Invest Like the Best with Patrick O'Shaughnessy
· Patrick O'Shaughnessy
[00:00:00] S0: My guest today is Karim Atiyeh. Karim is the co-founder and CTO of Ramp, the fastest growing finance automation platform in history, reaching over a billion in revenue in just over five years. Ramp is of course also our presenting sponsor, so I'm obviously very biased in how highly I think about Ramp and about Karim, but this interview was not a part of that sponsorship. I simply view Karim as one of the best operators active today. Ramp is building what Karim calls self-driving finance, using AI agents to automate everything from expense policy enforcement to invoice processing, eliminating the bureaucratic waste that plagues modern businesses. Karim shares his framework for moving from using AI as a productivity tool to programming AI as your actual product with policy agents that understand context better than humans and improve continuously. Our discussion captures the relentless iteration, speed and technical depth required to build generational companies in the age of AI. We explore his systematic approach to building consumer-grade experiences for business software, the psychology behind his divinely discontent management style, and why he believes technical founders will dominate this era because they can see the possibilities that others miss. Please enjoy my conversation with Karim Atiyeh. Karim, we're going to start this conversation at the very end, which is today, because you have one of the most unique perspectives on what's going on in the world as this technology paradigm is shifting where you're not an upstart anymore, you're a big established company, but at the same time you run the company like an upstart, like it's day one and you're up against massive incumbents. And so the reason I'm interested in this cocktail is this is going to happen in every industry where there's big, stodgy incumbents that have been doing things a certain way for a long time. There's going to be fast, talented young companies that challenge them. And so I'm curious for you to detail what it's like to be in that position right now where you're past a billion in revenue, you're one of the fastest companies ever to go to that size in five, six years, and you're up against Amex and companies like this. What does that feel like? What does the competitive battlefield feel like to you today?
· Karim Atiyeh
[00:00:00] S1: It's very exciting. But even when you are describing us as no longer an upstart, I still haven't internalized this, to be honest with you. The most exciting aspect of being a startup is that you get to grow very fast, make decisions quickly and move quickly. So we are trying to maintain that as much as we can. And we do this by essentially breaking up all the different problems we're trying to go after and give small teams full autonomy over these problems. So every time I spent some time with a small part of the company, it does feel like they have full autonomy over the problem they're going after and the way they're building. And I'd say with what's going on with AI advancements and the different ways that companies are starting to build is that every company is trying to figure out how to adopt those new technologies and benefit from them. And you're still in the phase where you have these articles coming out every couple weeks that are mixed. Some will say, oh, companies are adopting AI but not seeing any benefit. And another week you have an article from someone you think highly of or you trust saying that their company has adopted AI in some part of the organization. It's had an immense impact on them. So it's still early. It's obvious to me that the impacts of adopting the technology are transformative. But a lot of companies are still stuck in the very early phases. And 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. So if you're a developer, for example, you are using AI to help you write a little bit more code. So you're not sure on how to write the next couple of lines of code. You go to a ChatGPT, you go to an LLM and have it give you advice. You get some code, you copy it, you paste it, maybe you go a little bit further and you start using these agents where you describe your problem in English, and you use Cognition or Cursor, and then you get a lot more code written for you, and then you're reviewing it. And that's cool. But I think that next phase that we're entering, and the one that I'm seeing in our company, is one where you start thinking about these LLMs really as part of your product. And you're programming them so you're no longer writing the same old code that you used to write with the help of LLMs.
· Karim Atiyeh, Patrick O'Shaughnessy
Your code is the LLM now. Your code is the LLM plus instructions and an infinite loop. So you're essentially writing those agents. We're in the middle of that right now. [00:00:00] S0: Do you have a favorite example of that so far that you've actually deployed in this work?
· Karim Atiyeh
[00:00:00] S1: Yeah, of course. I'd say the most obvious one is what we call internally the policy agents. So most companies have a travel expense management policy. It's generally a document. Sometimes it's well written and clear, and sometimes it's not. But that's essentially a document that you write to drive the behaviors of people in the company and how they manage their expenses. In the old world, people make transactions, and before they make the transactions, sometimes they'll go and check the expense report, sometimes they try to remember it from memory. They'll make a transaction, they'll file an expense report, and some manager will try to make sure that some manager or someone on the finance team will try to make sure that the expense that was made by the employee actually abides by the rules of the expense policy. It's very manual. Generally, the transactions are missing context. So there's a lot of back and forth between the people enforcing the policy and the people who made the transaction. Takes a lot of time. So we've built our policy agents in a way that it has more context about the transaction than most people reviewing those transactions today. So it's integrated with your calendar, it's integrated with your email, it knows your expense policy, et cetera, and more context about the policy than any employee. So it can run 24/7 as transactions come in and apply the logic of the expense policy against these transactions and decide whether something is in policy or not, and do that a lot more efficiently than any human. So you have a 24/7 live enforcement of your policy. And not only that, but as it runs over time, it gets better at advising how to make the policy a little bit clearer or easier to interpret, and your policy becomes better and better over time. Maybe to make that a little bit simpler. Instead of having deterministic codes tell software what transactions are in policy or not, you essentially have this living, breathing text document that can evolve over time that's guiding an agent that has access to tools like your calendar, email, et cetera on how to classify transactions. And that same principle applies in so many different parts of the company. Now the good thing about expense policies is most companies have them.
· Karim Atiyeh
But companies do all sorts of things that sometimes are not written down that can be very easily automated by agents. So for example, when you receive an invoice as a company, what do you do? It's like, well, you do a couple of things. One, you make sure that the invoice is not fraudulent. Maybe you make sure that you're being invoiced for a product that you've actually ordered, a product that you've actually received. You make sure that the price matches what you had negotiated. So there's a lot of steps that actually happen before and after any payment that a company makes. And most of the time they're not very well documented. A lot of work that we're doing right now is because we have so many interesting customers and they're using the product to run their finances. We can infer a lot of these policies through their behavior. And a lot of those policies that we're inferring are driving the next generation of agents that we're building.
· Patrick O'Shaughnessy
[00:00:00] S0: I was with an old friend last week who was in town for, I think Visa had some sort of big conference in New York last week and all the various players from your competitors and people in your industry were there. And he told me that he had no affiliation with you or with me, that everyone there was bitching about Ramp taking all their customers. And I'm curious what you think the most common reasons are for that when you're beating whoever it is. What are the most common attribution reasons when you study why a given customer is picking you over somebody else? Because it seems a little bit slowly, then suddenly thing is happening with Ramp. You and I have talked about this offline and I'm curious why it feels that way to you. I think it relates to the speed of product and everything else, but I want to hear it from you.