Run the Three Outcomes Filter (and Build This Quarter's List)

You're the owner. Last month someone sold you a smart receptionist. The month before that you bought seats. Next week a vendor wants 45 minutes for an agent that does intake, reviews, and ads. You do not have an AI problem. You have a pile problem. Too many ideas, no filter, nobody whose job it is to say no.

This playbook is the filter for one shop — yours. You are the champion. If you will not protect the list, there is no project. Staff fill a five-question form. You walk one real process for 45 minutes. Bring de-identified process notes into your AI workspace. Walk out with a scored list and a one-page roadmap: this month, next month. Then execute the least that actually pays.

Every idea on that list has to hit at least one door:

• Effectiveness — makes money. Booked jobs, filled chairs, consults that show, close rate, capacity you can sell.
• Efficiency — saves time. Named hours, named person, named task. Not "we'll be faster."
• Quality — the customer can feel it. Fewer errors, less "depends who was working," replies that sound like the shop.

If an idea cannot fill one of those three with a straight face, it does not make the quarter. Pretty dashboards do not count. Same six steps in dental, HVAC, law, agency, or retail. Examples change. The filter does not.

Who this is for

Owners of a dental practice, HVAC / home-services shop, law firm, agency, or a handful of retail locations. You have staff. You already pay for software. You are tired of buying tools that sit. If you want someone else to "own AI" while you stay out of it, stop here. The champion is you.

What done looks like

On one page, visible to you and the office manager:

1. A scored list of 6–10 ideas, each tagged Effectiveness / Efficiency / Quality, with a 1–5 money score, a 1–5 time score, a 1–5 CX score, and a "this month / next month / later / never" label.
2. This month: one workflow. Named owner. Named definition of done. Named "least that pays."
3. Next month: one backup, in case this month ships.
4. A never pile: no customer-facing auto-send, no fake reviews, no custom software before a small manual loop works.

You can hold the page in a Monday huddle without a projector.

Time and tools

• Owner: half a day across the week — form + champion note, 45-minute walk, 90 minutes in Claude, 30 minutes to freeze the one-pager.
• Staff: 8 minutes each. Everyone here more than 90 days. Four people means four answers. You do not need thirty.
• Tools you already have: Google Form, a phone, Claude, a Google Doc. Do not buy a strategy platform.

Prerequisites

• You, the owner, will say out loud: "I am the champion. If this stalls, it is on me."
• Staff will not get punished for writing "the estimate follow-up is a mess." You say that in the note.
• One process you can walk in 45 minutes. HVAC: call → dispatch → job → estimate → follow-up. Dental: new-patient form → confirm → visit → recare. Law: web form → intake → consult → engagement. Pick one. Not five.
• Shop-owned business AI workspace approved for the data in this exercise. Use synthetic examples or de-identified patterns, not staff records.

Safety & data

Use a business AI workspace the shop owns, with access limited to the people doing this job. A paid plan alone is not permission to upload records. Strip names, phone numbers, addresses, dates of birth, account numbers, staff HR details and privileged material before pasting. Removing a name alone may not de-identify a record. Use synthetic examples when the remaining facts could identify someone.

For PHI, your responsible privacy lead must verify the exact product, features and integrations are covered by a signed BAA and your required safeguards before use. Otherwise keep the data out. Customer records never go into a free or personal chatbot account. Keep legal matters and payment-card data in their approved systems. AI drafts; the named workflow owner approves, and the named sender sends.

Set the shared boundary with the customer-data Field Note and the one-page AI Use Policy.

Step 1 — You are the champion, or this dies

Why. There is no innovation committee. There is you. If this is a side quest for the office manager, she will collect a form and nothing will change. Kill two enemies in one note: fear ("this replaces me") and theater ("we're going to AI the whole company").

Clicks.

