In 30 seconds
Ramp co-CEO Eric Glyman says an AI model is far more useful when it can work with the data and tools a person is allowed to use. He describes internal coding and knowledge systems, shared skills, and a quarter-long hiring pause intended to give employees room to learn new workflows. He argues that AI spending should be judged against results, not a fixed token budget. The same principle shaped Ramp’s earlier financial products: finding waste is useful, but helping a customer stop an unwanted payment is more powerful.
- 1Spotting waste and stopping a payment are different jobsParibus could ask for refunds through customer service. Ramp’s founders wanted the ability to act closer to the movement of money.
- 2Uneven adoption can move the bottleneck to another teamGlyman says engineers sped up at a different rate from design and product, exposing friction in existing handoffs.
- 3A model needs permissions and integrations to use company contextHe contrasts a simple chat box with Inspect, which he says can work with an engineer’s authorized code, data, virtual machines and browser.
- 4A token budget says little without the resulting valueGlyman contrasts an AI expense that earns a return with the same spend on work that creates little value.
- 5Shared skills can carry a useful workflow across a companyHe describes Glass connecting business systems with individual permissions and reusing practices learned by one team.
- 6Customer-facing staff can propose fixes themselvesGlyman reports 7,000 pull requests from non-engineers in a quarter and describes shorter paths from a customer’s bug report to a proposed change.
Takeaways
- Connect an agent to the information and tools it is permitted to use, rather than expecting a chat window alone to change a workflow.
- Judge AI spending against the outcome it produces, Glyman argues, rather than a single flat budget.
- Watch for new bottlenecks when one function adopts AI faster than the teams it depends on.
- Look for demonstrated initiative in hiring and ask for help explicitly when a problem is unclear.
Read the edited text by chapter
From a flight refund to Paribus
A price drop prompted Glyman and Atiyeh to build a service that found and requested refunds. Customer conversations helped them explain it and design referrals.
Eric Glyman and his co-founder Karim Atiyeh started with a way to recover money when an online purchase fell in price. Their first company, Paribus, connected to email, found receipts, checked prices and requested eligible refunds on the customer's behalf. They applied to Y Combinator twice before being accepted.
Glyman says the company grew quickly and was acquired by Capital One. At the bank, he learned how a large financial company used information, but also saw processes and incentives that could prevent good ideas or put the company's interests ahead of customers'. That experience shaped his interest in products that help customers spend less.
The Paribus idea came from Glyman's own life. After friends booked a flight, he saw the price had fallen by $100. He read the airline's policy, called within the allowed period and says the group recovered nearly $1,000. He wondered how often the same thing happened without people noticing.
The founders spent about a year experimenting with where to find purchase information. Card records did not contain itemized receipts, but email did. Their first scripts checked prices and sent refund requests; they then asked friends to connect their accounts so they could test them.
Early users were hard won: the founders carried signs and spoke with friends. One Reddit post in a community looking to earn small amounts brought in a few hundred users. During YC, they focused on user growth and tried to explain the service simply: stores owed people money, and Paribus would help recover it.
The offer also sounded too good to be true. Glyman says the team observed how users explained both its benefits and the catch—that Paribus retained 25% of savings. They introduced a two-sided referral offer that lowered the rate for both people, giving customers a concrete answer when they recommended it.
Customers naturally mentioned both the benefit and the 25% fee when describing Paribus to friends. Glyman says listening to those explanations showed the founders that a referral should address the hesitation instead of merely asking users to promote the service.
A card designed around customer savings
Glyman describes how financial incentives and the ability to act on spending shaped Ramp. The company grew from cards into a broader financial workflow.
Ramp began with a similar promise of saving time and money. At Capital One, Glyman had seen profitable financial businesses whose incentives did not always align with what customers wanted. He and Atiyeh wanted to build a company where the customer and the provider both benefited from reducing waste.
Traditional card businesses seek more spending while managing the cost of rewards, Glyman says. In customer research, people talked less about points versus cash back than about wanting more money in their accounts. Cancelling an unused $100 subscription, he argues, can matter more than earning 1% back on it.
Glyman explains the distinction with a simple example: eliminating a $100 subscription a company no longer uses can leave far more in its account than a small card reward. He says this made savings a more compelling product goal than maximizing spending volume.
Glyman calls this first a business-model insight: use money generated by the financial product to invest in software that helps customers, rather than competing only by dividing a small margin into larger rebates. Ramp expanded from cards and expense reports into bill payment and other financial workflows.
