Not Spending $100 Is a Hundred Times Better Than a Dollar Back

In 2015, Eric Glyman and Karim Atiyeh built an email plugin called Paribus. If you bought a TV from Amazon or Best Buy and the price dropped a week later, the store would refund the difference — if you asked. Almost nobody asked. Paribus scanned inboxes for receipts, detected what you had bought, and wrote to customer service as you to claim the refund, taking a cut. Within a year they had almost a million members, and Capital One made an offer they could not refuse.

After the sale they studied credit cards and found a large, profitable industry deeply misaligned with customers. Issuers obsessed over getting people to spend more and earn more points. Glyman says that if you actually listen, people do not want points or cash back; they want more in their bank account. "Not spending $100 in the first place is a hundred times better than getting a dollar or 1% back on it." That math became the seed of Ramp.

They incorporated in March 2019. Five and a half years later, Ramp is chasing "self-driving money": issue cards, make every kind of payment, manage approvals, even automate accounting from one place. Tap a card and the receipt is pulled from email; you do not tag every transaction by hand. About 25,000 businesses now use it, from early startups to Shopify, the Boys and Girls Club of America, and farms. Glyman is CEO; Atiyeh is CTO, and he also oversees marketing and other functions they see becoming more technical.

The shift from consumer to business was almost accidental. In the ideation phase they talked to friends, many of whom were starting or joining early-stage companies, and heard larger versions of the same problems they had solved for shoppers. Atiyeh noticed something odd: the better and larger companies got, the more wasteful they became, and the less they knew about their spend. That made the problem more interesting, not less. Waste scaled with success, so the best customers were the biggest opportunity.

They also saw an experience gap. Consumer products like Instagram and Robinhood obsess over every interaction. The same people then show up at work and are asked to use tools designed in the 1980s — slow and clunky. Nobody had brought consumer-grade UI obsession to complicated business problems. That gap was the opening.

The first hook was cultural as much as financial. Most of the credit card ecosystem was designed around ego and excess: lounge access, points, exotic trips. Glyman contrasts that with what Amex sold in the 1980s — mahogany, airport lounges — versus luxury in the 2020s: a yoga class at 2 p.m. because you own your time. People were missing time, so Ramp tried to save money and hours from day one.

Early products were blunt. Promising startups would share a card statement, and Ramp would find four project-management tools or five BI products doing the same job, because someone tried a tool, passed the card along, and finance only asked for receipts. They would cut about 2% of spend a year, then suggest using Ramp's card so the company would keep doing that. The mechanism was cutting redundant spend and automating the tedious — expense reports, transaction tagging — to give people Friday hours back with their families.

Atiyeh says the early thesis was that saving time and money requires context: card statements, inbox receipts, even the ERP. The data is unstructured; the job is to structure it and automate workflows. Internally, AI shows up as a high-level algorithm with constraints, applied to repetitive tasks so Ramp does not have to custom-build every company. Closing the books, paying bills on the last day they are due or earlier for a discount, and three-way matching for fraud all have a right answer that can be evaluated.

Ramp also used AI early to augment sales, marketing, and underwriting. Years ago, with few resources, one sales rep booked far more meetings than anyone else, so a few engineers sat with him instead of hiring two more copies of him. He had the right mechanism — scan who raised, who changed jobs, guess emails, try copy — but every step was manual. "Let's give him an Iron Man suit," they decided: signals pulled automatically, lead list done, emails found, copy A/B tested. Flash forward, and Ramp SDRs book multiples of the meetings of their next-closest competitors each month.

Sarah Guo asks why most companies cannot do this: engineers are rarely interested in tooling for SDRs. Glyman says Ramp engineers are measured against business problems and a P&L — unlike Capital One, where technology was a cost center with only an L, no profit, and meetings were about ripping cost out. They hire for spikes and for entrepreneurial, somewhat contrarian people. His job, as he sees it, is to recruit raw ambitious talent and keep the company interesting enough to hold them.

The other framing is that time is money. In the salesperson example they were not asking how many dollars they were spending, but where the hours were going and what could be automated. Glyman argues AI's best use case today is productivity: not where cost lives and how to rip it out, but how to make every hour go further. Elad Gil adds that early Google built internal tooling for online sales and operations so that by the time the company had 10,000 people, that function would not have required another 50,000 or 100,000.

Elad asks about build versus buy, and whether Ramp has considered spinning internal tools out as products. Atiyeh's twist: companies can usually tell if they built well, but almost never if they bought well. People get praised for hiring; almost nobody is evaluated on picking vendors. Ramp treated procurement as a skill early, interviewing vendors the way you interview candidates, talking mainly to engineering teams, and caring about slope — the rate of progress — rather than today's feature gaps.

Why not sell the SDR automation and marketing stack to customers? Glyman says the simple answer is simplicity of sale. Everything Ramp ships is centered on saving money and time for finance: expenses and cards, then bill pay, procurement, and travel, always for the same economic buyer. Selling to different buyer groups inside a company adds a lot of go-to-market friction. He admits they may not have thought deeply enough about productizing the internal tools. He also flips the old line: nobody gets fired for hiring IBM — at Ramp, you might get fired for picking IBM.

Guo observes that selling productivity is hard because many organizations do not value it the way they value the top or bottom line. Glyman is fine being called a fintech, but he thinks Ramp is a productivity company selling time to finance teams whose CEOs demand G&A fall every year with no new headcount. They refuse to lead with AI — he compares 2024 to the late 1990s, when ".com" in a name juiced a stock until Pets.com made it a liability — and instead show dollars and hours saved. Atiyeh says they are starting to see statistical significance: underwriting and bankruptcy risk look lower for Ramp customers than any cohort they have seen, partly selection, partly, they hope, the product.