1. Open a blank email or a printed page for the Monday huddle.
2. Name yourself as the person who will pick the first workflow and protect it for 30 days.
3. Promise two things: nobody is getting fired for filling out a form, and you are not automating the whole shop this quarter.
4. Send it. Then send the form in Step 2 the same day. A champion note with no form is a speech.

Paste this, then put it in your voice. Cut anything you would not actually say.

Subject: 8 minutes, not a new job

I'm not replacing anyone with a chatbot. I am tired of buying tools that sit.

This week I'm collecting what's actually eating the day. There's a 5-question form. It takes 8 minutes. Be blunt. If the estimate follow-up is a mess, write that. If customers wait because we don't confirm, write that. Answers are for fixing work, not grading people. Nothing here goes in a review. Use the anonymous option if you prefer.

I will read every answer. Next week I'll walk one process with [NAME] for 45 minutes. Then we'll pick ONE thing for this month. Not twelve.

If you're skeptical because the last "AI thing" wasted time, good. That's useful. Put it in question 5.

— [OWNER FIRST NAME]

Definition of done. The note is sent. You can point to a date when you will pick one workflow. Staff heard "one thing this month," not "transformation."

Step 2 — Send the five-question staff form

Why. People will not always tell you in the hallway what's broken. They will tell a short form you don't grade. You are not mining for AI excitement. You are mining for weekly tasks that take too long, make customers wait, or get copy-pasted from last Tuesday.

Clicks.

1. Go to forms.google.com → Blank form.
2. Title: What's eating the week — [SHOP NAME].
3. Offer an anonymous version: turn off email collection, do not require sign-in, and leave name, tenure and location optional. Check your form’s actual settings and explain any remaining access logs. Offer a paper drop box too. A separate optional contact field is for volunteers who want follow-up, not identification of anonymous answers.
4. Add the five questions below as paragraph answers; allow people to skip. State: “Answers are for fixing work, not grading people. Nothing here goes in a review.” Do not ask for customer details or complaints about named colleagues.
5. Send the link in the same email as the champion note, or print it and put it in the break room with a due date: seven days, then you close it.
6. You fill it out too. If the owner skips the form, staff will too.

The form (copy the questions exactly, then add a line that names your shop).

0. How long have you worked here?
   [ ] Under 90 days   [ ] 90 days–2 years   [ ] 2–5 years   [ ] 5+ years

1. List the work you actually do in a normal week. Not your job title. The tasks.
   (Example: confirm tomorrow's hygiene, call no-shows, post claims, answer the overflow inbox.)

2. What is the most time-consuming or stupid-feeling task you still do by hand? Walk through the steps like I'm new.

3. Where do customers (or patients / clients / homeowners) get stuck or complain — the thing that makes us look sloppy even when the work was fine?

4. What do you copy-paste every day? Emails, texts, estimate paragraphs, chart notes, status updates. Describe a synthetic example; do not paste a chart note, customer message, privileged material or staff detail.

5. What have you already tried with ChatGPT or any "AI" tool — and did it help, or did it create more work? If you haven't tried anything, write "haven't." That's fine.

If location would reveal the respondent in a small team, omit it. Ask for an optional broad process category instead.

People closest to the work have evidence regardless of tenure. Weekly tasks reveal where help might be useful. "Stupid-feeling" = Efficiency. Customer stuck = Quality. Copy-paste = the first draft a model can take. Question 5 = who already burned themselves on a bad tool. Those people need a win they can see, not a pep talk.

Definition of done. Form is live. Due date is on the calendar. You submitted your own answers. Start Step 3 even with three honest answers.

Step 3 — Walk one process for 45 minutes

Why. A survey tells you what people hate. A walk tells you how a job actually moves — extra login, photo that never hits the estimate, sticky note on the monitor. One process. 45 minutes. Phone, Loom, or a notebook.

Sit with the person who does the work — office manager, dispatcher, intake, front desk. Not the associate with opinions and no queue.

Clicks.

