends, one turned out to be likely Delian Asparouhov, apparently on some sort of garden leave before Founders Fund; “I think I’m going to stop playing Fortnite… Eric and I are starting another company” led straight to pitching likely Keith Rabois — Square/PayPal pedigree, “uniquely positioned to understand exactly what we were trying to do.” They wanted him “more for the advice than the money, which is I guess the best way to meet investors.”
The investor philosophy since: investors are employees who pay in capital — “you can’t buy Ramp stock on the open market… these are the two ways you are able to get Ramp equity” — so onboard them, educate them continuously, and let excited funds build positions across rounds (“might be all of them have participated in multiple rounds”).
The hardest round, ~end of 2022/23: not hard to raise, hard to justify — “we don’t really need the cash and the valuations are down, so what the hell is the point?” His resolution: “the act of raising a round only just makes it known… it’s just a price discovery mechanism” — set the checkpoint, build from it, “we’re never really trying to maximize what the valuation is at every single round.”

12. Card economics break at scale — the SaaS transition and the agent-pricing puzzle

The land-and-expand specimen: an unnamed aerospace company famous for building everything in-house took only the card and the API — “all the other software we want to build ourselves” — then six months later adopted the AP automation, then more, until the expansion motion “really started to become real” around year 2.5-3.
Why pure TPV had to end: Ramp wants customers to spend less — “the best ways to help them save money is to help them not make transactions that they should have never made, as opposed to giving them points” — but card spend “does not scale linearly with the complexity of the business”; large companies route spend through bill pay and procurement, so “our mechanism for capturing value… breaks for large businesses.”
The feared move paid off doubly: charging for software “not only did it not [hurt conversion], in many ways it’s accelerated growth” — it incentivized sales to pitch the products and made customers “a lot more demanding and responsive,” a market feedback signal a free product never generates.
Still unresolved: agent pricing. He worries the emerging charge-for-agent-time/tokens model “incentiviz[es] the engineering team to not be as efficient as possible with their use of AI… that’s why I don’t really love that model. I’m not quite sure exactly how it’s going to look.”

13. Technical founders see the car; everyone else asks for faster horses

His argument for why technical builders win this era: customers ask for “an additional widget here and a button there” — the Ford problem — while “the people who came up with the technology… see the possibilities in terms of product” better than domain experts.
The second leg: “the gap between not having subject matter expertise in a domain and having it is the smallest it’s ever been.” His joke that carries the point: “I’m a better doctor than I’ve ever been, but I’m not a doctor. I’m a better lawyer than I’ve ever been, but I’m not a lawyer” — engineers without procurement or accounting experience can now build the future of those workflows.

14. The CTO takes marketing: fix the system, not the creative

The trigger was leading-indicator paranoia while “all the lagging indicators were going extremely well”: conversion above 50% in many segments, monetization improving — both with hard ceilings — while lead generation was slowing and Ramp remained “sub 2% of corporate card alone.” “Conversion can’t get better than 100%.”
The attitude shift he imposed: when a podcast ad flops, the old read was “podcasting doesn’t work for us”; his read — “no, it has to work, we just haven’t figured out how to make it work. You picked the wrong person or the wrong format.” Applied across direct mail, paid, brand, product launches, “the Elon algorithm” applied to the work-generation system itself.
The clock-speed fix: the brief-to-brand-team pipeline meant everything took two weeks minimum; he rebuilt it so an idea reaches the world in 10 minutes, not two weeks — “I didn’t go into it with ‘I have better creative ideas’… I just want to put their ideas in front of the world as quickly as possible,” enabling more shots on goal and more risk per shot.
The billboard hack, as told: a New York billboard costs ~$100K, a second one $200K — but changing an existing one costs $1,000. “One billboard changing seven times in a week [$107K] is a much more powerful way to drive a message than two static billboards [$200K].”

