How Ramp Ships Production Features in 5 Minutes#
Published on March 22, 2026
Ramp's CPO Geoff Charles opened his real tools on camera and showed how a $32B corporate card and spend management platform runs on AI agents. No keynote. No pitch deck. Live demos, real data, real PRs. Here is what matters for your engineering org.
Before diving in, a note on the source material: Peter Yang's interview with Geoff (article) is one of those rare pieces of content every engineering leader and PM must watch. Peter has a knack for teasing out the operational details that 99% of interviewers overlook in their pursuit of high-level insights, and Geoff genuinely walking through actual tools, real data, and real workflows on camera makes this a uniquely accessible resource. Accept what you see here as the new baseline -- companies like Ramp are already operating this way.
The 5-Minute Feature That Changes the Calculus#
During the interview, Geoff typed a one-line spec into Ramp's internal tool Inspect -- "build a report with overdue payments, upcoming bills, and total outstanding" -- and five minutes later had production code. Back-end, front-end, real component library, PR ready for review [1].
Teams using Claude Code and similar tools against well-structured codebases reproduce this daily. If your company hasn't arrived at that point yet, the blockers are predictable: no codebase and organizational knowledge accessible to the AI (context files, design system docs, or access to a knowledge system), messy deployment pipelines, no automated PR review process with escalation paths for critical changes, too many manual gates. Fix those, and the 5-minute feature becomes table stakes.
Ramp now has PMs, designers, operators, and account managers shipping production PRs. Most auto-approved changes are quality-of-life fixes -- copy, naming, small UX tweaks -- the kind of improvements that sat in backlogs for months. Big changes and critical areas still get full human review with the same rigor as before [1].
AI Agents as Operational Infrastructure#
Ramp built two internal agents that changed how the company works [1]:
Voice of Customer Agent -- lives in Slack, synthesizes 90 days of Gong recordings, support tickets, Salesforce notes, and in-app chats in eight minutes. PMs then go conversational: "go deeper on this pain point, bring customer quotes, draft an outreach email." Eight days of research collapsed to eight minutes.
Analyst Agent -- lets anyone query Snowflake in plain English. The unexpected result: lowering the barrier to data access made the entire company more data-driven. Sales, marketing, and support teams started asking questions they never would have routed through an analyst.
This echoes the pattern described in the agentic engineering playbook -- agents become first-class teammates, not just tools.
50% AI-Written Code, Heading to 90%#
Up from 30% in December 2025, Ramp's codebase is now 50% AI-generated. Geoff aims for 80-90% and calls it "coding escape velocity" [1]. Everything now centers on whether your codebase is designed for AI to write code well -- clean abstractions, strong component libraries, and comprehensive context files accessible to AI.
Encode Feedback Into Systems, Not Slack Messages#
One of Geoff's most tactical points: he told his team to put the CTA above the fold a hundred times. Then he realized the real fix was encoding it into the design system and automated review process. AI now enforces it before anything hits his desk [1].
Every piece of repeated feedback should become a prompt rule, a skill instruction, or a design system constraint. The same principle drives structured task workflows -- encode your standards into the system so neither humans nor agents keep making the same mistakes.
Top 3 Lessons for Every Engineering Org#
1. Structure your codebase for AI, not just for humans. Component libraries, context files, clear abstractions, and well-documented conventions are the difference between a 5-minute feature and a 5-day feature. Your codebase knowledge needs to be machine-accessible.
2. Encode repeated feedback into automated systems. Every time you catch the same mistake twice, you have a system design problem. Turn style guidance, review criteria, and domain rules into prompts, design system constraints, or automated checks with clear escalation paths.
3. Lower the barrier to data and tooling for everyone. Ramp's biggest surprise was that giving non-technical teams access to data through natural language queries transformed company culture. The compounding effect of everyone asking better questions dwarfs the cost of building the agent.
What This Means for Your Org#
Ramp plans only 3 months out. Their PM role is forking into builder-PMs who ship code and strategy-PMs who think like GMs. They track token usage per employee -- to find who needs help, not to cut costs. Geoff's framing on spend: "Token consumption per employee is not even close to double digits. If agents can do 10x more work, why wouldn't you pay them twice as much?" [1]
Your stack, your processes, and your culture either enable this pace or block it. The companies that figured this out are already operating at a fundamentally different speed.