TL;DR
On this page
You typed "make me an invoice" into ChatGPT, filled in your hours, and got back a tidy-looking invoice in about ten seconds. Then you pasted it into Word to send it, and it looked like a college essay: a wall of text, a table with mismatched columns, no header, no logo. You are about to send that to a client paying you real money.
This walkthrough is part of the complete guide to generating client documents with AI.
The content the AI produced is usually fine. The problem is everything around it: the number it invented, the tax line it guessed at, and the fact that a block of chat output is not a document a paying client should receive. This post gives you a prompt that gets a clean draft, then the three things you fix before it goes out.
The prompt that drafts a real invoice
Paste this into ChatGPT, Claude, or Gemini. The rules at the bottom are what stop it from inventing the parts it has no business inventing.
You are an expert at creating professional freelance invoices. Create an
invoice from the details below.
FROM: [your name / business, address, email]
BILL TO: [client name, company, address]
INVOICE NUMBER: [your next number, e.g. 2026-014]
INVOICE DATE: [today]
DUE DATE: [e.g. Net 14 from today]
LINE ITEMS:
- [description] | [qty or hours] | [rate] | [amount]
- [add more lines]
PAYMENT TERMS:
- Accepted methods: [e.g. bank transfer, Wise]
- Late fee: 1.5% per month on overdue balances
Rules:
1. Lay it out as a clean invoice with the total and due date near the top.
2. State the payment terms, accepted methods, and the late fee explicitly.
3. Do NOT invent a tax rate. If tax applies, insert a [TAX %] placeholder
for me to fill in for my own jurisdiction.
4. Use the exact invoice number I gave you. Do not generate your own.
5. Keep it plain and professional. No marketing language.
The difference between this and a bare "make me an invoice" is rules 3 and 4. Those two lines stop the two errors that quietly cause the most trouble. Even with the rules in place, read the draft once against the three checks below before you send it.
The three things AI gets wrong on an invoice
1. The invoice number. AI has no memory of what you sent last month, so left alone it makes a number up, and a duplicated or out-of-sequence number breaks the paper trail you need if a payment is ever disputed. Stripe's guidance on what every invoice requires is blunt about why the number matters:
Assign a unique identification number to each invoice. This helps both parties reference the transaction in the future, in case of a dispute.
Source: Stripe, "Invoice Requirements"
Keep your own running sequence and feed the AI the next number every time. Never let it choose.
2. The tax line. Ask for an invoice and the model will often add a tax row with a confident percentage that may be wrong for your country, your client's country, or your registration status. The prompt above forces a placeholder instead. Fill the real rate yourself, or delete the line if you do not charge tax. A wrong tax figure is the kind of error that gets an invoice bounced back and your payment delayed.
3. The format. This is the one that makes you look amateur. The model returns text or a plain table, and pasting that into a Word document produces something that reads like a memo, not a business invoice. The look matters more than freelancers like to admit. As SolidGigs notes in its rundown of invoicing errors, the trouble usually starts with DIY tools:
I'm sure you've made at least one of these mistakes, especially when creating invoices in Excel.
Source: SolidGigs, "Common Invoicing Mistakes"
A polished invoice is not vanity. About 29% of invoices are paid at least a day late per Bonsai's data, and 44% of freelancers have been stiffed by a client at some point. A clear, professional invoice with firm terms is one of the cheapest ways to stay out of both groups, because it signals you run a real business and you track what you are owed.
Here is what AI hands you versus what a sendable invoice actually needs:
| Element | What the AI draft gives you | What it needs before sending |
|---|---|---|
| Invoice number | A made-up number | Your next number in sequence |
| Tax | A guessed rate, or none | Your real rate, or the line removed |
| Payment terms | Vague or missing | Due date, methods, and late fee stated |
| Format | A text blob or rough table | A clean, branded PDF |
| Totals | Usually correct, still check | Verified against your line items |
Which AI should you use: ChatGPT, Claude, or Gemini?
For drafting an invoice, the choice matters less than you would expect. All three produce the same kind of output from the same prompt: a structured invoice in plain text or a markdown table. Claude tends to return the cleanest tables, ChatGPT adds the most boilerplate you will want to trim, and Gemini sits between them. None of the three outputs a finished, branded PDF, and all three will invent an invoice number and guess a tax rate if you let them. The three rules in the prompt above matter far more than which model you paste it into.
Where the model choice does matter is privacy. If the billing details are sensitive, prefer the business or team tier of whichever tool you use, since those generally exclude your inputs from training, and keep real client data out of the free consumer tiers.
Tax deserves a specific word, because it is the field AI is most confidently wrong about. A US-based service provider often charges no sales tax at all; a UK provider who is VAT-registered must add VAT at the current rate; and a US provider billing a UK client should not add UK VAT to the document. The model has no idea which of these describes you, so it guesses. The payment terms guide covers how to phrase terms cleanly, and a field-by-field invoice template shows what the finished structure looks like once the tax line is right.
Speed is the point, so do not lose it to formatting
The reason to use AI for invoices is speed. FreshBooks' 2026 survey of small operators found that those who spend five or more hours a month on invoicing report cash-flow problems at 28% versus 10% for those who keep it shorter. Slow, manual invoicing is correlated with getting paid slowly. AI drafting helps, but only if you do not then burn the time you saved rebuilding the layout in Word on every single invoice.
pro tip
Use a placeholder client and round numbers when you draft, then fill in real figures afterward. One analysis found that sensitive data makes up 11% of what employees paste into ChatGPT, and a client's billing details are exactly that. Generate the structure with dummy values, then enter the real amounts privately.
Or close the format gap entirely
The prompt gets you correct content. The work it does not do is turn that content into the document a client actually receives, and that gap is where the time and the amateur look both come from.
If you would rather skip straight to a sendable invoice, FreelanceDesk takes the same details and produces a clean, numbered PDF with the payment terms and late fee already in place. You keep your own invoice sequence, you set your own tax rate, and the document never leaves your browser. It is free, and the layout problem disappears. The invoice anatomy guide covers every field the document needs, and the post on late-paying clients covers what to do when a correct invoice still goes unpaid. Once the invoice is out, the contract that should have come first is the document that makes the late fee enforceable.