Back to blog
Customer Success

How Our CS Team Cut Churn by 18% Using AI to Spot At-Risk Accounts Earlier

We weren't missing the signals — we were drowning in them. AI didn't replace our CSMs' instincts; it gave them the bandwidth to actually act on what they already knew.

LV

Lena Vasquez

Head of Customer Success, HubSpot

March 17, 2026 6 min read

Most CS teams aren't missing signals — they're drowning in them. CRM activity, product usage metrics, support ticket volume, NPS scores, billing changes, stakeholder turnover — any one of these can indicate churn risk. The problem is that no individual signal is reliable enough on its own, and most CSMs don't have the bandwidth to monitor all of them for all accounts simultaneously.

What we built

We built a health score model that ingests data from our CRM, product analytics platform, and support system and runs it through an AI model trained to identify early churn indicators. The model doesn't replace CSM judgment — it surfaces the accounts that need attention so CSMs can apply that judgment where it matters most.

What the AI found that humans were missing

The most predictive signals weren't the obvious ones. NPS scores and support ticket volume were lagging indicators — by the time those moved, churn risk was already high. The leading indicators were more subtle: declining feature adoption in month two of a contract, a reduction in active users relative to licences purchased, and stakeholder email response latency increasing. The AI identified these patterns across hundreds of accounts simultaneously. A CSM covering forty accounts couldn't do the same.

The results

In the twelve months after deployment, we reduced churn by eighteen percent. The reduction was concentrated in medium-risk accounts. High-risk accounts — where churn was often already decided — didn't move as much. That finding reshaped our intervention strategy: AI-assisted early warning for medium-risk accounts is where the leverage actually lives.

What to consider before building your own

  • Start with the signals you already have reliable data on — don't try to build a perfect model; build a useful one
  • Involve your CSMs in defining the signals — they know what actually precedes churn in your customer base better than any model built from scratch
  • Build for triage, not prediction — the value isn't in a churn probability score; it's in a prioritised list of accounts to call this week

The best AI in customer success doesn't make CSMs redundant. It makes them more effective — by giving them the visibility to act on what they already know before it becomes a lost account.

Ready to apply this?

AI for Customer Success Track

Go from reading about AI to using it in your daily work — with role-specific lessons built around your actual job.

View track