Traditional banks tell Guo that AI is too unreliable for the workflows they care about, and that regulators and compliance make it worse. Atiyeh does not think there is that much risk if you constrain the problem. Categorizing a transaction is not "make up categories"; it is "which of these known categories is it?" Every form can be pre-filled and editable, every list pre-ordered, and the product can ask only the question that is actually missing — was this lunch with a candidate or a customer — because the calendar may already answer the rest.

He splits risk into known and unknown. Known risks you can fix; unknown risks are what blow up companies. Ramp's use cases sit in a sweet spot that shrinks the unknown. Guo's summary: the ability to define tasks that make sense and measure performance against them internally is a strong competitive advantage in applying AI today. Translation from accounting-speak to business-speak is a problem they keep solving; they do not lean on models for the kind of basic math that still breaks.

They spent time fine-tuning and quickly decided their time was worth more waiting for the next generation of models. What they hardened instead is infrastructure: switch models easily, evaluate them on the subtasks that matter. When an investor asked what GPT-4o mini's price cut changed, Atiyeh's answer was that it was already in production and good enough for about 90% of tasks — a one-day switch, because the evals already existed.

What is still missing? End-to-end navigation of Ramp's own site or app is a cool demo and a hard production problem, because the product changes too fast. Atiyeh is iterating on something closer to CI for outbound email: every first-time email runs through an AI review for brand guidelines and a clear CTA, as instant and deterministic as they can make it. He wishes models had better knobs than ever-longer prompts. He singles out Claude's artifacts as a different interface: a mini web app you can tweak, which is how they are thinking about the right surfaces for their own tools.

It is unusual for a CTO to own marketing. Atiyeh says the thing that will remain for a very long time is taste: someone still has to decide if the work is good. His obsession is giving marketers time to focus on that differentiator and taking the mundane, repetitive process off their plates. He looks at marketing as a system with bottlenecks, parallelism, and dependencies, then builds systems to remove the bottlenecks — closer to refactoring an API than sprinkling on AI.

Ramp sits on 23rd Street in New York; six blocks south is Andy Warhol's old studio, the Factory. Before Warhol, painting a Mona Lisa meant decades of craft; he printed on silk screen with commercial methods and obsessed only over what was striking. People joked about "art by telephone," where it was unclear who had made or signed a piece — it did not matter, they were all Warhols. He abstracted production so he could focus on the striking, and decades later the work is still recognizably his. Glyman uses that as a lens on marketing's future: you can drive production complexity toward zero, but you cannot replace taste.

The marketing org's structure is roughly the same; the work is leverage and tooling so the team can match product's shipping speed. Many companies, faced with that pressure, slow to quarterly or yearly releases for more control. Ramp would rather keep weekly speed and give teams better tools. Glyman notes that a good idea can now become tens or hundreds of thousands of API calls — a hundred-person team in spirit — and the human mind does not work that way. Systems thinking and the ability to build and pick tools is becoming an interview bar for every role.

Gil recounts a Stitch Fix marketing lead who skipped a year-long agency cycle and handed video tools to hundreds of stylists already in the building: not Coke-ad quality, but a thousand times more effective and essentially free. In commercial creative work, he argues, you care about outcome, so people with taste and direct manipulation beat a perfect pipeline. Ramp inverts the usual order — teach the system the brand, let anyone draft video or social, start from what works, then iterate toward on-brand — because starting from on-brand and then trying to make it work is slower and more expensive. Atiyeh stresses they will not buy control by slowing weekly shipping to a quarterly cadence.

A Wall Street Journal story a year or so earlier had treated a million fewer U.S. bookkeepers over a decade as a crisis: wages stagnant, financials at risk. A few months later someone asked how many accountants and financial advisors there were; that number was up by about a million. Jobs were not vanishing so much as changing shape. Glyman wants automation pointed at the tedious and monotonous. Nobody should be the receipt-chaser who cannot have a normal conversation at the water cooler; that is a system job, which is where Ramp should sit.

Internal research on "strategic finance" found that only about 4% to 9% of finance jobs qualify, depending on how you measure. That implies 91% to 96% is non-strategic, and even people in strategic finance tell him the truly strategic slice of their own week is small. Finance is unusual in that people actually want automation: G&A has to fall, and they would rather decide where to invest than chase receipts. Salespeople should be selling, not looking up email addresses.

The catch is plumbing. Glyman had just met a CFO of a company with more than $10 billion in revenue and hundreds of finance tools for books, payables, and receivables — "Byzantine," in the CFO's word. Fewer people ask how to build the pipes and primitives: card payments, bill pay, accounting, the low-level operational tasks. Overlay intelligence on a hundred disconnected systems and you get insights you still have to execute by hand. Overlay it on real primitives and the vendor gets turned off, the contract gets renegotiated, and the work is orchestrated.

One internal experiment made that concrete. Ramp generates tens of thousands of hours of customer conversations that no team can listen to, so the applied AI group built a pipeline that produces a five-minute podcast — Glyman or Atiyeh asking questions about the week's most interesting customer moments, with plans to let anyone drill into a segment, persona, or topic. They used sentiment analysis to find the happiest calls and sent a "voice of the customer" episode to other founders. Every founder asked for the other cut: the voice of the angry customer, because as companies get bigger, nobody wants to tell you bad news. LLMs have time to listen; people do not.