← Back to all work

AI Active Case Agent

Helping teams spot risky cases earlier

A predictive risk score built from signals support leads already watch, with alert rules that respect CRM status, working hours and repeat noise.

Role
Product owner
Scope
Risk signals, alert rules, CRM status handling

The problem

Support leads often learn a case is in trouble when the customer escalates. The goal was to flag rising risk while the team still had time to act.

My role

I defined the signals behind the risk score and how the agent behaves when it raises an alert.

The signals I defined

SignalWhat it adds to risk
SentimentFrustration that builds across messages, not a single sharp reply.
SeverityBusiness impact the customer or agent has set on the case.
ComplexityCases that need several teams, products or handoffs.
Response SLATime left before the next response is due.
Resolution SLATime left before the resolution target is missed.
Account tierThe contractual and commercial weight of the account.
Case ageHow long the case has been open relative to similar cases.

Alert behavior

Different CRM statuses

Status changes what risk means. A case waiting on the customer should not build risk the way a case waiting on us does, and closed or resolved cases leave scoring.

Time when the team can act

Alerts arrive when the support team can respond. A risk raised outside their working hours is held for the start of their coverage instead of paging no one.

Cooling-off for repeat alerts

After an alert, the same case stays quiet for a cooling-off period unless its risk rises further, so leads see new problems instead of the same one again.

Illustrative exampleCase 52190 · Enterprise account · fictional

Signal levels

  • Sentimenthigh
  • Severityhigh
  • Complexitymedium
  • Response SLAmedium
  • Resolution SLAhigh
  • Account tierhigh
  • Case agelow

How the agent behaved

  1. Tue 22:40Risk crosses alert thresholdHeld: outside team hours
  2. Wed 09:00Alert sent to support leadTeam coverage starts
  3. Wed 11:30Risk unchangedSuppressed: cooling-off
  4. Wed 15:10Sentiment drops, resolution SLA nearRe-alert: risk rose
  5. Thu 10:00Status set to Waiting on customerResponse clock paused

Metrics and outcomes I set

  • Risk alerts acted on by support leads
  • Alerts raised while there is still time before an SLA is missed
  • Repeat alerts that add noise
Outcome I aimed for

Alerts that arrive early enough to act on and rarely enough to trust.

Watch and read