1. Pick the process that showed up most in the form, or the one you already know leaks money. HVAC: a new estimate from first call to follow-up. Dental: new patient from web form to kept visit. Law: inbound consult from email to booked conversation. Retail: a special-order from the floor to pickup text.
2. Book 45 minutes on their calendar, during a real day, not a Saturday theory session.
3. Use a synthetic record and tight notes by default. Record only with everyone’s informed permission and the shop’s approved recording/storage setup; do not capture customer screens or conversations. Tell them you want the clicks, not a presentation.
4. Ask the questions below. When they give a mushy answer, push once. Then shut up.

Interview questions (print this).

A. Show me the last real one. Not a perfect one. Last Tuesday is better than "typically."

B. What starts it? Call, form, walk-in, Facebook message, missed call, referral.

C. What are the steps, in order, including the ugly ones. Who touches it. Which screen. What gets copied.

D. Where does it stall? Be specific: "it sits in my drafts until I have a quiet hour" is useful. "It's busy" is not.

E. How often does this happen in a week? How long does one take, start to send?

F. What does the customer feel at the stall? Do they chase us? Do they book the other shop?

G. If this broke in public — wrong price, wrong clinical note, wrong legal sentence — who gets the call?

H. What have we already bought that is supposed to do this, and do you actually open it?

I. What would "done" look like in 30 days that you would notice without a dashboard? (Example: estimates out the same afternoon. New-patient forms answered before lunch. Consult recaps same day.)

Push line, use it:

You said that takes "a while." How many minutes, on the last one? Which tab were you in? Who else had to touch it before it left the building?

Quantify while they talk: frequency per week, minutes each, whose time, what a miss costs. Guess if you must. Label guesses. Written-and-wrong beats "a lot."

Definition of done. You have a recording or tight notes for ONE process. You can list 5–15 micro-steps on a legal pad. You know which step is repetitive, time-expensive, and not the glamorous one. That step is the insertion point — not the whole workflow.

Step 4 — Dump it into Claude and score the list

Why. Form answers, a walk transcript, your hunches. Claude will turn that into a scored table. It does not know your shop. The first draft will offer twelve initiatives. You cut it to a quarter list.

Clicks.

1. Make a folder on your computer: ai-filter-[month].
2. Keep raw form responses locally with restricted access. Put only de-identified workflow patterns in staff-form.md; remove staff identifiers, HR complaints and customer details.
3. Put de-identified steps and timings in process-walk.md. If a permitted recording exists, review it locally and extract only process facts; do not upload an unreviewed transcript.
4. Write ten lines in shop-truth.md: what you sell, who you don't serve, tools you already pay for, what you refuse (no fake reviews, no auto-send of legal or clinical, no custom app yet).
5. Go to Claude. New project or a single chat. Upload the three files. Paste the prompt below. Run it twice if the first pass is mush. The second pass is where you say "kill anything that needs a new platform."

Copy-paste prompt (the one that outputs the scored list).

You are helping a local shop owner build a 90-day AI list. You are not a strategy consultant. You do not invent enterprise programs. You do not recommend a custom platform.

Read the files I uploaded:
- staff-form.md (staff answers)
- process-walk.md (one process, walked)
- shop-truth.md (what we sell, tools we already pay for, hard nos)

Build a scored list of 8–12 candidate workflows.

RULES
- Every candidate must hit at least one of: Effectiveness (makes money), Efficiency (saves named hours), Quality (customer can feel it). If you cannot fill one door honestly, drop it.
- Prefer the tools we already pay for. Default is buy/configure, not build.
- A named human approves and sends every customer-facing message. Never recommend customer-facing auto-send.
- Never recommend fabricating Google reviews or any public proof.
- Break the walked process into 5–15 micro-steps. Mark the ONE repetitive, time-expensive insertion point. That is usually this month.
- Kill anything that requires a new "AI operating system," a 12-month roadmap, or hiring a developer this quarter.

