Health Score Model

Health Score Model

What this does

Builds a health-score model for a whole book of business: the inputs that actually predict retention, the weight each one carries, the thresholds that split green from yellow from red, and the one specific action attached to each tier. The output is a scored book where every color tells the CSM what to do next, not just how to feel.

What you'll need

You do not need to connect anything to get value today. Bring your account list and the skill runs now. Connect the tools below and it pulls the rest and adds signals you cannot paste by hand.

• Works today with: your accounts and whatever signals you have, usage, support load, sponsor status, renewal date, last touch. Paste it or upload a CSV.
• More powerful connected to a CRM: it reads the book, ARR, renewal dates, and open cases automatically.
• Sharper with a product-analytics tool: real adoption becomes the strongest input instead of a guess.
• Sharper with a support or ticketing tool: turns ticket volume and sentiment into a real risk input.

How this runs at your connection level

This skill is never reliant on a connector. It runs on the data you give it today and gets more powerful as you connect tools. It never invents a number it cannot see. A gap is a prompt, not a guess.

• Bring your data: paste or upload your book with whatever signals you have. The skill builds and applies the full model today on your real accounts. No connection required.
• Connect your tools: the same skill pulls usage, cases, and dates automatically and adds signals you cannot paste by hand. Same model, less effort, sharper.
• Just exploring: no book yet? Get the framework, the exact inputs it reads, the default weights, and a worked example on sample accounts, so you can see the shape before you feed it.

Every run ends with the one input that would most improve the model, a field to add or a tool to connect.

Customize this for yourself

This was built for a B2B SaaS org scoring a renewal book. Set these to your stack:

Pick inputs you can actually see. A model with a 40% weight on a signal you never measure is a model that lies confidently.

The method

Pick the inputs

Choose four to six signals that predict retention for your product: adoption, relationship depth, support health, engagement cadence, commercial trend. Fewer strong inputs beat many weak ones. Any input you cannot measure gets dropped or flagged, never faked.

Set the weights

Weight by predictive power, not by what is easy to pull. For most B2B SaaS, adoption carries the most and commercial the least, because usage moves first and the renewal number moves last. State the weights out loud so anyone can challenge them.

Draw the thresholds

Green, yellow, red are decisions about attention, not math for its own sake. Set cutoffs so green means "leave it running," yellow means "there is a plan," red means "escalate this week." Re-tune once you see the spread; if 80% of the book is red, the thresholds are wrong, not the book.

Action per tier

Each tier gets exactly one default action. Green: nurture and look for expansion. Yellow: a written recovery plan with an owner. Red: escalate, involve a leader, set a save play in motion. A score with no action is a dashboard nobody uses.

Explain every score

Every account shows the two inputs pulling it up and the two pulling it down. A CSM should never ask "why is this yellow." The model answers before they ask.

Quality gates

• Every input is measurable, or it is dropped from the model and noted.
• Weights are stated and sum to 100. No hidden factors.
• No account gets a color without the top reasons for it shown.
• Thresholds are sanity-checked against the book's spread, not set once and forgotten.

Output (example)

BOOK HEALTH · 5 sample accounts
Account      Score   Tier     Up                 Down
Account A    82      GREEN    adoption, sponsor  none material
Account B    61      YELLOW   adoption           champion quiet
Account C    58      YELLOW   engagement         support spike
Account D    34      RED      none material      no usage, no sponsor
Account E    73      GREEN    adoption, growth   one open case

Tier actions:
  GREEN  (A, E): nurture, scout expansion.
  YELLOW (B, C): written recovery plan, owner named, this week.
  RED    (D):    escalate to leader, open a save play now.

Where the numbers come from

The scores, weights, and thresholds come from the inputs you paste or the tools you connect. Nothing above is a real customer number; it is illustrative. The default weights (adoption 40, relationship 20, support 15, engagement 15, commercial 10) are a starting point tuned for one SaaS book, not a law. When an input is missing, the model redistributes weight and tells you which signal would sharpen the score.

Example chat

You: Build a health score for my renewal book, here's the account list with usage, support tickets, and sponsor status.

Claude: Built it on four inputs: adoption (40%), relationship (20%), support (15%), engagement (15%), commercial (10%).

Account      Score   Tier     Up                 Down
Account B    61      YELLOW   adoption           champion quiet
Account D    34      RED      none material      no usage, no sponsor

Account D looks red for the obvious reason: no usage, no sponsor. Account B is the one to actually watch. Adoption is fine, the flag is a quiet champion, which usage data alone would never catch. One sharpener: connect a support or ticketing tool so the support input reflects real ticket sentiment instead of the raw count you gave me.

Go further

The model works from a pasted book. Here is where the score stops being a snapshot and starts being a live signal.

• Rescore the book nightly. A scheduled Claude task recomputes every score against live Amplitude usage and Salesforce data overnight, so a color never lags reality by more than a day.
• Alert only on the flips. Have it DM the CSM in Slack just the accounts that changed tier since yesterday, not the whole book, so the signal doesn't get lost in the noise.
• Turn red into a tracked play. Feed every RED tier into an escalation or save-play skill automatically, so the score doesn't just sit there, it triggers the next move.

A score nobody acts on is a dashboard. A score that triggers the next move is a system.

Make it yours

Fork it. Change the inputs, the weights, the cutoffs, the action per tier. The point is not to run someone else's health score. It is to run one you can defend in a QBR and act on the same day. Built by an operator. Customize it, break it, make it better.