ChatGPT use cases for business: what actually works

Sales and marketing

• 1. Proposal and quote first drafts. Paste your last three winning proposals in as examples, then give it the notes from the new job. What makes this work is the examples: without them you get generic consultant-speak, with them you get something in your voice that needs twenty minutes of editing instead of two hours of writing.
• 2. Follow-up sequences. Give it the deal context, the objection you heard and the tone you want, and ask for three follow-ups spaced over a fortnight. The value isn't the writing, it's that the follow-ups get sent at all.
• 3. Repurposing one asset into five. A case study becomes a LinkedIn post, an email, a landing page section and a sales one-pager. Do it in one conversation so the facts stay consistent, and never let it invent a number that wasn't in the source.
• 4. Ad and headline variants. Ten versions in a minute, of which two are usually worth testing. Judge them against your own data, not against how clever they sound.

Operations and admin

• 5. Writing the SOPs nobody has written. Record the person who actually does the job talking through it, paste the transcript in and ask for a numbered procedure with the decision points called out. This is the single highest-return use I see in small businesses, because the knowledge was already there and just never got written down.
• 6. Meeting notes into actions. Feed it the transcript and ask specifically for decisions, owners and deadlines, not a summary. Asking for a summary gets you a wall of text; asking for decisions gets you something you can paste into a project tool.
• 7. Job ads, position descriptions and onboarding packs. Documents that take half a day and get written once every two years. Give it your existing ones as the pattern and it does the structural work.
• 8. Turning messy input into a clean document. Voice notes from the car into a scope of work, a photo of a whiteboard into a project brief, five email threads into one client update.

Customer service

• 9. Reply drafts from your own answers. Give it your policies and your best previous replies, then have it draft responses a human sends. Drafting is the safe version of this; letting it reply to customers unsupervised is not, and the gap between those two is where most support automation goes wrong.
• 10. Rewriting the hard email. The one about the overdue invoice or the mistake your team made. Ask for firm and professional, then read it once more yourself before it goes.

Analysis and decisions

• 11. Questions against your own spreadsheet. Upload the export and ask what moved and why. Treat every answer as a lead to verify in the source system rather than a finding you can act on, because it will occasionally read a column wrong and tell you so with total confidence.
• 12. Pressure-testing a decision before you make it. Describe the situation and ask what you've failed to consider, or ask it to argue the opposite case. Not for the decision itself, for the blind spots.

When one of these gets used the same way every week by several people, that's the moment to package the instructions and reference material into a custom GPT so nobody has to remember the prompt.

Use cases to avoid

These are the ones that create work rather than save it, or create risk you can't see until it lands.

• Regulated advice as a final answer. Tax, legal, financial, medical. It can help you understand a topic or draft an explanation for a qualified person to review, and that's where it stops.
• Final legal or contract wording. A draft to hand your lawyer, fine. Anything you sign without a lawyer reading, no.
• Client-facing output nobody reads first. The moment output goes straight out the door, one wrong figure becomes a customer problem.
• Facts, figures and citations it produced itself. If a number or a source matters, it comes from your systems or a page you've opened yourself.
• Hiring and performance decisions. Summarising applications to save reading time is reasonable. Ranking people, or deciding who gets an interview, is not something to hand over.
• Confidential material in personal accounts. Not a use case problem, an account problem. Fix it with a business workspace and a one-page written rule about what may be pasted in.

How to pick your first three

Score each candidate on four things: how often it happens, how long it takes now, how tolerant it is of a rough first pass, and how bad a mistake would be. The winners are frequent, slow, draft-friendly and low stakes. Start with three, give them to the people who own that work, and check in a month later on whether they're still doing it.

Costs and plan tiers are covered in ChatGPT for business. If you'd rather have someone work through your specific list with you, that's what the 1:1 session is for. It's advice only: I help you decide what to use AI for and in what order, and anything that needs building afterwards is a separate engagement with its own scope, price and contract.