Who Wants to Chat With an Expense Report?

Ravi Gupta and Sonya Huang of Sequoia sit down with Eric Glyman, co-founder and CEO of Ramp, a command-and-control system for company finances: issue cards, manage approvals, make payments, even automate closing the books. The average company on Ramp saves about 5% a year on expenses; over five years that has added up to billions of dollars and the equivalent of thousands of years of labor. Customers run from Shopify, Virgin Voyages, and the Boys and Girls Club of America to more than 25,000 other businesses. Glyman names the ambition in a phrase the hosts will keep repeating: self-driving money. The through-line of the episode is a cold shower for the industry's chatbot reflex. After ChatGPT, everyone slapped a chat box on the product. Glyman has a different take. Zero-touch automation that you never see can be worth more than a conversation.

Gupta, a Sequoia partner and former Instacart CFO, is especially interested in something Glyman posted: he and Karim built an AI agent in 2015, years before it was cool. Paribus was simple and, Glyman agrees, fair to call an agent. A decade ago, Amazon, Best Buy, and other retailers would refund the difference if a TV you bought for $1,000 went on sale for $900 the next day — if you noticed, wrote, and argued. Paribus plugged into Gmail or Yahoo, scanned receipts, tracked prices, and, when a drop made you eligible, generated an email that sounded like you. You woke up $100 richer. Millions of people used it. It was an agent that lived in the inbox.

The difference between then and now, Glyman says, is not the primitive. Software that does things on your behalf has been here a long time. Paribus was extremely narrow and hard to build: one job, price-adjustment guarantees. In 2024 the set of cases where a computer can work for you is far wider — not just refunds, but helping a finance team run, even negotiate, even run complex analyses. Generalized reasoning is what changed. For builders, the implication is the opposite of "invent a new Agent era." Look for the workflows where people are already being used as robots, and point the old primitive at a much larger surface.

Forget "Ramp AI," Glyman says. The mission is helping people spend less money and less time; AI is a new set of tools for that, and it only matters where the pain is. Finance teams are full of tedium he knows Gupta lived at Instacart: auditing expense reports by hand, downloading spreadsheets to tag vendors, rerunning the same analysis. The simple picture is two apps to buy one thing — American Express for the card, Concur for the expense report — with humans shuttling receipts between them. Ramp wants one tool. You tap the card; through zero-touch it pulls the receipt from the merchant and your email; the report is not easier, it is done. Large language models read the transaction and the general ledger, autocomplete the memo, and for the vast majority of customers categorize faster and more accurately than a person. The books, he says, are effectively doing themselves. Over time Ramp should keep pointing at better prices from the same vendors and the same processes run automatically. Scope is whatever saves customers time or money: higher yields, more efficient collections, cleaner record-keeping. He would rather call Ramp a productivity company than a money company. Finance is the carrier; the product is the intersection of time and the flow of capital. The end state is a finance department that improves itself — self-driving money.

How will people actually use it — prompting Ramp, or work happening in the background? After ChatGPT, the first phase was a chatbot slapped on everything. Glyman has never met anyone who said they wished they could chat with their bank account, or with their expense report. Internally, Ramp splits AI experiences into two buckets. The first is zero-touch. Paribus was an early sample: sign up, link email, no UX after that, wake up and it is done. Tap the card, the expense is filed; accounting runs on your behalf; memos are suggested so you click one-two-three. The question is which inputs you still need, and whether confidence is high enough to go all the way, from administration through analysis to closing the task. That surface, he thinks, will keep expanding.