FOR EACH CANDIDATE INCLUDE
1. Title (plain English)
2. Door: Effectiveness / Efficiency / Quality (one or more)
3. Who does it today (role)
4. Frequency per week and minutes each — use the files; if missing, write GUESS
5. Money sketch (booked jobs, filled chairs, consults). Label hypothetical. No fake ROI.
6. Money (1–5), Time (1–5), Customer (1–5): 1 = no evidence of benefit, 3 = plausible benefit with a measurable baseline, 5 = repeated material benefit supported by our records. Score Data risk Low/Med/High and Effort S/M/L separately. Unknown evidence stays UNKNOWN, not a confident score.
7. This month / next month / later / never
8. Least that pays: 14-day version in existing approved tools and a one-page SOP
9. Evidence: 1–2 de-identified process observations; do not expose a respondent
10. Risks if the model is wrong (price, clinical, legal, fake-sounding reply)

THEN OUTPUT
A) A markdown table ranked by Money + Time + Customer. Use lower effort as a tie-break. High data risk requires a privacy/qualified-owner review before it can be selected; do not let a high total override this gate.
B) A "never" list.
C) This month: exactly one workflow, with owner, definition of done, and the micro-step we automate — not the whole process.
D) Next month: exactly one backup.
E) Exactly 1 BUILD, 1 CANCEL and 1 POLICY: one bounded workflow, one candidate tool/process to retire after dependency checks, and one written rule that closes a data or approval gap. State evidence, owner and due date for each. A cancellation is a proposal, never permission to cancel an account.
F) Questions I still need to answer before we spend money.

Do not pad. If the files only support six real candidates, give me six.

Turn the CANCEL proposal into a checked decision with Tool Stack & AI Spend Audit. Turn POLICY into the one-page AI Use Policy. Keep BUILD to one workflow.

Run a second pass:

Beat this up. Cut anything I would not notice in 30 days. Cut anything that needs a new vendor. Make "this month" smaller. If two items are the same job in different clothes, merge them. Then rewrite the one-pager from Step 5 using only what survived.

Definition of done. You have a table you argued with. You circled one "this month." You do not have a 12-month Gantt chart.

Step 5 — Freeze a one-page roadmap (this month / next month)

Why. A Notion database of AI projects is how this becomes homework. A one-page doc is how a shop actually moves. This month is the only real commitment. Next month exists so you don't panic-add a second project in week two.

Clicks.

1. Open a Google Doc. Title: [SHOP] — this quarter's AI list — [MONTH YEAR].
2. Paste the skeleton below. Fill it from Claude's table, in your words. If a sentence sounds like a vendor, rewrite it.
3. Print it. Put it where the board already lives — dispatch wall, front desk, the fridge in the back. Digital-only lists die.

One-page skeleton.

SHOP: __________
CHAMPION: __________ (owner)
HELPER ON THE GROUND: __________ (office manager / dispatcher / intake)
DATE FROZEN: __________

THE FILTER
We only do work that makes money (Effectiveness), gives hours back (Efficiency), or makes the customer feel the difference (Quality). If it hits none, it waits.

THIS MONTH (one)
Workflow:
Door it hits:
Person who runs it every Tuesday:
Least that pays (14-day version):
Definition of done in 30 days:
What we will NOT automate:
Tool we already have:

NEXT MONTH (one, only if this month shipped)
Workflow:
Why it waits:

LATER PILE (parking lot — not a promise)
-

CANCEL (one proposal; verify export, dependencies and renewal date before acting)
Tool/process:
Owner and decision date:

POLICY (one written rule; owner approves before use)
Rule:
Owner and review date:

NEVER
- Fake or paid Google reviews
- No customer-facing auto-send; a named person approves and sends every message
- Custom software before the cheap loop works
-

SCOREBOARD WE WILL ACTUALLY LOOK AT
Number: __________
How we'll see it without a new dashboard: __________

Definition of done. The page is filled. One workflow for this month. One person besides you knows they own Tuesday. The never list is on the same page.

