The next phase of AI adoption is writing agents as the product, not using LLMs to write code faster.
Karim frames three phases: (1) using an LLM like ChatGPT for advice while still writing your own code, (2) using agentic tools like Cursor or Cognition to generate and review code, (3) making your actual code 'the LLM plus instructions and an infinite loop' - i.e., the agent itself is the shipped product. Ramp's policy agents (real-time expense-policy enforcement with more context than any human reviewer, via calendar/email integration) are his concrete example of phase three.
ai-as-product-not-tool
Undocumented company processes, not just written policies, are prime automation targets once you have enough customer data to infer them.
Most companies have a written expense policy an agent can ingest directly, but invoice review, fraud checks, price-negotiation verification, and similar steps are usually never written down. Ramp infers these implicit policies from aggregate customer behavior across its base and uses the inferred rules to train its next generation of agents.
ai-as-product-not-tool
Ramp's early competitive edge was building consumer-grade UX for a category built to please a single 'decision maker' and ignore everyone else.
Karim argues most prior-generation business software (he singles out Concur) was sold to and designed for the one person who signed the contract, leaving every other employee with a recurring 'small paper cut.' Ramp deliberately aimed for 'the user experience of an Instagram, applied to business software' - prefilling forms, cutting unnecessary questions, and self-resolving errors instead of asking the user to fix them.
company-building-speed
Karim's 'divinely discontent' management style is a deliberate defense against celebrating lagging indicators.
He cites a (misremembered) Jeff Bezos idea that today's results were baked in by work done six months to a year ago, so celebrating current performance risks under-investing in the problems that will determine next year's results. He described being furious over product gaps on the day Ramp announced its $13B valuation round rather than treating it as a moment to relax.
founder-psychology-and-drive
Ramp intentionally split its codebase into 'never break' and 'break constantly and fix fast' zones, a lesson carried over from Paribus.
At Paribus, Karim and Eric built on top of retailer websites that were actively hostile to their scraping, so they optimized for detecting and recovering from breakage quickly rather than preventing it. At Ramp, anything touching money movement or risk is engineered to never break; receipt-matching and UI-level experimentation are allowed to break often because the downside is low and fixing visibly and fast builds customer trust in a young company.
company-building-speed
Ramp shifted from pure card-transaction-volume revenue to software/SaaS pricing because TPV revenue doesn't scale with enterprise complexity.
Card spend as a share of a business's total spend plateaus as companies mature and shift more purchasing to bill pay and procurement systems, so a card-take-rate model caps out on revenue capture from large customers even when Ramp is delivering them enormous value. Charging directly for software also created a feedback signal - customer willingness to pay - that Karim says has improved product prioritization, not hurt growth as the team initially feared.
company-building-speed
Time-based or token-based pricing for AI agents risks rewarding engineering inefficiency.
Karim is skeptical of charging customers by agent runtime or token consumption because it removes the incentive for an engineering team to make an agent efficient - the more compute it burns, the more revenue it generates under that model. He says Ramp hasn't settled on the right way to price its own agents yet, but wants to avoid this misalignment.
ai-as-product-not-tool
Technical founders will disproportionately shape products in this era because they can see what's newly possible, not just what customers ask for.
Using the Ford 'faster horses' analogy, Karim argues customers can only request incremental improvements to what they know; seeing the 'car' requires understanding what the underlying technology (engines then, LLMs now) makes newly possible. He adds that the knowledge gap for a non-expert to reach expert-level competence in an unfamiliar domain (medicine, law, marketing) has shrunk dramatically via LLMs, letting technical people credibly operate outside engineering.
technical-founders-and-ai-leverage
Applying an engineering mindset to marketing meant fixing the system that produces creative work, not the creative ideas themselves.
When Karim took over marketing as CTO, the bottleneck was a two-week brief-review-assign cycle with the brand team for any asset, however small. He rebuilt the system for fast iteration and experimentation (internal AI tooling to generate on-brand images in seconds, accountability tied to the person who owns an idea rather than committee approval) without claiming better creative instincts than the marketing team itself.
ai-as-product-not-tool
Ramp chose yellow as its brand color specifically because no finance company used it, betting differentiation would compound into recognition.
Finance apps cluster around blue (trust) and green (money); the only yellow consumer brand at the time was Snapchat. Karim says the deliberate contrarian choice now means people assume any strongly yellow finance ad is Ramp's, likening it to recognizing a red Coca-Cola can from a blurry movie shot without reading the logo.
hiring-for-extremes
Karim hires for 'spikiness' by testing depth on whatever the candidate claims to be great at, not breadth across a checklist.
His interview approach: identify the one thing a resume claims strong expertise in, quickly get up to speed on that topic himself, then interrogate the candidate's actual depth there. He argues this reveals two signals at once - how good the person really is, and how good a judge of themselves they are - and that optimizing for broad, average competence across many dimensions tends to produce average hires.
hiring-for-extremes
A near-death moment (a legal or existential threat the company survives) is a more useful signal to investors than product-market fit as a stage.
Karim recounts Paribus receiving cease-and-desist letters from major retailers' law firms as a 12-13 person team fresh out of college, including one from Amazon over account security that resolved in Ramp's favor after a call with AWS's own architects. He argues 'almost better to invest right after someone's tried to kill them and they survived' since doing anything correct or disruptive draws attacks, and companies that survive by acting instead of endlessly deliberating tend to keep surviving.
founder-psychology-and-drive
Media referenced
Founders (episode on James Dyson) - podcast - Karim says he re-listened to David Senra's Dyson episode and found the same principle of 'differentiation for differentiation's sake' behind Ramp's early brand and marketing choices.