The second bucket is agentic AI: you name an outcome, maybe monitor along the way, and once trust is there you stop monitoring. A few weeks after OpenAI released GPT-4o, Ramp shipped a "tour guide." You can say, issue a card with a $50 budget that can only be used at Starbucks, go — and the model reads everything on the screen and walks you through clicks and fields. It was not a scripted demo; the reasoning of the large model is now capable of that. Most business work, Glyman argues, already looks like this. The interface is the one Gupta has joked about: you hire an analyst and say, I want this done, go figure it out. Status quo is strange by comparison. People are trained to learn your app. It makes more sense for them to think about running the business, selling, finding ROI. Done right, you should be able to say what you are trying to accomplish and have it fetched. Early on you may watch. Over time, as trust and reasoning rise, you will see the handoff. He draws the self-driving analogy: a decade ago the hype car still had a person in the seat ready to grab the wheel; now Waymos fill the Bay Area, and a human in the driver's seat looks odd because the machine is more accurate. Sets of tasks will get handed off the same way, an agent driving parts of the organization at high fidelity.

Huang asks whether buttons and knobs fade into the background. Glyman says yes, very much so. Great design is understanding the job to be done so well that you can cut steps until the thing is effortless. Users will still want to audit what the powerful tools did. Day to day, it will more often be a prompt: here is the outcome, go. Among Ramp's 25,000 businesses, some have sophisticated finance teams and some are small operators drowning in the baggage of stitched-together tools. Time that should go to clients, pitches, and the next product gets stolen by learning an amalgamation of software. Done right, the world should feel more frictionless. If the industry cannot accomplish that, he says, it has been a failure.

Ramp is known for speed, quality, and customer obsession. Gupta wants to know how Glyman and Karim built an engineering culture that actually wants to implement AI well, when so many places resist. Glyman goes back to positioning. They do not market an AI-driven finance tool, and they do not ask, do you want to adopt AI on your finance team? They exist to save you time and money. They lead with the benefit and put the outcome in simple terms. It is not possible to use Ramp today without using AI — when people say expense reports got easy or the books closed faster, that is AI inserted at many points in the process — but AI is a means, not the end. They did not want technology in search of a problem. Decompose where hours and capital are wasted, and process automation turns out to be the right tool for the job, again and again. You do not ask people whether they would like this model. You ask whether they would like this expense-report problem gone.

The second organizational rule is not to abstract the makers away from the problem. At a ten-person startup, everyone talks to users. At 500 to 1,000 people, you ask who spoke to a customer last week and, outside sales, you get silence from marketing and engineering. Glyman thinks the most important thing beyond empathy, if you want great products, is taste. The people making the thing have to understand the customer's pain not just decently but, in some cases, better than the customer does. Only then can you design well enough to automate the task. That is why, when Ramp ships, engineers are on the call with customers, accountable for the metrics. You will not find people at Ramp who have not talked to customers with any recency. He calls that a core of great product cultures.

Has the last year or two of foundation models been a discontinuous jump, or just more AI magic compounding in the product? Glyman, speaking as a practitioner rather than a researcher, calls it a step-function change — the biggest shift to productivity in his lifetime. The ramifications for anyone building software are blunt. Old moats were more features, more integrations, lock-in after ten years of data you cannot extract. Now human-level reasoning, and in some cases superhuman, is available through an API. That revalues every software business. Inside Ramp the same shift shows up off the product: AI automation in outreach, SDRs multiple times more productive than at competitors, a different customer-service motion, different copywriting, the ability to listen to 100,000 sales calls at once and ask what 100,000 people thought. None of that was possible a few years ago. Most people, he thinks, are still experimenting with ChatGPT. Far too few have put AI into the crevices of how they actually work day to day and felt it. He thinks that should accelerate.

What does an excellent finance leader's job look like in five years if they lean in? Glyman says it is incumbent on anyone — finance leader, engineer, designer — to automate the parts of the job they dislike or that are low value, because a whole class of those problems can be automated now. From conversations with Gupta he has internalized a definition of leadership: great leaders do not just create focus and urgency, they identify where there are outsized returns to time. In a lot of finance organizations the CFO may have that focus, but look at the rest of the team's calendar. The month closes, and the first six days — eight in bad cases, 15 in worse — go to tagging transactions, downloading spreadsheets, matching things. Only in the last five days of the month does the work they got into finance for actually start. People should be turning those rote streams into an automated process now, so that in five years the work is more strategic, more about where value is created.