Step 6 — Execute the least that pays

Why. The list is not the product. The first dull step is. Automate the repetitive, time-expensive micro-step — not the pet project, not the whole chain. A first-draft estimate email is a win. An AI dispatcher that runs the company is a science fair.

Default: buy or configure. The seat and the software you already pay for. Custom software waits until a cheap loop has been used for real. Don't start there.

Clicks.

1. Write the micro-steps of this month's workflow on the one-pager. Number them.
2. Circle one insertion point. Example: "tech notes → customer-facing estimate email." Not "the entire job-to-cash cycle."
3. Write a one-page SOP: trigger, paste, prompt, human edits, human sends, where the file lives.
4. Run it on the next five real jobs or visits. Not a demo.
5. After five, ask the person who ran it: did this give time back, make money, or make the customer better off? If they hesitate, the insertion point is wrong or the prompt is vague. Fix that. Do not add a second workflow.

SOP starter prompt (paste into the same Claude project, with the shop files).

Write a one-page SOP for this month's workflow: [NAME THE MICRO-STEP].

Audience: [office manager / dispatcher / front desk], not a developer.
Must include:
- Trigger (when they open the tool)
- Exactly what to paste
- The prompt they run
- What they check before a human hits send
- What they never let the model do (pricing promises, clinical claims, legal advice, fake reviews)
- Where they save a good example so next week is easier
Keep it on one page. Short sentences.

Definition of done. Five real runs. A human sent all five. You can say whether the door you picked (money, time, or CX) moved in a way you can see. The one-pager still has only one "this month."

Hypothetical worked example

Hypothetical. Not a client story. Numbers are napkin math so you can see the shape.

Shop: four-truck HVAC, owner still doing estimate follow-up at 9 p.m. Staff form comes back from the office manager, two techs, and the dispatcher. Copy-paste shows up in four answers: the "thanks for the visit, here's what we found" email. The walk is 45 minutes on Tuesday. Micro-steps on the pad:

1. Missed call or web form
2. Dispatch
3. Tech on site, photos, notes in the phone
4. Notes sit until the owner is in the office
5. Owner types the estimate email from memory and a price book
6. Follow-up if they don't reply — sometimes Friday, sometimes never

Claude's first table tries to sell a custom dispatcher and a review engine. Owner runs the second prompt. Survivors: (A) same-afternoon estimate draft from tech notes, (B) follow-up on unclosed quotes at 24 hours, (C) a human-sent missed-call follow-up that offers a diagnostic booking path. Scores: A hits Efficiency hard and Quality (complete estimate, first time). B hits Effectiveness. C hits Effectiveness and is probably next month because the after-hours vendor is a separate install.

This month: A. Least that pays: office manager pastes tech notes plus the never-discount / permit-line rules into Claude, edits five minutes, owner approves over a number they already use, she sends. No new platform. Hypothetical: eight of those a week at 25 minutes saved is evening time back — a sketch, not a case study. They'll know in 30 days if estimates leave the same afternoon. Next month: unclosed-quote follow-up, only if A shipped. Never: auto-send of a price, fake reviews, an AI dispatcher app.

Operator Standard check

Trace: Trace one real process without exposing records or grading staff.

Choose: Score Money, Time and Customer alongside Data risk and Effort; pick one build, one cancel proposal and one policy.

Build: Use the existing tools for a small draft-and-review workflow with a named owner.

Prove: Compare the chosen outcome and time spent at weeks 1 and 4; stop if risk or rework outweighs value.

Failure modes

The owner delegates the champion role. The office manager collects forms. Nothing gets picked. You get a folder called "AI ideas" and a new login in January.

The form has 22 questions. Staff ghost it. Five questions or paper.