Paribus could identify potential savings from emails, but it had to ask customer service for a refund. Ramp's founders wanted to be closer to the movement of money itself. Seeing an unwanted subscription and being able to stop a payment are different levels of action.
Garry Tan asks whether selling first to startups helped. Glyman says companies such as Cursor and Cognition started as small Ramp customers and grew. If forced to choose, he prefers building for the future rather than the past.
The next chapter: connect work and help growth
Ramp seeks to join fragmented financial systems and says it is expanding from reducing spending to helping customers use time and money better.
Glyman describes Ramp's first chapter as joining fragmented financial workflows. Companies often have separate systems for cards, expenses, bills, purchasing and travel between their payroll system and their accounting records. He says this fragmented view costs people time and makes waste harder to find.
Glyman says a finance team can end up teaching employees several separate systems just to buy software. An incomplete picture of spending compounds the administrative burden. Ramp's first chapter sought to put those processes into a simpler, connected view.
He says Ramp now measures money and hours saved and reports that the median customer saves 7.5% of expenses annually. Its next chapter, he says, adds tools intended to help customers collect revenue and make spending more productive, including accounts receivable and AI tools. These are the company's own measures and plans.
Glyman asks whether organizations have redesigned themselves for improved AI capabilities, rather than merely adding a chatbot to existing jobs. He sees Ramp's next stage as building tools for a different way of operating, although most organizations are still adapting to the pace of change.
Make AI useful across the company
Glyman describes uneven adoption, a hiring pause and the practical barriers non-engineers face. Integrations allow agents to work with permitted company context.
Tan notes a Ramp statistic that the top 1% of users account for 80% of token spending. Glyman agrees and sees a challenge in spreading useful practices beyond the most active users. Both discuss adoption as an organizational issue, not simply a matter of giving people access to a model.
Changing tools can be difficult even when the task is familiar. Glyman compares asking a finance professional accustomed to Excel on Windows to switch to a Mac with asking someone to change a workflow they have practiced for 20 years. Learning AI, he says, requires time and room to adjust.
Glyman says the transition needs patience. Telling someone to change a familiar routine is a much larger request than changing a device, and an organization has to give staff time to work out what the new tools can do.
Within Ramp, he says, a small group became much more productive while functions adopted AI at different rates. If engineers move much faster than designers or product managers, the ratios and handoffs used by the old organization can become bottlenecks. He describes this as a coordination problem rather than a uniform productivity gain.
The company paused hiring for a quarter even though it was growing, Glyman says. Hiring requires interviewing and onboarding, but he also wanted learning AI to become a core competency rather than a small side project. His aim was to create conditions in which people could discover a use they would not want to give up.
For a non-engineer, installing a command-line tool and connecting it to company data can be a substantial barrier. Glyman compares a basic chat interface to passing notes through a wall: the model may be capable, but it cannot act on information and systems it cannot reach.
Ramp's coding agent Inspect can use the data and code an engineer is permitted to access, along with virtual machines and a browser, Glyman says. He reports that it is involved in more than 85% of the company's pull requests and describes a possible workflow in which it reads a customer bug report, locates a change and responds.
He says Inspect is useful because it is connected to code and to tools an engineer already uses, rather than being an isolated source of text. In the example he gives, a customer complaint can be tied to recent code changes and a proposed correction.
The tool needs integrations and authentication as well as a model. Glyman says Ramp standardizes some internal workflows but lets employees choose other tools, and the company has experimented with products from multiple AI labs. He does not present one company's setup as the right answer for everyone.
Measure AI spending against outcomes
A flat spending cap misses differences in value, he argues. Shared data, permissions and reusable skills are central to Ramp’s Glass system.
On budgets, Glyman objects to judging AI use only by the amount spent. He would welcome a $1,000 expense that produces $10,000 of value, but question the same expense for work with little return. His goal is to connect model spending to outcomes, such as whether a new landing page improves conversion.
Tan asks about Glass, Ramp's shared company knowledge and workflow system. Glyman says it connects core systems, including data warehouses, analytics and customer support, while individual access follows the permissions a person should have. He reports that 1,925 of roughly 1,930 staff used it in the prior 24 hours.
Glass connects systems such as warehouses, analytics and support software to give a company a more complete view. Glyman says employees do not all see everything: access should reflect each person's permissions, which can differ between companies.
Shared skills let a useful workflow discovered by one team become available to others. The system also lets people select models and make cost-versus-latency choices. Glyman gives a background coding example where a process taking 10% longer may be acceptable for a 50% lower price.
