Monday Customer Health Brief

You're the owner. When the shop was small you could feel who was happy. You remembered the house where the part was late, the patient who sat too long, the invoice that never got a second notice. Then you added a truck, or a hygienist, or a second location, and the feeling went. You still think you'd "know" if someone was about to leave a three-star or ghost a $4,000 invoice. You wouldn't. You'd find out on Google, or when they called the shop down the road.

This playbook is a Monday habit, not a customer-success department. You prepare a permitted, minimized summary of last week's work for your approved AI workspace (Reactive). You get a one-pager of five accounts to touch (Proactive). You get drafted emails a human still sends (Active). That's the whole machine. Delivery is the truth serum. Not a sentiment dashboard, not a "health score" you can't explain to your spouse.

Map the idea into a shop: "did we ship" becomes jobs completed on time, callbacks, late parts, three-star reviews, unpaid invoices. If you didn't finish the work, didn't come back when you said you would, left them waiting on a part, collected a shrug on Google, or let the bill sit, they are not healthy. You don't need a model to invent a vibe on top of that.

Effectiveness. You keep the customers who already paid you, and you catch the unpaid invoices and the wobbling accounts before they become a lost replacement or a lost family of four. Efficiency. Monday is forty-five minutes, not a scavenger hunt through four logins. Quality. The person who had a late part gets a human note, not silence, and you never automate an apology you don't mean.

Who it's for

You run a service business — HVAC, dental, a shop with jobs and invoices — and you have enough customers that you can no longer keep the pulse in your head. You already have last week's jobs somewhere: Jobber, Housecall, ServiceTitan, a PMS, a spreadsheet the office manager hates.

You are not trying to "do customer success like a SaaS company." You are trying to start the week knowing which five humans deserve a call or a note, and having the note drafted so you actually send it.

What you get when done

A repeatable Sunday-night or Monday-morning dump. A one-page brief: five accounts, why they matter, what to do. Drafted emails or texts sitting in your drafts folder. A rule that nothing customer-facing sends itself. After two Mondays, you will know whether the signal is useful. If it isn't, you will change the prompt, not buy a platform.

Time and tools

First build: two hours to pick the source of truth, write the dump format, and run it once on last week. Then 45 minutes every Monday: 15 to dump, 15 to read the brief, 15 to edit and send.

Tools:

• The source system: field service management (FSM), such as ServiceTitan, Housecall Pro or Jobber; a dental practice management system (PMS); law practice management; and accounting, such as QuickBooks. Keep identities and detailed records there. Export only the approved columns below.
• Reviews. Google Business Profile — last 7 days, including 3-stars you hoped nobody saw.
• Invoices. Unpaid and overdue from the same system. Aging, not a vibe.
• ChatGPT or Claude. I use Claude for long messy dumps; ChatGPT is fine. One chat per Monday, named with the date.
• Email or the shop SMS number. Drafts only.
• Optional, later: Zapier or Make to email you the dump reminder. Not in week one. Week one is you, a CSV, a prompt.

Prerequisites

You can export or copy last week's completed jobs. You can see unpaid invoices. You can see new Google reviews. You have a definition of "done work" that isn't "we feel busy." You will name one human who sends — you, the office manager, the coordinator — not "whoever has a minute."

If your jobs live only in techs' heads and a whiteboard, fix that enough to list last week's internal account IDs and outcomes locally. The model cannot brief a week you refuse to write down.

Step 1 — Reactive: prepare last week's permitted data

Safety & data

Use the shop's own approved business AI workspace, with a named administrator. A paid plan alone does not make customer data appropriate to upload. Before pasting, strip names, phone numbers, addresses, dates of birth, account numbers, card data, staff HR details and privileged material. Use synthetic or properly de-identified examples by default; initials alone are not de-identification. Keep the identity key in the source system. For PHI, the exact product, features and workflow must be approved and covered by a signed BAA where required; otherwise leave the data out. Legal confidentiality and other contractual duties still apply.

A free or personal chatbot is not where customer records go. Read the customer-data Field Note and install the one-page AI use policy.

AI drafts; the named office manager assigns each flag and the relationship owner checks and sends outreach. Re-identify the random IDs locally after the brief. Do not paste clinical visit types, legal matter descriptions, free-text complaints or raw reviews linked to a person. A dataset with IDs, dates and balances may still identify someone; have the data owner approve the minimized fields and workflow. If not approved, use synthetic rows to test the prompt and review real records locally.

Export only these columns

Pull from the FSM, dental PMS, law practice-management system and accounting as applicable. Keep the period consistent: for example, Sunday through Saturday. The source owner reconciles IDs before export. Flag missing records instead of filling blanks from memory.

