Key Takeaways

  • Dumping raw feedback into large language models fails at scale: a one-million-token context window holds less than 0.5% of Ramp's Gong call transcripts.
  • Ramp started small by creating a dedicated Slack "hate channel" and an automated daily podcast compiling angry user feedback.
  • To solve data silos, Ramp built an internal customer insights agent combining traditional ETL pipelines, vector search, and contextual clustering.
  • AI synthesis does not replace customer interviews; it pinpoints the exact users product managers need to call.

The Method

Every growing company drowns in fragmented user feedback. Customer complaints scatter across Gong recordings, Zendesk tickets, LogRocket session replays, CSAT surveys, and direct emails to executives. When teams try to fix this with AI, they usually dump raw text into a prompt and hope for clarity.

That strategy breaks quickly. As Geoff Charles points out, “LMS are great, but a 1 million token window, that's less than 0.5% of gong transcripts at ramp. So, we started small. We created a hate channel and it every day it posts all the lovely things that our customers say.”

From that Slack experiment, Charles tested audio formats to make user frustration impossible to ignore: “My favorite one was the podcast, the hate podcast. Great way to start your day. You listen in 100 customers yelling at you about all the ways your product is broken.”

Raw feeds create emotional awareness, but they do not help product teams prioritize engineering work. To turn unstructured complaints into clear direction, Ramp built a dedicated customer insights agent. Here is how they structured it:

1. Centralize data ingestion: Build automated ETL pipelines that ingest every customer touchpoint, including sales calls, support tickets, user session logs, and executive escalations.

2. Apply vector search: Index unstructured feedback text across all sources into a unified vector database.

3. Cluster by context: Group related complaints around shared product features, workflows, and user personas rather than simple keyword matches.

4. Trace back to specific accounts: Link every clustered insight directly to the original customer record, contract size, and user profile.

Charles explains the architecture plainly: “We built an actual customers insight agent. That agent pulls from all the data sources at a company. It uses traditional ETL and pipelines. It uses actual vector search. It has clustering around the same context.”

Where This Breaks Down

The biggest failure mode with AI-generated insights is passive consumption. When product managers receive clean summaries, they assume they understand the customer without doing the hard work of live discovery.

Automated summaries strip away tone, hesitation, and workflow quirks. If a team relies entirely on clustered tickets, they build features for symptoms rather than root causes. Charles warns against this shortcut: “Get people access to this data as fast as possible. That doesn't mean don't talk to customers. It means you actually identify exactly which customers to talk to because the data is traceable.”

The pipeline exists to generate qualified interview leads, not to replace the interview itself.

What to Do With This

Pick your single largest feedback silo this week (whether that is Zendesk tickets or Gong recordings) and export the last 500 entries. Group them by specific product friction rather than broad categories, identify the five customers experiencing the highest severity issue, and book discovery calls with three of them by Friday.