Minimizing churn before it happens with AI

Retention is one of the most important levers in any recurring revenue business. As a go-to-market leader, I've long believed that deeply understanding why customers leave is as strategically important as understanding why they buy. If you can identify the warning signs early enough, you can act on them. If you can spot patterns across hundreds of accounts, you can change the conditions that create them in the first place.

The problem is that the answer to "why did this customer churn?" is almost never in one place. Some details are in the Gong call, where a champion first mentioned a budget freeze. Others are in the Zendesk tickets that go unanswered for too long. More clues are in the billing page behavior, where activity spikes three weeks before the renewal conversation. Yet more clues are in the Salesforce notes from a quarterly business review where the tone shifted. 

To piece together the full story, you have to pore through and make sense of data from each of those places. To understand trends and glean lessons from them, you need to analyze vast amounts of data, at scale.

As a sales VP, this was exactly the kind of problem I considered critical to solve: not just understanding churn after the fact, but giving our teams the context to intervene earlier, learn faster, and improve retention at scale.

[CTA_MODULE]

Analyzing churn by hand is valuable (but too slow)

Before we built anything, we had a dedicated team member tasked with analyzing scale manually — pulling data from hundreds of systems and synthesizing it into a coherent narrative for individual accounts.

That process built an invaluable baseline of knowledge: which data sources were actually predictive, which signals arrived early enough to act on, and which patterns showed up repeatedly across customers who left for similar reasons. The manual work was slow and expensive, but it was also deeply educational. The person doing it had spent years developing a sense of what mattered and what didn't.

As our customer base grew, I knew we needed to preserve the judgment and expertise our team had developed while removing the manual bottleneck. The goal became to encode that expertise into a repeatable system that could help every GTM team make better retention decisions, faster.

Why you need to combine data from multiple sources

Gong data is one of the richest signals available, because it captures not just what customers did but what they said and how they said it. We initially tried Gong's native AI search, but the results were unreliable, often misrepresenting a call or failing to answer the specific questions we needed answered.

This is because Gong did not contain all of our business context. Our business spans multiple data sources, each containing invaluable context for understanding and identifying churn patterns. AI embedded in a point solution is optimized for that solution's use case, not yours. 

By contrast, combining a Gong call transcript with Salesforce notes, Zendesk support ticket history, and product usage data yields a picture no single source could provide. The agent can see that a customer said they were happy in their last QBR, but billing page views doubled the following month, and they filed two tickets about pricing in the same window. That full story is only visible when the data is unified.

Fivetran turns fragmented data into centralized context for AI

Fivetran is what made the context available, reliable, and reusable. Using Fivetran, we centralized our customer data into BigQuery, creating a reliable foundation where signals from Gong, Salesforce, Zendesk, product usage, billing behavior, and ARR trends could be analyzed together. Centralization made the AI useful — once the data was available in one place, we could build an AI Skill in Claude on top of a complete view of the customer, rather than forcing the model to reason from isolated fragments. We turned information into decisions using our existing data centralization infrastructure.

For each account it analyzes, the agent pulls and synthesizes data across multiple sources: ARR trends, Gong call transcripts from the past 12 months, Salesforce interaction history, Zendesk support tickets, and in-product usage behavior, including which pages customers visited, how many active users they had, and whether their engagement with billing-related pages changed in the period before churn. Analysts used to check all of these manually — the agent now checks programmatically.

This fundamentally changes our operating model. Instead of waiting for a post-churn review or relying on anecdotal signals from individual account teams, we can surface account-level narratives as needed, compare patterns across customers, and give account teams a clearer view of where intervention is needed. That helps us move from reactive churn analysis to proactive retention management.

Business context makes or breaks AI

This project reinforced something we've come to believe strongly about enterprise AI: the model is rarely the limiting factor. The real challenge is giving AI access to the right business context: what constitutes a regrettable or non-regrettable churn; the difference between a dev and prod connection; what churn looks like.

For us, Fivetran provided that foundation by bringing relevant data together in a governed, queryable way. Our churn analysis agent is effective not because of the model it uses, but because it combines holistic data with years of human expertise encoded into its prompts, logic, and guardrails.

Every successful agentic system depends on those two ingredients: comprehensive data and business context. Without them, AI produces plausible but unreliable answers; with them, it can scale expert judgment.

For any revenue leader thinking about how to apply AI to retention, my advice is always to start by building a solid data foundation for AI.

[CTA_MODULE]