How to use ChatGPT at work: setup, ten workflows, and what to avoid

Start with the right account

Which ChatGPT you use matters more than which prompt you type. There are three situations.

A personal account (Free, Plus, or Pro). Fine for learning, for drafts that contain nothing confidential, and for your own career work. Consumer plans have their own defaults about whether your conversations can be used to train models, and you are responsible for the settings.

A workspace your company pays for (Team or Enterprise). Conversations are not used for training by default, an admin controls who is in the workspace, and you can share custom GPTs and Projects with colleagues. If your company offers this, use it for all work, even the boring work, because it is the account your policy covers.

No approved account at all. Then the rule is simple: never paste confidential company or client data into a consumer AI tool unless your organization has approved it. Use ChatGPT for the parts of your job that involve public information and your own thinking, and ask your manager or IT what is planned.

If you work in healthcare, law, finance, HR, or insurance, treat anything with PHI, privileged material, PII, or regulated financial data as off-limits on a personal account, full stop.

Three settings to check in the first five minutes

1. Data controls. On a personal account, open Settings, then Data controls, and look at the toggle that lets OpenAI use your content to improve its models. Decide on purpose rather than by default. On Team and Enterprise this is handled for you.
2. Temporary Chat. The toggle near the model picker starts a conversation that is not saved to history, does not use or update memory, and is not used for training. Use it for one-off questions about anything sensitive, and for experiments you do not want polluting your memory.
3. Memory. Under Settings, then Personalization, ChatGPT can save facts about you across conversations and can reference your past chats. This is useful (it stops you re-explaining your role every morning) and also a liability (it can carry a client's name from one chat into an unrelated one). Open the memory manager, read what it has stored, delete anything you would not want repeated, and turn off chat-history referencing if your work involves many unrelated clients.

Custom instructions: the ten-minute investment

Also under Personalization. Tell ChatGPT who you are and how you want answers. Something like: "I am an operations manager at a forty-person logistics company. Default to plain American English, short paragraphs, no walls of bullet points. When I ask for a draft, give me one version, not three. If my request is ambiguous, ask one clarifying question before answering." Every chat inherits this. Revisit it monthly.

Ten workflows that pay for the subscription

Each of these takes under ten minutes once you have done it twice. Short prompts are shown inline; a full worked example follows later.

1. Turn bullet points into a finished email

Paste five bullets and the recipient's situation. "Write this as an email to a customer who is annoyed about a late delivery. Apologetic but not groveling, under 150 words, one clear next step." Then edit the opening line; that is the part that most often sounds machine-made.

2. Rewrite for a different audience

Paste a technical update. "Rewrite this for the executive team: what changed, what it costs, what you need from them. Three short paragraphs." The same paragraph can be re-cut for a customer, a new hire, or a board.

3. Summarize a long thread or document

Upload the PDF or paste the thread. "Summarize in five bullets, then list every open question and who owns it. Quote the sender for any commitment." Verify anything you will repeat as fact.

4. Prepare for a meeting

"I have a thirty-minute call with [PERSON, ROLE] about [TOPIC]. Background: [PASTE]. Give me a three-item agenda, the two questions I should ask first, and the objection I am most likely to hear." Afterward, paste your rough notes and ask for a follow-up email with action items.

5. Turn messy notes into a status update

Paste the notes from your week. "Write a Friday status update for my manager: done, in progress, blocked, and what I need decided. Keep it to one screen."

6. Analyze a spreadsheet

Upload the CSV or Excel export. Data analysis mode writes and runs Python, shows you the code, and can produce charts and a cleaned file you can download. Ask it to describe the columns first, then ask your questions. Check the totals against the source.

7. Research with sources

Turn on web search for a quick question with citations. Use deep research when you want a longer cited report that takes several minutes: "Compare the three main vendors for [CATEGORY] for a company of [SIZE], covering pricing structure, integrations, and what customers complain about. Cite sources." Open the citations; a dead or misread source is common.

8. Explain something you do not understand

A contract clause, an error message, an accounting term, an acronym in a vendor's email. "Explain this the way a patient colleague would, then tell me what question I should ask the person who sent it."

9. Pressure-test a plan

Paste the plan. "Act as a skeptical CFO. Find the three weakest assumptions, explain why each could fail, and tell me what evidence would change your mind." This is the most underused workflow on the list.

10. Write the formula, macro, or script

"Write an Excel formula that returns the number of business days between [COLUMN A] and [COLUMN B], excluding the holidays listed on [SHEET NAME]." Or ask for a short Python script that merges twelve monthly exports into one file. Test on a copy.

Files and data analysis

Upload the file rather than describing it. ChatGPT reads PDFs, Word documents, PowerPoint decks, spreadsheets, CSVs, and images, including screenshots of a dashboard or a photo of a whiteboard. Drag the file into the message box or use the attachment button.