Glyman wants the system to preserve not only facts but the reasoning and practices that make the company distinctive. When a person leaves, a skill file can retain some of a useful workflow. Tan asks whether that shared view also helps executives see inconsistent procedures across the organization.
As an example, Glyman describes an expense-policy agent that reads a company's rules and recommends approval, rejection or clarification. He says it achieves 99% accuracy and learns from overrides and incomplete rules. The claim is his description of Ramp's product; the approach depends on permissions and how the organization defines and monitors accuracy.
Reconsider how large companies operate
The speakers discuss silos, coordination and whether AI can help large organizations move more quickly. They consider competing visions for the future of firms.
Tan and Glyman discuss how organizational silos can protect information and slow decisions. Glyman contrasts that with a young company focused on creating value before it has much to protect. As companies grow, coordination costs and reluctance to make hard calls can become serious obstacles.
Glyman believes AI may let some large organizations move faster again. He says Ramp exceeds $2 billion in annual revenue and is growing more than 25% faster than a year earlier. He presents that as the company's experience, while arguing that other firms can also reconsider their structures.
Glyman argues that companies sometimes become attached to the form of the product they sell instead of understanding the underlying purpose. He sees the distinction as useful when old tools no longer match customers' needs.
He uses American Express as an example of a company that changed its products over a long history while maintaining a broader purpose of trust. Glyman's point is that a business should understand the job it does for customers rather than define itself permanently by its first product.
Tan outlines two possible AI futures: a few laboratories capture most value, or agents enable thousands of highly specialized firms. He says these outcomes could differ across industries. Glyman favors the more distributed possibility but agrees that people will have to work to make it happen.
Glyman argues that AI consumption no longer necessarily rises enough to offset each fall in price, a pattern he calls a break in Jevons paradox. He qualifies this as a recent observation that may not last, and suggests society's ability to absorb capabilities may be one limit. He sees lower costs as opening possibilities for new startups.
He describes falling model costs as a chance to make previously uneconomic businesses possible. He also notes that the recent change he sees in AI consumption could be temporary and that new model families might change the pattern again.
Let staff closest to customers change the product
Glyman reports more contributions from non-engineers and a faster release cadence. He also describes plans to offer parts of Ramp’s internal tools to customers.
Tan returns to Glass and asks how internal tools affect Ramp's product. Glyman says non-engineers created 7,000 pull requests in one quarter, many from customer experience staff, compared with effectively none a couple of years earlier. He contrasts this with a bug report passing through a long chain of tickets and priorities.
A conventional support workflow can move from a customer complaint to a ticket, product prioritization, design and engineering, sometimes taking months or never being scheduled. Glyman says AI-assisted changes can shorten that path when the right people have access.
If staff closest to a customer's problem can propose a fix in minutes, Glyman says, the product can improve continuously and people gain more agency. Tan compares this to workers improving a manufacturing process directly. Glyman says expectations rise when colleagues know a change can be made quickly.
Glyman reports that Ramp has recently shipped a major product release every month, compared with one or two in a good year previously. He links that pace to growth and to the company's reported change in median customer savings from 5% to 7.5%; these are his accounts of the company's results, not a controlled comparison.
Tan asks whether companies that maintain records must eventually become AI harnesses. Glyman agrees in Ramp's case: he describes it as a productivity company trying to make customers' money and time go further. He says Ramp is beginning to onboard customers to Glass and expects agent-based interfaces to become a more important way to use its product.
Find evidence of building and ask for help
Glyman looks for demonstrated initiative in hiring. His advice to younger founders is to name problems and seek help rather than pretend to have every answer.
For hiring, Glyman looks for direct evidence that someone enjoys and can build things. A teenager who created a Minecraft server business, or a student who excelled in a relevant contest, may show a skill more concretely than a résumé. He adds that conventional hiring can also find people who have not yet had those opportunities.
His hiring approach is not restricted to one credential. Glyman says someone who has built a product that other people use and pay for has left evidence of initiative, while a contest may offer evidence for a different kind of technical problem solving.
Asked what advice he would give his younger self, Glyman returns to the difficulty of asking for help. People often project confidence even when stuck; founders can instead name the problem, explain their reasoning and ask someone for guidance. He remembers YC office hours around three questions: growth, the biggest problem and what to do about it.
More from this channel
1:00:06EN
1:00:06EN
1:21:42EN
44:15EN
47:14EN
48:45EN