A daily watcher with five inputs and one job.
The Customer Health Monitor is a scheduled co-worker that runs once a day, before the CS team's stand-up. It reads CRM activity, ticket sentiment, product usage, exec contact recency, and renewal proximity for every account. It ranks accounts by composite risk score, drafts a one-paragraph explanation for the top 3–5, and DMs the assigned CSM.
The CSM decides what to do — reach out, escalate, file a save plan, or note and watch. The co-worker doesn't write to the CRM, doesn't email the customer, doesn't post in shared channels. Quiet by default.
What it reads.
| System | Objects / tables | Access |
|---|---|---|
| CRM (Salesforce) | Account, Opportunity (open), Activity, Contact | Read-only, named objects |
| Tickets (Zendesk) | Ticket, ticket comments (last 90d) | Read-only, customer-bound |
| Product (warehouse) | workspace_usage_daily, login_events | Read-only, daily refresh |
| Calendar | Meeting metadata only (no body) | Account/contact join |
Nothing else. The co-worker does not need (and does not receive) access to email bodies, internal Slack channels, or contract pricing. Scope creep is the most common security review blocker — we scope tight from day one.
07:00 to 07:30. The CSM has the brief before stand-up. No surprise alerts during the day.
Each measured against the account's own baseline.
- Usage delta (14d). Rolling-sum product events, deseasonalized by day-of-week. Z-scored vs same account's prior 90d.
- Ticket sentiment. Count + severity-weighted score vs same account's median. One P0 outweighs five P3s.
- Login frequency. Distinct daily-logins drop-off as a fraction of the account's typical weekly pattern.
- Exec contact gap. Days since last meaningful interaction with a named exec sponsor.
- Renewal proximity. Multiplier on the composite. The same 2σ anomaly matters more 60 days from renewal than 360.
Per-account baselines are the magic. Industry-average thresholds catastrophically over-fire on heterogeneous customer bases. Each account is scored against its own history.
Three short paragraphs, one chart, three citation links.
Top risk: Northwind Bank · score 8.2
Usage down 38% w/w (vs 4% typical). Two P0 tickets unresolved >72h. Last exec contact 47 days ago.
Sources: usage_daily · zendesk #142 · sfdc.Activity.
[chart: 30d usage trend]
Brief, sourced, no recommendation. The first iteration had the model recommending next-best actions; CSMs felt it overstepped and lost trust in the rest. Removing recommendations was a one-line change that doubled engagement.
What the CSM does.
- Acknowledge — CSM reads, optionally clicks through to citations, logs "noted" with a one-line reason. Tracked.
- Open save plan — CSM opens a CRM task with the alert pre-filled. Co-worker stays out of CRM.
- Escalate — one-click forward to CSM manager Slack channel. Used for accounts trending bad for 3+ consecutive days.
- Suppress — CSM can mark an account "in active save mode" for 14 days; co-worker holds back further alerts on that account during the suppression window.
Every action is logged. The audit log is what you use during quarterly retrospectives to ask "did the co-worker actually catch the accounts that churned?"
"By month two, the CSMs were checking the morning DM before their CRM. That's the inflection point — the moment a Watch co-worker becomes the front door."
What we measure to know it's working.
- Time-to-first-action on at-risk accounts. Baseline pre-rollout. Target: cut in half by month 2.
- CSM trust score (weekly micro-survey: would you miss the DM if it stopped?). Target: > 7/10 by week 4.
- Catch rate vs churned accounts. Of accounts that churned, what % showed up in the daily top-5 in the 60 days prior? Target: > 70% by quarter 2.
- Alert noise. What fraction of DMs result in "noted, no action"? Above 60% means we're over-firing — tune up.
Build time: 2–3 weeks.
If you have a CS team, a CRM, and a product usage signal, this pattern fits a standard 2-4 week engagement. The first 45 minutes is the assessment call.