How to build customer health signals

What matters here

• Choose signals because they predicted churn in your data, not because they were easy to get.
• Never collapse to one number alone. Two accounts scoring 62 can need entirely different conversations.
• Back-test against customers who actually left. A score nobody validated is a shared opinion with a colour attached.
• Alert on a category moving, not on a total crossing a line. The total moves slowly and hides the change.
• Track whether flagged accounts actually churned. If they did not, the model is decoration.

Who this is for

You run customer success operations. There is a health score somewhere, nobody quite trusts it, and the accounts that churn were usually green the month before.

How it works in practice

The path from a customer going quiet to somebody doing something about it.

1. 1

   Signals are gathered per category

   Usage, support, engagement and commercial, kept separate rather than merged on collection.

2. 2

   Each category is scored on its own

   So a healthy product user with a support problem does not average out to fine.

3. 3

   Weights come from a back-test

   Fitted against accounts that actually churned, not chosen in a workshop.

4. 4

   The categories stay visible on the record

   Alongside the total, because the total says whether and the categories say why.

5. 5

   Alerts fire on a category moving

   Not on the total crossing a threshold, which happens late and slowly.

6. 6

   Every flag is tracked to an outcome

   Did the flagged account churn. That number decides whether anybody should trust this.

Paste these into Claude Code or Codex with the Tray Headless plugin installed. Each stage runs on its own. The systems named in them are the worked example rather than a requirement, and every prompt says so.

Once per project, run /tray-workflows:set-workspace to pick the workspace these build in. Point it at a sandbox first.

1. 1

   Set up, then look at who actually left

   The history is the only honest source of weights.

   Headless skills build-workflow tray-patterns

   Copy

   Use build-workflow. The systems in play are Snowflake and Salesforce,
   or whatever we run in those seats.
   
   Before building a score, pull the last two years of churned accounts and
   show me, for each, what the six months before looked like: usage trend,
   support volume and escalations, meeting frequency, payment behaviour, and
   whether a named champion left.
   
   I want to see which of those actually moved before churn and how far
   ahead. That is where the weights come from. Anything that did not move
   should not be in the model regardless of how sensible it sounds.

2. 2

   Score categories separately, and keep them separate

   Averaging is how a real problem disappears.

   Headless skills build-workflow

   Copy

   Use build-workflow. Score four categories independently:
   
     Usage: active users against seats, breadth, and direction of travel
     Support: ticket volume against the norm for that size, escalations,
     and satisfaction where we have it
     Engagement: days since a meeting, executive contact, champion still
     present
     Commercial: payment behaviour, contract value trend, open expansion
   
   Score each 0 to 100 and store all four on the account alongside the
   total.
   
   Do not let a good category hide a bad one. An account with usage at 90
   and support at 20 averages to fine and is not fine, and the categories
   are the only thing that shows it.

3. 3

   Fit the weights against real churn

   Otherwise the model encodes what people believe instead of what happened.

   Copy

   Fit the category weights against the churned cohort rather than choosing
   them in a workshop.
   
   Then back-test: run the scoring over historical data and check how many
   accounts that churned were flagged, how far in advance, and how many
   that were flagged did not churn.
   
   Report both error types. A model that flags everything catches every
   churn and is useless. A model that flags nothing looks precise and is
   also useless.
   
   Show me the back-test before this goes near a customer success team. A
   health score nobody validated is a shared opinion with a colour attached,
   and it takes one wrong call for people to stop looking at it.

   The false positive rate is the one that decides adoption. If CSMs chase five healthy accounts for every real one, they stop chasing.

4. 4

   Alert on movement in a category

   The total moves late and slowly. Categories move first.

   Headless skills tray-gotchas

   Copy

   Use tray-gotchas, then alert on movement instead of level:
   
     A category dropping more than a band in a period
     Two or more categories declining together, which is a stronger signal
     than one falling further
     Any decline on an account with a renewal inside two quarters
     A champion leaving, which should alert on its own regardless of score
   
   Route to the account owner and the customer success manager with which
   category moved and what changed inside it.
   
   Never alert only on the total. It moves last, by which point the
   conversation is a rescue, not a check-in.

5. 5

   Track outcomes, and be willing to delete signals

   A model nobody measures becomes decoration within two quarters.

   Copy

   For every flag, record what happened: nothing, intervention, saved,
   churned.
   
   Report quarterly: churn rate among flagged versus unflagged accounts,
   average warning time before churn, false positive rate per signal, and
   accounts that churned without ever being flagged.
   
   That last list is the most useful thing here. It is where the missing
   signal is, and it is the only way the model improves.
   
   Retire any signal whose weight fits to near zero on the back-test. A
   score built from twelve inputs where four do the work is harder to trust
   and no more accurate.

6. 6

   Prove it works, then hand the weights over

   Because the model changes as the product and the segments do.

   Copy

   Run the per-step schema checks and the whole-workflow audit before this
   touches production.
   
   Then open the same workflow in Tray Build so customer success operations
   can adjust the category weights and the alert bands in the visual canvas,
   and re-run the back-test after every change rather than trusting that it
   still holds.

What it connects to

Health is assembled from four systems and lands in the one where relationships live.

Named systems are the ones most teams run, not the only ones that work. Each is an authentication in your Tray workspace, referenced by name, so the workflow never holds a credential. Where we have a connector page, the name links to it.

Questions people ask

Related guides