15. Attention has alpha decay; differentiation is the durable move

The paid-media lesson from Paribus: when Facebook made video ads stop playing sound by default (~2014-15), effectiveness cratered, ads got cheap — and the winners were videos legible without sound. Their answer: Eric and Karim in banana costumes holding text signs. “It worked very well for three or four months and then it stopped working” — every channel edge decays, so hire people obsessed with platform changes; same logic for SEO’s shifting reward functions. “Speed super important.”
On brand, channeling a Senra episode on Dyson — “seeking differentiation for differentiation’s sake.” Every finance app sat in blue/green on the color wheel (“green for money, blue is trust”); the only yellow was Snapchat. “The primary reason why we chose the yellow was because it was different. That’s it.” The payoff is Coca-Cola logic: “you see a red can in the distance… you know it’s a Coca-Cola can” — now a very yellow ad or bus reads as Ramp. Patrick’s adjacent complaint stands as the cautionary tale: launch videos went from Cognition’s breakthrough to “a sea of slop” nobody watches.

16. Hire the Avengers: spikiness over checklists, speed over estimates

The recruiting edge at Paribus was asymmetric information: unable to outbid Google, they hunted freshmen taking the hardest Harvard/MIT classes they personally knew — “in year one, if you’re taking that class and got a really good grade, you are extremely talented.” Exhibit A: Calvin Lee, hired off a one-month January internship, who left high school early for the informatics olympiad, finished college in 2.5 years, and has since spanned engineering, sales, and forward-deployed engineering at Ramp.
His interview method: no 10-item checklist — take whatever the resume claims greatness at, research it for a couple of hours, and test it. Two signals: “how good of a judge are you of how good you are,” and how far the skill actually goes. “The more things you are looking to vet someone on, the more likely you are to get average people.” He wants “the Avengers… where everyone has a clear superpower.”
The speed doctrine, via the story of Jeff, Ramp’s first PM, who was “appalled” nobody sized tasks in story points: Karim’s pushback — a “Shreddinger’s principle” of estimation: “you can get a lot of precision on how long things take, or you could do them very fast. It’s hard to get both,” because rewarding accurate estimates incentivizes padded ones. Broader claim: “best-practice process is a good way to move you to average” — the fastest people do things “in a very odd, nonstandard way.”

17. Rails don’t matter, self-driving finance does — and the reward is the journey

On payments infrastructure: the misconception is that Visa’s cut makes card payments expensive — “a lot of that expense comes in the form of rewards” Americans are very attached to. Stablecoins promise merchants faster/cheaper, “but it’s still not very clear what it is for consumers.” The genuinely interesting shift: if agents make the payments, they might care less about rewards and “might optimize the rails” on a different maximization function. His position: “I couldn’t care less whether that runs on likely ACH rails or the card rails or the stable coin” — Ramp benefits from whatever wins, and yes, it’s “kind of crazy that payments cannot settle on a weekend or outside business hours in certain cases.”
The 2030 vision: “my hope is that people don’t have to log into Ramp at all” — a self-driving-car analogy, from mechanical cars through lane assist to sitting in the back seat: finances that are self-driving, with the flywheel being both better models and the accumulating behavioral data (“we can infer a lot of their intent from their actions and the way they decide to fill forms”).
The retrospective lesson across Paribus and Ramp: he used to assume challenges ended in a reward that would feel great; now — “the reward is just the journey… the reward for solving challenges is just more complicated challenges over time” — so optimize for the people you solve them with. What he’s proudest of: the talent (an intern won an International Physics Olympiad gold this summer) and the diaspora — “I would love for Ramp to be the last job that they ever have to apply for,” even knowing that means everyone tries to poach them. Plus the origin-story coda: at 16, war broke out in Lebanon during RSI, and his friend Zach’s family took in a stranger — labeled meals, pocket money, hidden key — “they’re already treating me like family.”

How Kong Was Born: APIs, Hustle, and the Future of AI Infrastructure The Singularity is Here: AI is Solving Math, Sora Outpaces Chat-GPT & AI is Designing Chips | EP#201