Brief the owner of a local [business] on this permitted minimized weekly sheet.
Columns: random customer ID, last visit/job date, open estimate amount, AR days,
complaint Y/N, membership status, assigned owner role. [paste approved rows]
Use only these fields. Missing means unknown. Do not infer medical conditions,
legal issues, reasons for late payment or employee performance.

Output:
A. One paragraph describing the observable week and gaps in the data.
B. Flags, with columns: customer ID | source fact | risk | owner-per-flag |
   proposed by-when | human action | information to verify locally.
C. Unassigned or overdue flags at the top. Every red flag needs one named human
   and a confirmed by-when before the brief is acted on. If only a role is
   available, mark NAME REQUIRED; do not invent a person or deadline.
D. List IDs with AR aging over 60 days and total supplied open estimates separately.
   AR days do not reveal balances; never infer AR dollars from open estimates.
E. Questions the source owner must answer. No health scores, diagnoses or guesses.
Do not draft customer messages, send, or change records.

The office manager verifies each flag against the source, resolves the ID and role locally, and writes the person's name and agreed deadline on the internal brief. Copy the aggregate AR-over-60-days dollar total directly from the accounting aging report to the scoreboard locally. Open estimates are not unpaid invoices; the model cannot derive that total from this seven-column export. A flag without an owner is unfinished work. For dental data, the privacy lead must confirm actual HIPAA de-identification or the covered BAA workflow; random IDs and exact visit dates alone do not qualify as de-identified.

Done looks like. One week, one minimized sheet, one verified flag list. Every red flag has a named person and a by-when. Customer identities have stayed in the approved systems.

Step 2 — Proactive: Monday one-pager of 5 accounts to touch

Why. A complete dump still loses if it takes forty minutes to read. You need the Presidential Daily Brief of a two-truck shop: five account IDs, why, and who will act by when. Once you trust the smoke signal from Step 1, you stop lighting the fire by scrolling Jobber on Sunday night and hoping you remember the late evaporator coil.

Exactly what to do.

Same chat, or a second prompt in that chat after correcting unsupported flags against the source ("drop account A17; that issue is resolved"). Paste the one-pager prompt.

Turn the check-in list into a MONDAY ONE-PAGER the owner will read in five minutes.

Hard cap: FIVE accounts. If you have more, keep the five that mix (a) money at risk, (b) a customer who had a bad delivery last week, (c) a review that needs a human, (d) a good account we could wow this week — at most ONE wow. Do not fill the page with wow if delivery is on fire.

For each of the five:
- Random account ID (re-identify locally after the brief)
- Owner-per-flag and by-when; mark NAME REQUIRED if not yet assigned
- Why this week (point at last week's fact)
- Risk type: delivery / review / unpaid / opportunity
- Suggested human action in one sentence (call, text, email, wait)
- Do NOT draft the message yet
- Do NOT score them 0–100. High / medium / low is enough, and only if you can say why from the dump

Then a 5-line Executive Summary at the top:
- 3 risks
- 1 opportunity (or "none this week")
- 1 pattern supported by these columns (complaint flags, lapsed memberships, AR aging) or "no pattern"
- Never infer late parts, no-shows or an underlying cause from absent fields

If five accounts would require inventing problems, give fewer. Empty is allowed.

Put this one-pager where you will see it: printed on the desk, a note in the Monday huddle, an email you send yourself. The point is that Monday starts with names, not with inbox. Keep the named, re-identified brief inside the approved shop system; do not paste it back into AI.

You do not need Zapier yet. If after four Mondays you still trust the prompt, you can have a tool remind you to dump the CSVs. I would not wire Jobber directly into a model until you have watched it lie at least once and caught it. Automation of a sloppy dump just mails you slop at 8 a.m.

Done looks like. A one-pager with five or fewer names, each tied to a fact from last week. You read it before you unlock the rest of Monday. Someone on the floor agrees the five are real.

Step 3 — Active: drafted emails, human sends

Why. Insight that doesn't become a call is a diary. The last move is hospitality prep: the messages are ready so you spend your judgment on tone, not on a blank page. The model does not get to hit send. That's Quality. Do not auto-send an apology; a wrong recipient would create a new problem.

Exactly what to do.

Before any customer text: ask your phone/SMS provider which registration applies. US application-to-person business messages sent over 10-digit local numbers generally require A2P 10DLC brand/campaign registration; toll-free and short-code messaging use different processes. Registration is not consent. Record the required consent, respect opt-outs, and verify the provider's STOP/HELP handling with a test. Do not infer SMS consent from a missed voice call alone. Obtain and retain consent appropriate to the sender and message purpose; permission for one reply is not permission for recurring follow-up. If permission is unclear, use the approved human callback path instead.

