Job page

What not to paste into ChatGPT at a small business

Someone on the team already pasted a customer list. You still want to use the tool. You do not want a forty-page policy.

This is a one-page rule set. Three buckets. Public. Internal. Never. You keep using AI. You stop feeding it names, money, and IDs on a free account.

Is it safe to put customer data in ChatGPT? Not like that. Not on a shared tab. Turning training off is not zero risk. Treat that as a practical note, not a legal brief.

We design the work the AI does and the system that keeps it honest. The Clean beat is the feed: what the system can access and what it will not. Stay is teaching the team so the next hire does not upload the export again.

You can keep using it

The fear is real. The need is also real. Staff already draft emails, rewrite a service description, or ask "how do I say this" in a chat box. Banning the tool does not work. They will use it on a phone in the parking lot.

Write the line instead.

Use it. Bound it. Do not dump the books.

An agent on a job you named is different from a shared ChatGPT tab. A tab has no owner, no field list, and no "will not access." A feed does. That is the whole difference.

Three buckets

Print this. Tape it near the computer people actually use.

Public: fine

Anything you would pin on the door, put on a flyer, or read to a stranger on the phone.

• Hours, address, how to book
• Service names and public prices
• How you describe a cut, a repair, a package
• A blank invoice or RO template with no customer on it
• A process written so a new hire could follow it, with names stripped

If it is already on the street, a model seeing it does not create a new problem. Draft the rewrite. Then you post it, not the chat.

Internal: anonymize, then ask

Process, patterns, and stripped examples. Not a person. Not a live ticket.

• "A job waiting on parts more than five days. What do we tell the customer?" with no name, no phone, no amount
• Your refund rule, late policy, how you quote a roof, opening checklist
• A sample invoice line: SKU, qty, unit price. Customer column deleted. Invoice number swapped for INV-EXAMPLE
• Labor versus revenue as totals for a week, not payroll for a named person
• A count: 40 quotes still sent, oldest 18 days. No export attached

Ask first if a real person is still recognizable. A rare job in a small town plus a date is often enough to name them. If you would not pin it on the break-room wall, it is not internal. It is never.

Never: names + money + IDs on a free account

Do not upload a customer list to AI. Do not paste the export. Do not "just this once."

• Customer names with phones, emails, or addresses
• Full invoices, repair orders, or estimates with a person on them
• Card numbers, bank details, government IDs, dates of birth
• Payroll for a named person, offer letters, write-ups
• Health notes, kids in the household, anything a lawyer would ask for
• Passwords, vendor logins, other people's price lists
• The whole Drive, the whole mailbox, the whole POS export "so it has context"

Names plus money plus IDs is the stack. Any two of those on a free account is already too much. All three is how shops get into trouble they cannot unsay.

A paid business account with a contract is still not a dump. It is a different risk, not zero. If the job needs real customer rows, do not paste them. Curate a feed you control. How to clean that feed is the field version: one job, six fields, a hard line around the rest.

Shop floor, not a seminar

Three things that hit the paste box in a real week.

The invoice. Someone wants a cleaner description, or a late-fee sentence, or "make this look less angry." Paste the template or one redacted line. Not the PDF with the customer, the phone, the last four of the card, and the balance. You already know how to say the work. You do not need the model to see who owes you.

The repair order. Status is the job: received, in progress, waiting, ready, picked up. You can ask, "Customer keeps calling. Job is waiting on a part. What do we say?" You do not paste the RO with the VIN, the cell number, and the promise date next to their name. Stage and waiting-on are enough. If status lives on a whiteboard, type the stage. Photographing the wall into a chat is still a dump.

The customer phone. Follow-up copy is a good use. The list is not. Write three example texts with {{Name}} and {{Job}}. Do not upload last quarter's Booksy or POS export so the model can "personalize." Personalizing from a live list is how a free account gets every regular you have. If you need follow-up to run on real quotes, that is an agent job on a bounded sheet, draft-first, with you still sending.

Same instinct as high-volume ops: messy item names create rework. A paste of the whole file creates a second, worse mess, one you cannot pull back. Invoice work at production load only holds when the fields mean the same thing and the wrong fields stay out. A three-person shop is not a warehouse. The boundary is the same size: named columns in, everything else out.

Turning training off is not zero risk

Providers publish terms. They change. Staff will not read them. You should not copy a lawyer's page onto the break-room wall and call it a rule.

The owner rule:

1. A free or personal ChatGPT (or any consumer chat) is for public and stripped internal only.
2. Switching "do not train on this" to on is not a vault. The prompt still left your shop. Support, bugs, logs, a screenshot, a shared laptop. None of that is zero.
3. If you would not email the file to a stranger to "take a look," do not paste it.
4. If the job needs the real rows, it needs a system you control, not a tab.

No percentages. We have not measured your risk. Nobody honest has a number for "how likely is a leak from one paste." The practical move is not to paste.

The card you print

Fifteen minutes. One page. Owner writes it. Everyone who uses a chat tool sees it.

Read it out loud once. If a sentence needs a lawyer, you are writing a policy. Cut it. This is AI boundaries for a small team, not a compliance binder.

Put a copy in the new-hire packet. Stay is teaching. A feed goes stale. A new person will paste unless you said not to.

If it already happened

Do not run a shame meeting.

• Stop. The file does not go in again.
• Note what it was: list, invoices, one RO, a screenshot of the board.
• If it was cards, IDs, or passwords, treat it as you would any other credential mistake. Rotate what you can, tell the person who already advises you on that kind of problem.
• Write the card the same day. The paste is how you found out you had no rule.

You cannot unsay a prompt. You can stop the next one.

Clean + Stay, not a scare story

Boundaries are a curated feed. Not a lecture.

Clean and organize: pick the job. Name the fields. Name the files. Name the date range. Name will-not-access. Deduplicate the key. Status is a list. The model sees that, and only that.

Stay: who may change a rule. What to do when the system does not know. When to take the phone. The printed card above is the first Stay artifact. Care, later, is watching the connections you agreed to, not watching every chat tab.

Connect → Clean and organize → See → Do → Stay. You do not need all five beats to stop a bad paste. You need the card, then a feed if the job is real.

We build it. You help hone it for your business. You direct it. Hone, here, means the first weeks you catch a draft that still has a phone number in it, and you cut it. The system learns the cut. You still own the exception.

If you want a feed instead of a shared tab

A PDF of this page is not a system. A shared ChatGPT tab with a stern sticky note is not a system.

If the work needs real quotes, real status, real invoices (bounded, logged, directed), that is an Intelligent Business Layer. Priced in writing around how you actually run. Not a chatbot dropped on a folder.

A website can be a front door later. It is not what this page is for.

How we work if you want the short version of who is talking.