Catch the chatbot's mistake before the customer service rep does

You get a bill with a charge you don't recognize, or a renewal notice that says you're locked in for another year. You paste it into a chatbot, ask what it means, and get a tidy, confident answer. Then you call the company, repeat what the chatbot said, and the representative tells you that's not what your agreement says.

The problem is rarely that the chatbot can't read. The problem is that a vague question invites it to fill gaps. If a due date is cut off, it may supply a typical one. If you ask whether you can cancel, it may describe how cancellations usually work instead of how yours works. Either way, the answer sounds equally sure.

This guide gives you two prompts that close those gaps. The first reads an administrative document and isn't allowed to invent numbers. The second takes any claim, whether it came from the chatbot, a friend or a forum, and checks it word for word against your contract before you rely on it. Together they take about five minutes.

Why "what does this bill mean?" goes wrong

A short question gets a short, smooth answer. The chatbot has no way to show you which parts it read and which parts it assumed, so it blends them. A total printed on the page and a deadline guessed from a common billing cycle look the same in the reply.

Most of the trouble comes from three gaps in a loose prompt. It doesn't say what to do when information is missing, so the chatbot guesses. It doesn't ask for a list of assumptions, so the guesses stay hidden. And it doesn't tie advice to the source text, so you can't check where a claim came from.

You don't need a smarter chatbot to fix this. You need a prompt that sets rules: use only what is on the page, mark what's missing and show your work. Any free chatbot follows those rules well enough to be useful, as long as you give it the actual text.

Step one: get the facts out of the document without guesses

Start with the document that's worrying you. It might be a utility bill, a medical statement, a letter about a fee change or an insurance explanation of benefits. Copy the text from the PDF, or type out the relevant lines if it's a paper letter.

Before you paste, replace anything that identifies you. Your name becomes [Name], the account number becomes [Account Number] and your address becomes [Address]. Leave the amounts, dates and fee descriptions alone, because those are what the chatbot needs.

I am pasting an administrative document below. I have removed my personal identifiers and replaced them with bracketed placeholders.

Analyze this document and provide:
1. A 3-bullet summary of the core issue or charges.
2. The exact total balance due and the payment deadline.
3. Any disputed fees, unexpected line items, or required next actions.
4. A list of any assumptions you made that are not explicitly written in the text.

Do not guess any missing figures. If a number or date is not in the text, write "NOT SPECIFIED".

Document text:
[PASTE REDACTED TEXT HERE]

Paste the redacted text where it says [PASTE REDACTED TEXT HERE] and send the prompt unchanged. The opening line tells the chatbot the brackets are deliberate, so it won't try to work out who you are or treat [Account Number] as an error.

A good answer comes back in four numbered parts. The summary names the charges in plain words, for example a maintenance fee added for local grid upgrades and an unpaid balance carried over from last cycle. The total and the deadline should match the page exactly. The third part flags anything that needs action, such as a late penalty if you miss the date, or a charge you might be able to contest if you qualify for an exemption.

Read part four first. That's where the chatbot admits what it filled in. If it says your account tier is NOT SPECIFIED, it's being honest, and you now know what to look up before you call. If part four is empty but the summary mentions things you can't find on the page, ask the chatbot to quote the line each fact came from.

Step two: check the advice before you say it on the phone

The first prompt tells you what the document says. The next question is usually what you can do about it: can you cancel, can you dispute this fee, are you owed a refund? This is where chatbots, friends and forum posts most often give advice that's close to right but wrong for your contract.

Before you act on a claim like that, check it a second time. The check works best in a fresh chat, so the chatbot isn't defending something it told you a minute ago.

I need to verify the accuracy of a statement against my original service contract text.

Statement to verify:
"[INSERT SPECIFIC CLAIM OR GENERATED ADVICE HERE]"

Original contract text:
[PASTE RELEVANT CONTRACT SECTION HERE]

Evaluate the statement using only the provided contract text:
1. Status: State whether the statement is [FULLY SUPPORTED], [PARTIALLY SUPPORTED], or [UNSUPPORTED / FABRICATED].
2. Verbatim Evidence: Quote the exact sentence(s) from the contract text that prove or disprove the statement.
3. Practical Risk: If I rely on this statement when speaking to a representative, what specific error could they call out?

Put the exact claim you plan to rely on inside the quotation marks. Write it as one sentence you'd say out loud, such as "I can cancel by phone without a penalty within 30 days of renewal." A vague claim gets a vague verdict.

For the contract text, paste only the section that applies, such as the cancellation clause, the fee schedule or the early termination terms. To find it, search the contract for words like cancel, terminate, notice or fee. Pasting the whole agreement works, but a focused section gets a sharper result and makes the quoted evidence easier to check.

The verdict comes in three parts. The status tells you how far to trust the claim. The evidence quotes the contract word for word, so you can find the same sentence in your copy. The practical risk is the most useful part, because it tells you what a representative could push back on. Here's a typical case. The claim says you can cancel by phone, but the clause requires written notice by certified mail or through the official portal. The risk is that your phone request gets rejected and you miss the window.

If the status is PARTIALLY SUPPORTED, rewrite your claim to match the quoted sentence and use that version when you call or write. If you search your contract and can't find the quoted evidence, don't trust any of the answer.

What to remove before you paste anything

Both prompts assume you've already redacted the document. Do this every time, even if the chatbot says it doesn't store your conversations.

Remove your full name, account and policy numbers, street address, phone number, date of birth, any government ID number and the last digits of a payment card. Keep the amounts, dates, fee names, clause numbers and the wording of the terms. The chatbot doesn't need to know who you are to read a fee schedule.

Use the same placeholder every time something repeats. If the bill mentions your account number three times, write [Account Number] all three times, so the chatbot can still see that the references match.

Turn it into a five-minute habit

These prompts pay off when you use them on every document that asks you to pay, sign or act, not just the ones that already look wrong. Billing errors are easy to miss because the total usually looks plausible.

Here's a simple routine. When a document arrives, redact it and run the first prompt. Check the total and deadline against the page yourself. If the answer suggests any action, write that claim as one sentence and run the second prompt against the matching section of your contract. Only after that do you call, pay or write back.

Keep both prompts in a note on your phone with the placeholders left in. Save the chatbot's output next to the document, with the date. If a dispute comes up later, you'll have a record of what you checked and which clause you relied on.

What this cannot do

• It can only read what you paste. If a fee is explained on a separate page or in an earlier letter, the chatbot won't know about it. NOT SPECIFIED means the information isn't in your excerpt. It doesn't mean the information doesn't exist.
• It doesn't make the chatbot correct. The rules cut down on guessing but don't stop it completely. A chatbot can still misread a table or quote a sentence that isn't there. Always find the quoted evidence in your own copy.
• It can't interpret the law around your contract. Consumer protection rules, state regulations and regulator decisions can override what the contract says. The second prompt checks only the text you give it.
• It can't tell you what the company will actually do. A representative may waive a fee the contract allows, or refuse something the contract permits. The prompt prepares you for the conversation but can't predict how it ends.
• It isn't legal or financial advice. For large amounts, threatened collections or anything involving court, take the document and your notes to a qualified professional or a free consumer advice service.
• Scanned documents need clean text. If you copy text from a blurry scan, digits can come through wrong. Compare every number in the output with the original.

Related reading

• Stop Typing, Start Asking — full first chapter, free, no email required
• Full contents and what is included
• The three documents you signed without reading
• The three emails you keep putting off, each under 120 words
• All Avalon Store products