For each of the five, decide channel: phone call (no draft needed beyond a 3-line talk-off), SMS from the shop number, or email. Then paste the outreach prompt once with all five, or once per account if the context is long.

The outreach prompt (copy this):

Draft outreach for the five accounts on my Monday one-pager. I am the owner of [shop], [city]. Voice: first person, short, specific, no "just circling back," no "your business is important to us."

One-pager:
[paste only the permitted de-identified fields for the five]

Extra facts I know that were not in the dump (optional):
[permitted de-identified facts — or "none"]

For each account write:
- Channel: call talk-off (5 bullets) OR SMS (under 50 words) OR email (subject + 80–120 words)
- The actual draft, using placeholders for names, contact details and amounts to fill locally
- Assigned relationship owner and by-when from the approved sheet
- A note labeled HUMAN MUST CHECK: what fact in this draft could be wrong
- A line we will NOT say (don't blame the tech by name, don't promise a credit you didn't approve, don't diagnose, don't offer a discount I didn't list)

Rules:
- If last week was late or a callback, own it. Do not hide behind "unforeseen circumstances."
- If the invoice is overdue, be direct and kind. Name the amount I provided. Do not round. Do not threaten collections unless I wrote that policy.
- If the review was 3-star, draft a *private* reply asking to make it right, and a short public reply that does not argue. I will post the public one myself.
- If this is the "wow" account, do not sound like a coupon. Offer a real next step I listed (tune-up, recare, a thank-you).
- NEVER send. Label every draft DRAFT — HUMAN SENDS.
- SMS uses the consent and registration checks above; include STOP/HELP wording.
- Public healthcare replies must not confirm a patient relationship or treatment.
- Do not invent a job we completed, a part that arrived, or an apology for something that isn't in the facts.

Paste the drafts into the approved email, PMS message window or SMS tool — unsent. The relationship owner fills placeholders from the source record and checks recipient, facts and permissions. Public dental replies must not confirm a patient relationship or mention treatment; use “Please call our office manager at [number] so we can look into this.” Walk them with the person who owns the relationship. Change the sentence you would never say. Then send, or pick up the phone and ignore the draft.

Examples of what this is for, in a shop:

• Late part: "The coil we ordered for [address] landed [day]. I can be there [window]. If that window is wrong, reply with two that work."
• Callback: "We came back because the first visit didn't finish it. I want to hear whether the house is doing what we said it would."
• 3-star: public = thank them and ask to continue offline. Private = make it right. Never argue the star count.
• Unpaid: the amount, the job, how to pay, a human name. Not a novel.
• Wow: the family that paid a replacement on time and left a specific 5-star — a thank-you from you, not a review-beg in disguise. If you want a review ask, that's a different playbook, and you still never fabricate one.

Done looks like. Up to five checked drafts, each either sent by its human owner or consciously skipped. Record corrections; do not manufacture an edit just to tick a box. No automatic customer sends.

Worked example (hypothetical dental practice)

Hypothetical. Not a client. A coordinator has an overdue-balance list and several unresolved complaints. The practice data owner approves a minimized workflow; if the exact AI product and agreements do not cover the data, the coordinator performs the real review locally and tests prompts on synthetic rows only.

A synthetic demo sheet uses random IDs A01–A05, coarse visit periods, complaint flags, AR aging and owner roles. There are no names, initials, procedure details or review quotations. The prompt surfaces an older balance and an unassigned complaint; it cannot explain why either occurred.

Monday, the coordinator checks the real source records locally, assigns the complaint to the office manager by noon and the balance to the billing lead by Tuesday. Each accepts the task. Drafts contain placeholders. Staff fill them in the approved system, verify the recipient and send or call themselves. No public reply identifies someone as a patient.

At the next brief, “attempted” means a human actually contacted the account. “Kept” means the office can point to an agreed, completed next step, not an AI prediction. An unresolved case stays open with its owner. The brief is useful because responsibility is visible.

How this maps if you don't "ship software"

If you are HVAC, late parts and same-week callbacks are the truth serum. If you are dental, wait time, remakes, recare holes, and balances are the truth serum. If you are an agency, stalled deliverables and unsigned change orders. Same three steps.

Operator Standard check

Trace: Review one consistent week of permitted account fields.

Choose: Choose at most five factual flags; assign a named owner and by-when.

Build: Save the briefing and draft prompts; fill identity placeholders locally.

Prove: Count attempted and completed saves and track overdue balances.

Failure modes / what NOT to automate

Do not auto-send. Not the apology, not the invoice nudge, not the review reply.

