How we cut customer churn by 38% using an AI agent that never sleeps

Churn doesn't happen overnight

In our experience, customers who churn show clear warning signs 3–6 weeks before they actually cancel. Login frequency drops. They stop using core features. Support tickets spike. They go quiet on Slack. The signal is there — the problem is that no one is watching all of it at once, all the time.

We built an AI agent to watch it for us. It runs 24/7, monitors 14 behavioural signals per account, and triggers intervention before customers reach the point of no return.

The 14 signals we track

We categorised signals into three groups:

Engagement signals

• Daily active users (rolling 7-day average)

• Core feature usage frequency

• Session depth (pages per session)

• Time since last login (by seat)

Relationship signals

• NPS score trend (last survey vs previous)

• Response rate to CS outreach

• Slack community activity

• Executive sponsor engagement

Commercial signals

• Renewal date proximity

• Expansion vs contraction of seat count

• Payment method expiry approaching

• Open support tickets (severity weighted)

• Contract document views (have they opened the contract?)

• Competitive tool detected in tech stack

How the agent scores risk

Each signal is scored weekly and fed into a risk model that outputs a health score between 0 and 100. Accounts below 60 are flagged as at-risk. Below 40 are considered critical.

function calculateHealthScore(signals) {
  const weights = {
    engagement: 0.40,
    relationship: 0.35,
    commercial: 0.25
  };

  return (
    signals.engagement * weights.engagement +
    signals.relationship * weights.relationship +
    signals.commercial * weights.commercial
  );
}

function calculateHealthScore(signals) {
  const weights = {
    engagement: 0.40,
    relationship: 0.35,
    commercial: 0.25
  };

  return (
    signals.engagement * weights.engagement +
    signals.relationship * weights.relationship +
    signals.commercial * weights.commercial
  );
}

function calculateHealthScore(signals) {
  const weights = {
    engagement: 0.40,
    relationship: 0.35,
    commercial: 0.25
  };

  return (
    signals.engagement * weights.engagement +
    signals.relationship * weights.relationship +
    signals.commercial * weights.commercial
  );
}

"We used to find out accounts were churning when they submitted the cancellation form. Now we know six weeks before they've even made the decision."

The intervention playbook

Results after 6 months

• Churn rate reduced from 6.2% to 3.8% monthly

• At-risk accounts saved: 34 out of 41 flagged

• Average intervention lead time: 38 days before renewal

• CSM capacity freed up: 11 hours per week per person

• NRR improved from 104% to 118%

The 38% churn reduction came almost entirely from catching at-risk accounts earlier — not from changing the intervention itself. Timing is everything in retention.