Companies
Ramp - Finance automation platform Karim co-founded and runs as CTO; passed $1B in revenue in about five years and is building toward 'self-driving finance.'
Paribus - Karim and Eric Glyman's first company, founded 2014; automated price-match and refund claims for online shoppers, sold to Capital One (now Capital One Shopping).
Capital One - Acquired Paribus after partnership talks turned into acquisition talks; Karim and Eric ran the product there for two years before starting Ramp.
Founders Fund - Ramp's first investor; Karim met Delian Asparouhov (about to join Founders Fund with Keith Rabois) while playing Fortnite, leading to Ramp's first pitch meeting.
Virage - Boston startup where Karim did his MIT Research Science Institute research project, converting TV news speech to text and classifying it with Markov chains - an ancestor of LLM-era NLP.
American Express - Incumbent business-card competitor Ramp displaced; Karim's early pitch was literally 'this is not worse than Amex.'
Cognition - Cited via Scott Wu's launch video as an example of a novel attention-getting format that got copied into oversaturation ('a sea of slop of launch videos').
Techniques and frameworks
Policy agents - Ramp's internal term for AI agents that read expense policy documents plus calendar/email context and enforce them on transactions in real time, continuously, and get better at both enforcement and at improving the policy document itself.
Programming the LLM, not coding around it - Karim's framework for the shift from using LLMs as a coding aid to writing 'the LLM plus instructions and an infinite loop' as the product itself.
Hiring for spikiness over checklists - Interview philosophy of finding people with one extreme, provable strength (an Olympiad, a hard class aced as a freshman) rather than broad but average resumes; explicitly modeled as assembling 'the Avengers.'
Break it fast, fix it fast (vs. build it not to break) - Engineering philosophy from Paribus: deliberately ship systems that fail predictably and get fixed quickly, reserved for low-stakes surfaces; money-movement and risk systems get the opposite treatment - built to never break.
Mutual accountability culture - Ramp's stated alternative to top-down management - teams are accountable to peers/other functions rather than only to managers, which Karim credits for cross-functional obsession with the customer.
Summary
Karim Atiyeh, co-founder and CTO of Ramp, walks Patrick O'Shaughnessy through the company's evolution from a scrappy corporate-card upstart to a finance automation platform that passed $1B in revenue in roughly five years. The conversation opens on what it feels like to compete against Amex-scale incumbents while still trying to run Ramp "like it's day one," and quickly turns into Karim's framework for AI: most companies are still in the phase of using LLMs to write code a bit faster, while Ramp is trying to operate in the next phase, where the code itself is an agent - illustrated by its "policy agents," which read a company's expense policy plus calendar and email context and enforce it live, 24/7, improving both their own judgment and the underlying policy document over time.
A large stretch of the episode traces Karim's personal arc: growing up in postwar Beirut, a summer research program at MIT that exposed him to the ancestors of modern NLP, and Paribus, the price-tracking company he co-founded with Eric Glyman that automated retailer refund claims and got so aggressive with email scraping that it drew cease-and-desist letters from major retailers' legal teams - including a scare from Amazon that ended with a validating call with AWS's own architecture team. Paribus sold to Capital One, where Karim and Eric spent two years learning the economics of the card business before deciding businesses, not consumers, were the bigger opportunity, which became Ramp.
The middle of the conversation is a tour of Ramp's build philosophy: split engineering into "never break" zones (money movement, risk) and "break fast, fix fast" zones (UI, receipt matching) inherited directly from Paribus's experience building on hostile third-party websites; a deliberate move from pure card-transaction-volume revenue to software pricing once TPV growth was shown to plateau against enterprise complexity; and a "mutual accountability" culture where teams answer to peers across functions rather than only to managers. Karim is candid about his "divinely discontent" management style, tracing it partly to a Bezos-style discomfort with celebrating lagging indicators and partly to simply not knowing what he'd do with himself if there were no more problems to solve.
A significant segment covers Karim personally taking over marketing as CTO roughly a year and a half before the recording, applying an engineer's instinct to fix the system that produces creative work (cutting a two-week brief-to-output cycle down to minutes) rather than trying to out-create the marketing team. He connects this to Ramp's early, deliberately contrarian choice of yellow as a brand color (the only other yellow finance-adjacent brand was Snapchat) and to his broader belief that technical founders will disproportionately shape products in the AI era because they can see what's newly possible, the way engine designers - not horse-and-buggy customers - could see the car coming.
The episode closes on hiring philosophy (interviewing for "spikiness" in one provable area rather than checklist breadth, illustrated by an early hire who'd left high school early for the Informatics Olympiad), a reflection on how his view of company-building has shifted from chasing an end reward toward valuing the people he does the work with, and Patrick's traditional closing question, which Karim answers with a story about a Lebanese-American MIT friend's family taking him in unconditionally when war broke out back home and he was stranded in the US at 16.
Notable Quotes
"Your code is the LLM now. Your code is the LLM plus instructions and an infinite loop." - Karim Atiyeh
"I would challenge you to find anyone who enjoys using Concur, for example. There's a lot of business software that was built in that way. No care for the user experience." - Karim Atiyeh
"If you're doing anything that's correct or right, people are going to try to kill you multiple times." - Karim Atiyeh
"If you want an additional widget here, an additional button there... they don't realize that the car to their faster horse is possible." - Karim Atiyeh
"The more things you are looking to vet someone on, the more likely you are to get average people, essentially." - Karim Atiyeh