Inside Ramp, the $32B Company Where AI Agents Run Everything | Geoff Charles

Overview

Core thesis: AI agents are the primary accelerant across discovery, analytics, spec-writing, coding, QA, releases and enablement. Ramp has re-architected processes and org design to make AI ubiquitous and to scale human impact.

Ramp (CPO Jeff, interviewed by host Peter) operates as an AI-native, high-velocity product organization. Key characteristics:

• ~50,000+ customers and >1,000,000 end users.
• ~25 product managers and a focus on rapid delivery.
• Shipped 500+ features in the last year and reached >$1B in revenue.

Key metrics & targets

• Revenue: >$1B ARR (last year).
• Customers / users: 50,000+ customers; >1,000,000 end users.
• Feature velocity: 500+ features shipped in one year.
• PM headcount: ~25 PMs supporting the above output.
• AI-generated code adoption:
  • 30% (Dec) → 50% (current) → projected ~80% by March (near-term target).
  • Potential 90–100% longer-term.
• Beta penetration: ~10% of customers opt into beta.
• Time-savings examples:
  • Voice-of-customer analysis (90 days of tickets/chats): ~8 minutes (vs days manually).
  • Data Q&A responses: ~2 minutes.
  • Prototyping / feature built (example): ~5 minutes.
• PR process: a “double-digit percentage” of PRs automatically approved via automated review.
• Planning horizon: practical and reliable planning for ~3 months.

Frameworks, processes & playbooks

AI Proficiency Ladder (L0–L3)

• L0: occasional ChatGPT users (likely low retention).
• L1: custom GPTs / basic agents / Notion agents.
• L2: built apps/automations that materially augment work.
• L3: systems builders who create skills/agents that scale.
• Goal: migrate everyone upward; expect L0 attrition.

Voice-of-the-Customer agent

• Ingests Gong, Salesforce notes, support tickets, in-app surveys, email, and Snowflake analytics.
• Outputs synthesized themes, representative quotes, links to sessions, recommended outreach, and draft outreach emails.

Ramp Research → Snowflake CLI + cloud code + skills

• Natural-language data analysis across company schemas.
• Auto-generates results, interpretations, full HTML reports, and experimental growth ideas.

“Inspect” code-generation workflow

• Conversational spec → agent plans → uses codebase context and design component library → produces full PR (frontend + backend as needed) → automatic PR review + routing.

Release automation (Ramp Releases)

• Automated release report: product preview, analytics impact (Snowflake), Slack summaries, help-center content, enablement content.
• Routes releases for staged rollout (alpha → beta → GA).

Staged rollout taxonomy

• Doc-fooding / alpha → beta (10% opt-in) → monitored GA with automated impact metrics.

No-committee, no-signoff principle

• Empirical gating: prove value in a small release or beta; escalate only by complexity/impact.

Hiring & interview requirement

• New hires must demonstrate AI proficiency; product candidates must present an actual prototype built with tools.

Token & usage governance

• Open access to tools (no budget/token gatekeeping).
• Internal tracking of token usage per employee and across tools.

Concrete examples / case studies

• Voice-of-customer demo: analyzed procurement feedback by scanning 90 days of tickets/chats and surfaced prioritized themes (purchase order management, approval routing, exports, currency constraints) plus links in ~8 minutes.
• Ramp Research demo: answered metric questions (e.g., open rate of automated emails) by querying database schemas and returning interpreted results in ~2 minutes.
• Inspect demo: PM asked the agent to build an accounts-payable metrics report (overdue, upcoming, buckets). The agent inspected the codebase, used the component library, created front-end and back-end code, opened a PR — production-ready in ~5 minutes.
• Automated triage: for escalations or confusing support tickets, AI creates a diagnosis and a ready PR for remediation for PM/engineer review and ship.
• Internal adoption showcases: finance built treasury automation; legal used agents for contract review; marketing automated website creation.

