Your own audit

Section A · Map your repetitive text work

List jobs where the words repeat more than the judgment.

Prompts to jog memory

• What emails do you send every week that start the same way?
• What reports are 70% last month's report with new numbers?
• What social captions, menu notes, donor updates or recall reminders follow a pattern?
• What intake forms get re-typed into nicer language?

Rule of thumb. If a trained high-school hire could draft it with a good checklist, it belongs in Section A. If a mistake could cost a license, a grant or a lawsuit, park it until Sections C and F are clear.

Section B · Map judgment-only work

List jobs where the value is the decision, not the paragraph.

Examples from the five shops, for calibration, not copying

• Ray: which route gets the last emergency slot on a 110° day.
• Dr. Ahn: clinical recommendations and anything that smells like diagnosis.
• Teresa: who gets scarce housing help when the list is longer than the rooms.
• Ruiz family: cutting a shift when labor is high but a regular event is tonight.
• Hartnett & Cole: fiduciary advice and final document language.

If you catch yourself writing “AI could probably do this” next to a Section B item, rewrite the sentence as “AI could hide who is accountable.” Then leave it human.

Section C · Map sensitive data

Check every type of data that appears in your weekly workflow.

• Customer, patient or client names and contact info
• Payment details, account numbers, SSNs, Tax IDs
• Health information, even “just scheduling”
• Children's information
• Employee records, wages, disciplinary notes
• Donor or grant beneficiary case notes
• Attorney work product, confidential client facts
• Passwords, badge codes, vendor logins
• Unpublished financials
• Other

If you cannot name the person in question 2, your answer to question 1 is No until you can.

De-identify checklist, use before any draft help

• Names removed or replaced with role labels (Patient A, Donor 12)
• Dates shifted or generalized if they identify someone
• Account, chart or case numbers removed
• Addresses and school or employer details removed
• Unique story details that only one person would know removed
• Second human scanned the prompt before submit

Section D · Baseline metrics, two weeks

Pick three to five metrics. More than five and you will quit. Example — metric: minutes on after-hours text triage. Count: timer plus tally.

Suggested metric menu, steal only what fits

• Minutes spent rewriting the same email type
• Number of no-shows or late cancels
• Hours to produce the weekly schedule
• Hours to produce the monthly donor or board update
• Number of review replies older than 48 hours
• Count of wrong-list or wrong-person contact errors
• Hours spent fixing incomplete field logs
• Number of drafts that needed factual correction before send

Do not track feelings as your only metric. “Team seems happier” is allowed as a note, not as proof.

Section E · Red-flag list

If any of these show up, pause the experiment the same day.

• Real sensitive data pasted into a tool that is not approved in writing
• A citation, statistic, case cite or policy claim you did not verify from a primary source
• Customer-facing copy published without a human tone check
• Staff saying they feel watched, replaced or set up to fail, and leadership shrugging
• A near miss in legal, medical, financial or safety language
• Owner radio-silence after announcing a new process
• Lists used for mail, email or text not reconciled to the current system of record
• Free-tier output treated as grant-ready or court-ready
• Family or partner conflict turning into secret pilots
• Anyone saying “the AI said so” as a reason to skip review

Section F · Go / no-go thresholds

Fill these before week 1 ends. Do not invent them after you see results you like.

Go, proceed to a tiny pilot, if all are true

• Sections A, B and C complete
• At least 10 business days of Section D data captured
• One repetitive text task identified with a named human reviewer
• Sensitive-data rule posted and initialed by the owner

No-go, do not pilot yet, if any are true

• Owner will not put their name on the data rules
• Staff who must review drafts are already past capacity with no relief
• The only proposed task is judgment-only (Section B)
• You cannot explain the pilot in one paragraph to a new hire
• Week 1 and 2 baseline is missing or fake
• A red flag from Section E is already active

Pilot scope sentence, write one

For days, we will use AI only to draft . A human named will review every output before it is used. We will not touch data. We will stop if .

If that sentence is hard to write, you are not ready. That is a successful audit result. Readiness is the point.

One-page Monday starter

If you are overwhelmed, do only this tomorrow morning.

1. List five repetitive text tasks (Section A).
2. Circle one that does not involve sensitive data.
3. Start a two-week tally of how long that task takes.
4. Write the sensitive-data rule in six lines and tape it near the keyboard.
5. Put a 30-minute meeting on the calendar for Day 16 titled Go / No-Go. Invite the people who actually do the work.

That is the whole job for week one of your own audit. Your version only needs to be honest.