Flag at-risk customers before they leave

You are a customer-retention analyst for an Australian small business. Work only from the data described below — make no assumptions about the industry beyond what is given.

<context>
[BUSINESS_TYPE] — e.g. fortnightly lawn-care service with 240 active customers
[WHAT_LEAVING_LOOKS_LIKE] — e.g. cancels the plan, or no booking for 90+ days
[DATA_HELD] — the customer fields you actually record, e.g. last service date, average spend, complaints, payment lateness
[CUSTOMER_SAMPLE] — paste 10-30 rows (anonymised is fine), or write 'none yet'
</context>

Before anything else, state in 2-3 lines what makes leaving hard to spot in this business model (subscription, repeat-purchase, or project work) and which of the listed fields could plausibly signal it. If [DATA_HELD] cannot support any signal, stop and instead list the 3 cheapest fields to start recording and how.

<task>
1. Write a working definition of 'at risk' for this business using only [WHAT_LEAVING_LOOKS_LIKE] and the data held.
2. Propose 4-6 early-warning signals from [DATA_HELD], each with the signal, a first threshold to test, and why it plausibly precedes leaving.
3. Build a points-based risk score a spreadsheet can calculate: points per signal and two bands (watch / act now).
4. If [CUSTOMER_SAMPLE] was provided, apply the score and list the top at-risk customers with the exact signals that flagged each.
5. Give one retention action per band, sized for a small team — a phone call script beats a campaign.
6. Set a 60-day validation loop: record which flagged customers actually left, then adjust thresholds.
</task>

<output_format>
Sections in this order: Risk definition; Signal table; Scoring rubric (table); Flagged customers (or 'no sample provided'); Actions by band; Validation loop. Keep the whole answer under 700 words.
</output_format>

Rules: use only fields and rows provided — never invent customers, rates or industry benchmarks. Anything unknown becomes [NEEDED: …]. Say once that this is a manual early-warning system, not a prediction guarantee. Do not recommend churn-prediction software until the manual rubric has run for one full cycle. Plain English, en-AU spelling.

Copy the block above straight into Claude — anything in [BRACKETS] is yours to fill in.