Three habits make file work reliable:

• Ask for structure first. "List the sheets and columns and tell me what each seems to contain." You find out quickly whether it read the file correctly.
• Ask it to show its work. For numbers, "show the calculation" reveals whether it summed the right column.
• Spot-check against the source. Open the original to the page or cell the answer cites. Scanned PDFs, merged cells, and tables that span pages are where misreads happen.

For repeated analysis of the same kind of file (a weekly export, a monthly report), build a Project or a custom GPT with instructions on how to read it, so you stop re-explaining the layout.

Projects and custom GPTs

Projects are folders. Each holds files, a set of instructions, and the chats that belong to that piece of work. A Project called "Q4 vendor renewal" can hold the contracts, your negotiation notes, and the instruction "always quote clause numbers." Every chat inside starts informed. Projects are for you, and on Team and Enterprise they can be shared with colleagues. Check the project settings if you want its chats kept separate from your general memory.

Custom GPTs are packaged assistants. You write instructions, attach knowledge files, optionally connect actions to outside systems, and share the result with a link, within your workspace, or publicly. A GPT is right when other people should use it without seeing how it works: "Proposal formatter," "Onboarding Q&A," "Tone checker for customer emails."

The decision rule: Project for your own recurring work, GPT for something you hand to others. Both are covered in Custom GPTs vs Claude Projects vs Gemini Gems and How to build a custom GPT.

On business plans, connectors to Google Drive, SharePoint and OneDrive, Dropbox, and other work systems let ChatGPT search your files without uploads. Which connectors you get depends on your plan and on what your admin has switched on.

Voice, mobile, and the desktop app

The mobile app is the part most people underuse. Voice mode holds a real spoken conversation, which is the fastest way to think out loud on a commute: "Help me decide how to structure tomorrow's kickoff." Ask for a written summary at the end and it lands in the chat history. The camera lets you photograph a form, a label, or a page and ask about it.

The desktop apps for Windows and Mac add a keyboard shortcut that opens a small chat window over whatever you are working on, which beats switching to a browser tab. On some plans the desktop app can also record a meeting and produce notes; if you use it, get consent from everyone in the room and check your company policy first.

Voice and camera are convenient and easy to misuse. Do not dictate customer details in a coffee shop.

What to avoid

• Trusting a number, name, or quote without checking. Verify every fact that leaves your hands. See How to fact-check AI output.
• Pasting what you should not. Credentials, customer PII, unreleased financials, anything under NDA, unless your plan and policy cover it.
• Letting it be the last reader. For anything with legal, medical, or financial consequences, a human who understands the consequences signs off.
• Uploading a file and asking a vague question. "Thoughts?" gets a generic answer. Ask for what you actually need.
• Mental math. Use data analysis mode for calculations, since it runs code, rather than asking the model to add columns in its head.
• Sending the first draft. Read it aloud. Cut the filler sentence at the top and the summary sentence at the bottom; that alone makes most drafts sound human.

A worked example: the weekly sales export

You are a sales manager. Every Monday you get a CRM export with a few hundred rows: rep, account, stage, amount, close date, last activity. You want a one-page summary before the 9 a.m. pipeline call.

Create a Project called "Monday pipeline." Put a short instruction in it describing the columns and what "stalled" means at your company. Then each week, upload this week's and last week's exports and paste this:

You are a sales operations analyst. I am attaching this week's pipeline export from our CRM, plus last week's for comparison. The columns are described in this project's instructions.

Tasks:
1. Describe the data first: row count, date range, and any columns with missing values. Do not skip this step.
2. Total pipeline by stage and by rep, in a table.
3. List deals over [DOLLAR THRESHOLD] with no activity in the last [NUMBER] days, sorted by amount.
4. Compare this week's total to last week's and name the five biggest movers up or down.
5. Write three sentences I can say at the top of the pipeline call.

Format: tables for 2 through 4, plain text for 5. Show the calculation for any total. If a column looks different from what the instructions describe, stop and ask me before continuing.

Read the row count against the export. Open two of the "stalled" deals in the CRM to confirm the activity dates. By the third week this is a ten-minute job that used to take an hour, and the Project remembers the column layout so you never explain it again.

Next steps

• Compare the alternatives in ChatGPT vs Claude vs Gemini vs Copilot and read the ChatGPT tool page.
• Build your first packaged assistant with How to build a custom GPT.
• Read AI privacy at work before you upload anything sensitive, and learn the terms AI memory and hallucination.
• Grab ready-made prompts for writing and email and data and spreadsheets.
• Sharpen your prompts with the prompt builder or go deeper with the course.