Huang asks the classic question: when the system is more efficient, is everyone out on the golf course? Glyman thinks a lot of purpose in life is creation; people will find more time for leisure, but they will also want to move things forward. A statistic making the rounds a few months earlier is the one he reaches for. There was supposed to be a crisis: the United States had about a million fewer bookkeepers over the last ten to twenty years. Look at job descriptions for financial analysts, strategists, CFOs, and those roles had grown by almost a million. People had not vanished. They had moved from tagging and tabulating to a higher level of abstraction, work that is uniquely human, high value, and in many cases more fulfilling. He expects the same pattern in the AI era. Some creative work computers will do better. Creation itself, the part that gives him excitement, he still thinks is a very human thing.

Advice for founders, in this moment, starts with the length of the race. Great companies are built over years and decades; Glyman hopes Ramp is the last company he ever works on. There is always a question of what is changing and which industries get reset, but what has made Ramp work is starting from timeless truths — things that will not change in ten years or a hundred. Paraphrasing Jeff Bezos: a hundred years from now nobody will say they wish Amazon had raised prices, or delivered a little more slowly. Same for Ramp. If you can create great work with less effort, less time, fewer dollars, that will always be in style. Be curious first about real customers, real businesses, problems that are not going away, then uncover the technological shifts that let you solve them in a new or disruptive way. Gupta quotes a friend: try to be timeless rather than timely, because the timely evaporates. That, he says, is how Ramp is being built.

Rapid fire. Favorite AI app is not Ramp, though he jokes it. From a UX angle, Cognition's Devin has bent his thinking. They are trying to build an AI engineer, and they gave the agent the kit an engineer would have: a notebook, a planner, a browser to hit Stack Overflow when it gets confused, shell access, connections to your tools. What is most fascinating is that you can watch what Devin does, what it gets right and the mistakes it makes. That taught him a new way to look at how these tools may feel. Outside finance, the industry he most hopes goes "self-driving" is healthcare. Diagnosis is already changing; wellness is large sets of data, not an annual checkup, but continuous measurement over years. For many doctors the job is glorified note-taking, with too little diagnosis and listening. AI could return some of the humanity to care. Design and creation, too: the question stops being whether you understand the tools and becomes whether you can make something interesting. Gupta's analogy lands: companies stare at hourly dashboards, then check their own health once a year and often skip it. You would not run a business that way.

Inside financial services, Glyman thinks the sector has a long way to go. Competitors like American Express, Chase, and Citi are led by wonderful people, but their founders wore top hats. The world went from no phones to flip phones to iPhones and a lot of the underlying logic never changed: if you were a bank you could move and store money, and if you were not, you were the rest of the world. That is a big part of why the industry moved so slowly. It is wonderful, he says, that it is no longer only banks. Ramp is a fintech, but he thinks of it as a productivity company, interested in how time intersects the movement of money. Disruptors will come from adjacent places. Apple would blush at the label, but he thinks Apple is a true fintech innovator right now, connecting the movement of money to who people are, cutting fraud and friction online and in person. A lot of the next wave will not come from traditional finance companies.

Who does he admire most in AI, present company excluded? Satya Nadella. He partnered with OpenAI to take what was primarily research into production, into almost every aspect of computing, and in doing so helped create the category. He has been a mentor to Glyman and the Ramp team on how agentic AI can be designed over much larger cycles — not weeks and years, truly decades. Glyman mentions sincerity and kindness, and a book, *Hit Refresh*, plus a gift for partnering deeply with all kinds of companies. Huang closes on the phrase the conversation added to the lexicon: self-driving money, self-driving finance. The combination of innovation and applicability, she says, is the rare part.