The world is too loud. Read what matters.

Y Combinator

Ramp Paused Hiring for a Quarter to Force AI Mastery

Ramp discovered engineers' productivity increased 100x with AI, but salespeople and designers couldn't keep up and would drag down the organization, so they froze hiring for an entire quarter and forced everyone to make AI a core competency rather than an ancillary skill.

Enterprise softwareAI agentOrganizational transformationProductivity toolsFintechYC

The video won't play here. Listen to the audio instead:

Ramp's co-CEO explains how a $6 billion company embeds AI into financial workflows, internal tools, and organizational structure—not abstract concepts, but specific mechanisms and numbers that make this worth listening to in full.

The argument · tap a timestamp to hear it

22:01

1% of users consumed 80% of tokens, structural friction emerged

A few engineers saw productivity surge to its extreme with AI—not exaggerating, 100x—while designers, product managers and other roles couldn't keep up. This destroyed the traditional ratio of one product manager and one designer per six to eight engineers. Engineers began outpacing the rest of the company, creating structural friction within the organization.

— Eric
23:05

Pausing hiring for a quarter forces teams to truly master AI

Eric noticed a counterintuitive correlation: the more people used AI, the less they needed to hire; the less people used AI, the greater the hiring need. So earlier this year, while business was growing over 100%, Ramp decided not to hire for an entire quarter. The reason wasn't just that hiring is time-consuming—more importantly, they didn't want AI learning to consume only 5% of employees' time, but to make it a core competency.

— Eric
24:09

The real barrier for large companies to adopt AI is terminal access

Small companies can get everyone on AI easily, but at big companies when sales wants to use Claude Code, they first need to open a terminal, install packages, then request access to Databricks, Snowflake, Slack, Stripe—and often get stuck right there. Eric believes most people experience AI like ‘passing notes through a wall’ rather than the way top engineers do it: AI connected directly to systems with permissions, integrated into the team.

— Eric
32:45

AI expense approval makes purely ceremonial gatekeeper positions redundant

Companies must have someone review every expense for compliance reasons—a requirement stemming from Sarbanes-Oxley—but most approvers actually don't care about policies and either rubber-stamp everything or go through motions, often getting it wrong. Ramp built a policy agent that can automatically approve, reject, or flag incomplete policies with 99% accuracy, and every manual override teaches the system, even retroactively refining the policies themselves.

— Eric
40:06

Coase's theorem debate: AI converges to one firm or enables thousands

There are two extreme views on how AI reshapes firm size. One is ‘doomsday’: some lab wins everything, the optimal number of companies approaches one. The other: AI enables hyper-specialization, letting thousands of companies each become billion-dollar champions in narrow verticals. Eric explicitly backs the second view and says Ramp is ‘betting’ on it.

— Eric / Gary
41:06

Jevons paradox broke for the first time in the past two months

When intelligence got 10% cheaper, usage would grow faster than price fell—like how cheaper railways increased transport demand. This held for AI too. But Eric directly points out it broke in the past two months: model capabilities keep climbing, but society's capacity to digest that capability has hit a ceiling. It may be a local maximum, not necessarily permanent.

— Eric
44:33

Non-engineer pull requests jumped from zero to 7,000 in a single quarter

Ramp had 7,000 pull requests created by non-engineers in Q2, while two years ago the number was basically zero—no one would let people who hadn't learned git submit code. Most came from the customer experience team. They used to file bugs with product managers for scheduling, wait on designers, wait on engineers, and get fixes two months later. Now they can fix it themselves and ship the same day.

— Eric
48:22

System-of-record companies that don't become AI harnesses will die

Gary brought up a slightly controversial tweet he'd posted: ‘Either die as a system-of-record company or live long enough to become an AI harness.’ Eric explicitly agreed and said Ramp doesn't consider itself fintech but a productivity company, where the core question is always ‘How do you make your money and time go further?’

— Gary / Eric

In their own words · checked verbatim

there's this app called Paribus. It's really great. It saves people money. Whenever you buy things and it goes on sale, you should know though, it keeps 25% of what it, of what it saves you.

Eric7:23

the future is here. It is not evenly distributed. That is for sure.

Eric18:58

If someone spends $1,000 to make $10,000, I'm trying to throw money at them. If someone's spending $1,000 to do their weekly prep for no value, I don't know.

Eric28:27

over the last two months, Jevin's paradox broke.

Eric42:10

either you die a system of record company or you live long enough to become an AI harness.

Gary48:22

because that's where the money is

Eric50:26

I think one of the most powerful words you can say is a founder, CEO or just person is, is actually like, I need your help. I'm trying to figure this out. Here's the problem I have.

Eric52:37

Figures

Ramp current valuation$60 billion14:43
Ramp annualized revenueOver $2 billion per year37:02
Ramp revenue growth rate (year-over-year)Over 25%37:02
Median customer annual cost savings7.5%15:45
U.S. average corporate profit margin8%47:18
PRs generated by InspectOver 85%24:09
Ramp engineers' monthly PR growth~40% per month24:09
MCP usage weekly growth~10% per week49:22
Non-engineer pull requests in Q27,00044:33

Glossary

Jevons paradox
When resources become cheaper, total consumption increases faster than the price falls.
Coase's theorem
Firm size is determined by the balance between internal coordination costs and external transaction costs.
AI harness
Middleware that connects AI models to real system permissions and workflows, enabling them to execute tasks.
MCP
Model Context Protocol—a standard for connecting AI models to external systems and data.

How to listen

Who it's for

Founders rebuilding organizational workflows with AI; investors following enterprise software and agentic products; engineering and product leaders who want to understand the real friction in deploying AI at scale.

Skip

0:00-8:00: Paribus founding story; skip if you care only about Ramp and AI organizational transformation.