The thread is not a filing cabinet
Someone on the team pasted a customer's full record into a chat "just to summarize it," and now the summary is clever and the record is in a consumer product the company never reviewed. That is the failure this guide is written about. ChatGPT can draft, reformat, and explain. It is not your system of record, your lawyer, or your accountant. Used with a source you control and a rule about what stays out, it removes blank-page time. Used as a place to think out loud with confidential data, it creates a problem your IT lead will not find funny.
OpenAI sells more than one plan, including business and enterprise offerings with different terms from the consumer app. Read the current terms on OpenAI's site for the plan you actually pay for. A screenshot of a setting from last year is not a policy. This page is not legal advice.
Decide the job list, in writing
A business does not need twenty use cases. It needs the ones that repeat and that survive a privacy read. A durable list looks like this:
- Turn notes you already wrote into a follow-up email.
- Rewrite a paragraph for a reader you name, without new facts.
- Propose an outline for a document whose headings a human will approve.
- Explain an error or a spreadsheet column to the person who owns the file, after identifiers are removed.
- Draft a checklist for a process the team already follows, which the team then edits.
Notice what is absent. The model does not set prices, approve discounts, hire, fire, diagnose, or tell a customer their case is resolved. Those are decisions. Decisions stay with people who have the authority and the information.
Write the allowed list on an internal page. If a task is not on the list, the person asks before they paste. The small business guide is the shorter version of this page for a shop that does not have an IT handbook yet. The rule is the same.
What never goes in the box
Make a deny list that is boring and specific.
- Passwords, keys, tokens, and recovery codes.
- Full payment card numbers and government identifiers.
- Health information, unless counsel and the plan's terms say the workflow is acceptable. When in doubt, leave it out.
- Customer data you can replace with "Customer A" and a description of the issue.
- Unreleased financials, board materials, and anything marked under a contract.
If the task needs the real identifier, do the task in the system that already stores it, or strip the identifier and keep the problem. "Customer A was double-charged for the annual plan and the receipt shows two lines" is usually enough to draft a reply. The email address can be added by the human at send time.
Read your own privacy commitments to customers, and the vendor's. If you promised customers you would not share their data with a model provider, a helpful employee can break that promise in a minute. The deny list is how you stop that minute.
A pattern that produces usable drafts
Use the four-part brief from the prompt guide: job, source, shape, limits. For business mail, the source is the note you took, not the model's memory of "how companies sound."
Example you can adapt:
"Draft a reply from a support lead to a customer. Source: we billed twice on May 2, we will refund the second charge, the refund takes three to five business days, we do not know why it double-posted yet. Shape: under 130 words, no bullet list, no apology longer than one clause. Limits: do not offer a discount, do not blame a person by name, do not invent a cause."
Then the human checks the refund window against the real policy. If the model wrote "immediately," and the source said three to five days, the human fixes it. The error is ordinary. The check is the job.
For documents, ask for headings first. Approve them in the meeting. Only then ask for prose under the headings you kept. This stops a long draft from arguing a strategy nobody selected.
Team conventions that keep quality up
Pick one workspace or one team account if the product offers it and your admin has reviewed it. Personal accounts for company work scatter data across people who may leave. Shared custom instructions can hold the voice and the deny list, and they are not a substitute for pasting the source on each task. Instructions drift. The source should be in the thread.
Agree on a review line for anything a customer will see. For the first month, a second person reads it. After that, sample. Keep a short channel or doc where people paste failures: the draft offered a discount, the draft invented a feature, the draft used a slur of corporate cheer. Add the failure to the limits block. This is training the team, not fine-tuning a model.
Do not require people to pretend they did not use the tool if your culture also requires them to use it. Disclosure inside the company can be as small as a footer on the doc: "Drafted with ChatGPT from the May 2 note, edited before sending." Clients get disclosure when the contract or the relationship calls for it. The freelancer guide treats that as a scope term. Employees should ask their manager if they are unsure.
Meetings, code, and other departments
Meeting notes belong to the meeting tool or the doc. A transcript, if everyone consented and your policy allows it, can be summarized into decisions and owners. The summary is useful when each action has a name and a date. It is useless when it is a polite paragraph. Ask for a table you can paste into the tracker, then have the facilitator correct it before it is official. The productivity guide is about that handoff.
Engineers using ChatGPT for code should follow the same review they would use for Cursor or Copilot: read the diff, run the tests, do not add a dependency you did not intend. Secrets do not go in the prompt, including keys that "are only in staging."
Analysts can ask for a formula or a critique of a chart title. They should compute the number in the sheet. A fluent explanation of a wrong total will be repeated in a meeting as if it were the total.
When Claude or another model is the better window
Some days the work is a long policy and a question about what it requires. Claude is often the window people prefer for that reading job. The comparison helps a team assign seats without pretending one subscription covers every department. Standardize the deny list across both. Do not standardize in a way that forces a long-document task into a tool that truncates the file and then guesses the rest.
If the company already lives in Microsoft or Google, look at the assistant bundled there before you add a third login. Bundled does not mean safe. It means the admin may already have a switch. Ask the admin.
A rollout that fits on one page
Week one: publish the allow list, the deny list, and one worked example from your own mail, with the private bits removed. Week two: everyone uses it on a real low-risk draft and brings one failure back. Week three: decide whether the plan you are on matches the data you are actually pasting. If people are pasting more than the plan allows, stop and change the plan or stop the pasting. Do not "see how it goes" with customer files.
Revisit the page when OpenAI changes terms, and when you change what you sell. The page is short on purpose. A handbook nobody reads is not a control. A deny list on the wall, and a person willing to say "take that out of the chat," is a control.
Your next draft starts from a note you wrote and a limit you can point to. If you cannot point to the source, you are not ready to open the window.