Organizational & operational tactics (actionable)

• Open access to AI tooling (multiple LLMs, cloud code, tokens); remove budget friction to accelerate discovery and track usage to identify adopters and laggards.
• Build an internal skills repository and reusable “domain skills” that codify best practices and company context (Notion + skills + design components).
• Instrument and synthesize customer signals (calls, tickets, surveys, analytics) into a searchable voice-of-the-customer agent to prioritize roadmap and outreach.
• Automate routine spec-to-code cycles: enable PMs/designers/operators to produce PRs and reserve engineers for complex/scale tasks and agent-management.
• Automate release artifacts: product previews, analytics impact, help-center content, internal enablement, and Slack communications generated by bots.
• Implement staged rollouts with automated monitoring and gating (alpha → beta → GA) to maintain quality and control risk.
• Use automated complexity-routing to send high-complexity changes to senior engineers/product directors for deeper review.
• Require interview assessments demonstrating tool fluency (build a prototype with agents/cloud code).
• Create public internal channels, office hours, and designated experts to evangelize adoption and provide hands-on help.
• Leaders should focus on fixing broken processes (identify which prompt/skill/design/system failed) rather than applying one-off fixes.

Product management, engineering & talent implications

Product management

• Role shift: from spec-writing to product-building and systems thinking.
• Two PM career tracks:
  1. Builders: expert at rapid iteration using AI tools.
  2. GM/business-focused PMs: own positioning, distribution, monetization, and long-term strategy.
• PMs must reserve focused IC time to build and learn; reduce meetings and committees.

Engineering

• Engineers may transition to building and managing many agents, focusing on complex system design and agent orchestration rather than routine coding.

Management & career advice

• Managers should spend more time in IC mode to re-skill and demonstrate workflows; reduce meeting load.
• Career advice: prioritize becoming an excellent builder and agent-user. Management is a less valuable short-term track than tooling/build skill.
• Learn to bake domain expertise into agents (accounting/CPA rules, domain-specific workflows), not just UI.

Quality control & governance

• Maintain quality through:
  • Complexity thresholds that automatically route code for human review.
  • Staged rollouts and monitoring analytics (Snowflake) to catch regressions.
  • Automated PR review and a release bot that compiles impact evidence and documentation before GA.
• Leaders should audit process failures (which skill/prompt/design component broke) and fix underlying systems rather than repeatedly giving the same feedback.

Costs & ROI stance

• Ramp tolerates token costs as an investment in discovery; token spend per employee is small relative to salary and the ROI from amplified output.
• Argument: aggressive investment now to capture competitive advantage before AI functionality becomes fully commoditized.

Actionable recommendations (quick checklist)

• Start a voice-of-customer pipeline: centralize and index calls, tickets, CRM notes and analytics; attach an LLM-based agent for synthesis and follow-up.
• Build an internal skills marketplace (reusable agent prompts/skills + design components) and require reuse.
• Open access to cloud-code/agent tooling and track adoption metrics.
• Add automated release documentation generation into your release pipeline.
• Shorten planning cycles to ~3 months; focus strategy discussions on trade-offs and customer segments.
• Require AI-tool fluency in hiring interviews (e.g., product candidates produce a prototype).
• Implement staged rollouts and complexity-based routing of PRs.
• Encourage leaders and PMs to spend dedicated IC time to learn and build with agents.

Risks & cultural notes

• Expect attrition among non-adopters (L0); cultural change is required: fewer committees, fewer signoffs, and strong emphasis on individual initiative and learning.
• Quality and collaboration risks can be mitigated via automated gating, staged rollouts, and code-review processes; humans remain required for complex decisions.
• Domain expertise becomes increasingly valuable: agents are powerful only when fed accurate, codified domain knowledge.

Presenters / sources

• Host: Peter (interviewer)
• Guest: Jeff (CPO of Ramp)

(Note: examples and timing are drawn from the Ramp CPO’s demos and remarks in the interview.)