You walk five processes. You learn nothing deeply. One process. 45 minutes.

You treat Claude's first table as the plan. It will flatter you with twelve initiatives. Beat it up. One this month.

You pick the glamorous step. A website chatbot while estimates still sit two days. Insertion point is the dull, expensive micro-step.

You skip the never list. Then someone auto-sends a clinical guess or a legal sentence. The filter failed in public.

You measure logins. Measure same-afternoon estimates, filled chairs, consult recaps, review replies inside 48 hours.

You start a custom build because the SOP felt too small. If the cheap loop is ignored, the app will be ignored with a nicer screen.

FAQ

Do I need every staff member to fill this out? No. People who do the work and have been here long enough to be honest. Two good answers plus one walk can freeze a one-pager. More answers help you see copy-paste. They do not replace the walk.

What if staff are scared this is about cutting heads? That's why Step 1 exists. Say it. Mean it. Pick a first workflow that gives them the ugly copy-paste back, not one that "replaces the front desk." If you were using this as cover to cut someone, don't. It will poison the form.

Can I skip Claude and just pick the thing I already hate? Yes if the leak is obvious — missed calls, unclosed estimates, empty chairs. Claude is there so you don't ignore the copy-paste the office manager is drowning in while you chase a receptionist demo. Use the prompt. Then override it as the owner.

We already bought [tool]. Does this mean we wasted the money? Maybe the tool is fine and the workflow was never written. Step 6 starts in software you already pay for. If the scored list says it hits none of money, time, or CX, stop feeding it. Don't buy the next one.

If it goes wrong (kill switch)

The owner can pause the pilot immediately: stop pasting, remove the workflow from the team’s active instructions, and return to the prior manual SOP. Do not cancel a production tool from a score alone. If a bad draft reached a customer, the named sender tells them which statement was wrong, gives the verified correction and next step, and logs it. Escalate privacy, clinical or legal errors to the qualified lead before contacting anyone.

Prove it: 4-week scoreboard

Record the baseline before changing the workflow. Use actual shop counts; leave blanks until measured.

Save it as a skill

Save the final prompt and never-list in skills/three-outcomes-filter.md in the shop’s shared drive, or the approved project instructions your AI workspace supports. Name its owner and review date. The file is reusable guidance, not a grant of sending or payment access.

ROLE: Help the owner choose one useful workflow.
READ: company.md, customer.md, rules.md, de-identified process notes.
PROMPT: Score Money, Time, Customer 1–5; Data risk Low/Med/High; Effort S/M/L. Mark unknown evidence. Produce 1 BUILD, 1 CANCEL proposal, 1 POLICY, each with owner, due date, baseline and a four-week decision.
NEVER: invent ROI; grade staff; expose records; cancel accounts; make customer-facing sends.
APPROVAL: Owner decides. Named person reviews and sends any customer message.
STOP: Missing evidence, high data risk or conflicting rules → draft questions and ask.
OWNER / REVIEW DATE: [name] / [date]

Read the reasoning in Three Outcomes and choosing the first workflow.

Screen-recording outline (8–10 minutes)

Record on a fake or anonymized shop.

1. 0:00–0:40 — Voiceover: "You're the owner. Filter, not logins. Money, time, CX. One workflow this month."
2. 0:40–2:00 — Screen: the champion email. Read the two promises. Send.
3. 2:00–4:00 — Screen: Google Form. Build the five questions live. Show the link going into the email.
4. 4:00–5:30 — Screen: a legal pad or a simple list of micro-steps from a 45-minute HVAC estimate walk. Point at the stall ("notes sit until the owner types").
5. 5:30–7:30 — Screen: Claude. Drop in three files. Paste the scored-list prompt. Scroll the table. Run the "beat this up" follow-up. Circle one this-month row.
6. 7:30–8:30 — Screen: the one-page Google Doc, filled. This month / next month / never.
7. 8:30–9:30 — 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.