Do not let the model invent a health score you cannot point at a row for. High / medium / low with a fact is enough. "72" is costume jewelry.

Do not paste clinical notes, legal files, addresses, initials, card data or raw customer complaints into the AI brief. Removing a name alone is insufficient. Use only the data owner's approved minimized fields, or synthetic rows.

Do not treat Jobber or the PMS as the whole relationship. A customer can have on-time jobs and still be done with you because your estimator was rude. The dump will miss that. That's why a human sends, and why empty "wow" slots are allowed.

Do not wire four tools in Zapier on Monday one. You will spend the month authenticating and skip the brief. Manual dump until the prompt is boring.

Do not "reach out" to everyone who paid you last week. That's how you become noise. Five accounts.

Do not have the model offer discounts, free visits, or refunds you didn't put in the facts. It will be generous with your margin.

Do not argue with a 3-star in public. Draft, human, short.

Do not skip a week because last week "felt fine." The weeks that feel fine are when the 61-day invoice quietly becomes 90.

If it goes wrong (kill switch)

The office manager can stop the export and pause every downstream draft. Disable any scheduled export if one was later added; do not delete source records. Verify IDs and facts locally before restarting. If a wrong message reached someone, the relationship owner calls: “Our message included an error about [verified fact]. Here is the correction and the person handling it.” Report any data exposure to the shop's incident owner immediately and follow the approved incident process.

Prove it: 4-week scoreboard

Use the same definition and observation window in every column. Fill the baseline before the install; leave unknowns blank, not zero. The thresholds below are decisions you set, not promised results.

Save it as a skill

Save the final prompt and never-list as skills/monday-customer-health-brief.md in the shop's shared drive, or in the approved project instructions your AI workspace supports. Keep one owner and one dated master copy. Copy this instruction block and replace every bracket before use.

Skill: Monday Customer Health Brief
Owner / backup: [office manager] / [name]
Read current company.md, customer.md and rules.md.
Use only approved minimized columns; synthetic examples if approval is absent.
Keep the identity key and re-identification local. Never infer diagnoses or motives.
Output source-linked flags with owner-per-flag and by-when; ask for missing owners.
Never-list: [shop's data exclusions, claims, discounts and escalation rules].
Prepare drafts with placeholders. The relationship owner verifies, fills and sends.
Pause by stopping the export and pending drafts. Review saves and AR each Monday.
Version / reviewed by: [date] / [name]

Use the Three Outcomes Filter, Quote Follow-Up and Shop Context Pack to connect the brief to action. Read the Field Note on measuring whether AI pays off.

FAQ

I don't have Jobber or a PMS export. Can I still do this? Yes. A local sheet of account IDs, approved dates, complaint flags and aging is enough. Use synthetic data with AI if the real export is not approved. If you cannot list last week's jobs at all, the playbook is not your problem. Your board is.

Should I connect Claude to my tools so I don't paste? Later, maybe. First, paste. You need to see the lie with your own eyes. Connectors make the lie faster. When the Monday output has been useful for a month and you are sick of CSV, then look at a scheduled dump. The thought pattern is the point, not the logo on the scheduler.

What if the five accounts are all unpaid invoices? Then that's the week. Money is a health signal. Don't force a "wow" onto the page to make the brief feel balanced. Pay the coordinator to work the aging. Put a wow on the list when you actually have a delivery win.

Is a 3-star really "customer health"? It's Quality, public, with a name on your Google page. It's also Effectiveness, because the next stranger reads it. Treat it as a check-in, not a brand crisis. Make the person's visit right if you can. Don't beg them to change the star. Don't write a fake 5-star from the shop to bury it. We never fabricate reviews.

Screen-recording notes (8–12 min, lo-fi)

1. 0:00–0:50 — Camera. "When you add trucks or chairs, you lose the feel. Don't buy a dashboard. Dump last week."
2. 0:50–3:30 — Screen: a synthetic CSV using only the seven approved columns. Explain that live identities stay in the source system. No fancy BI.
3. 3:30–6:30 — Paste the weekly prompt. Scroll the output. Cross out a invented score out loud. Circle three random account IDs that match the synthetic rows.
4. 6:30–8:30 — One-pager prompt. Five accounts. Kill a sixth. Read the executive summary.
5. 8:30–10:30 — Outreach prompt. One apology draft. You rewrite a discount the model offered. Label HUMAN SENDS. Do not click send on camera unless it's a dummy draft.
6. 10:30–end — Close: "Let’s talk about your business. Start with a free introductory conversation. The $1,500 AI Workflow Sprint helps your team use AI on real work; implementation is scoped separately." Direct viewers to /start.