Skip to content

Ramp VoC Agent Overview

How Ramp built a composite AI system that turns raw customer signals into actionable product and sales intelligence.

What It Is

  • Not a single tool — a composite system spanning multiple AI agents, data platforms, and third-party tools
  • Demoed publicly by Geoff Charles (CPO) — described as achieving "8 days of research in 8 minutes"
  • Answers strategic questions: Why did customers churn? Why are they angry in support? Why did they pick a competitor?

Three Core Internal Agents

AgentPurpose
Voice of CustomerSynthesizes feedback across Gong, Zendesk, Salesforce into insights
AnalystNatural language data queries ("get insights using English")
InspectInternal coding agent (30% of PRs — separate from VoC)

The Stack

  • LLM: Anthropic Claude (primary) via Claude Code and Snowflake Cortex
  • Data Platform: Snowflake + dbt + Hightouch CDP — "low petabytes" of unstructured data
  • Agent Orchestration: LangChain connecting to Salesforce, Gong, BigQuery, Apollo, Gmail, LinkedIn, Exa
  • Sales Intelligence: Actively.ai (Gong + Salesforce + third-party signals)
  • Revenue Ops: Momentum.io (auto-populates CRM from call transcripts)
  • Customer Support: Sierra AI (knowledge management feedback loops)

Impact

  • 500+ features shipped in 2025 with only 25 PMs
  • 80%+ of sales workflows run through the internal platform
  • 50% of code written by AI (targeting 80%)
  • Every function required to onboard and work with agents

Core Philosophy

"The moat is proprietary data. Not the model. Not the prompt." — Parth Gujare, Senior PM at Ramp

The AI flywheel: more customers using Ramp -> more spending data -> better insights -> more customers. AI commoditizes data interpretation, not data access.

Key Takeaways

  • VoC is a system-of-systems, not a single agent — it spans data infrastructure, agent orchestration, and multiple SaaS integrations
  • Data infrastructure (Snowflake CDP) must exist before agents can be useful
  • Ramp embeds VoC directly into PM and sales workflows rather than building standalone dashboards
  • Small teams win — 2-4 person GTM engineering teams with single ownership outperform large projects
  • Focus on automation metrics (tasks completed), not engagement metrics (sessions)

Sources