The account list: tiles for total, green, amber and red accounts, filter chips, and a table with account name, owner, yearly revenue, health score with a small trend line, usage, tickets, billing and a status pill, one red account selected.
Accounts. Every customer, its score and the trend. Sample data.

Why this sample

Customer teams in B2B software often learn that an account is in trouble when the cancellation email arrives. The warning signs were there, but spread over three tools: usage fell, tickets piled up, a payment failed. One score with a plain reason is a clean Feature-tier job, built on data the company already has.

This is a sample build. There is no client, and nothing here is a result from a real company.

What it does

  • Pulls product usage, support tickets and billing into one table per account, once a day.
  • Gives each account a score from rules and weights the team can read.
  • Marks accounts green, amber or red and shows the trend over time.
  • Writes a short note when an account turns red, with a link behind every sentence.
  • Alerts the account owner when one of their accounts changes colour.
One account page: a score of 38 that fell from 71, a note that says why it turned red in three sentences with links to usage, tickets and billing, a line chart of weekly active users, open tickets and a failed invoice, and a score breakdown that adds up to 38.
Account page. Why it turned red, and the data behind each line. Sample data.

The AI part

The score is plain rules and weights, not a model. The language model only writes the note on why an account turned red. It gets the signals that moved and links each sentence to its source: a usage chart, a ticket, an invoice. It never changes the score, never emails the customer and never guesses at reasons it cannot see in the data.

We would check the notes by hand on a set of red accounts: does every sentence match the data behind its link.

The score settings: weight sliders for usage, support and billing that add up to 100, a list of rules with point values, and a preview of how many accounts would be green, amber and red with the new weights.
Weights. The team sets the rules and sees the effect before saving.

Where it stops

No automatic outreach to customers and no churn prediction model. A trained model needs a real history of lost accounts first, so it is a later step.

Timeline

When What happened
Week 1 Read-only connections to product events, helpdesk and billing; a daily sync into one account table.
Week 2 Score rules with weights, the account list with filters, score history per account.
Week 3 Account page, the why-red note with links to its data, an alert to the account owner.
Week 4 Hand check of notes, weight tuning with the team, runbook and handover.