An AI second opinion is a second read of an answer you already have, by a model that sees the same question and the first answer. It helps most in the minutes before you act on that answer: sending the email, signing the order form, paying the invoice or walking into the meeting.
What people tend to ask for
Many requests start with a pasted ChatGPT answer and the words "is this right?" That question invites a yes. A model asked to confirm something usually finds a way to confirm it.
The requests that get somewhere name a reader and a risk. Someone has a price-change email and wants to know what an annual customer will object to. Someone has a contract summary and wants to know which lines the contract itself doesn't support. Someone has an AI recommendation between two vendors and wants to know whether it still holds if the cheaper vendor's uptime figure turns out to be wrong.
A worked example
This is an example request, not a customer case. A founder of a small software company has a draft email telling annual customers that renewal prices go up 20 percent in 60 days. A chatbot helped write it. Before sending, the founder puts three things into one request:
- The draft email, word for word.
- The facts it depends on: the current plan terms, the renewal and notice clause from the customer agreement, and the date the new price takes effect.
- One question: which sentences would make a customer cancel, complain or dispute the charge, and which sentences claim something the agreement doesn't support?
A review that helps quotes the sentence it is challenging. It might note that the email says the new price applies "at your next renewal" and ask whether 60 days meets the notice period in the clause. The founder can check that against the agreement in a minute. A review that calls the email "clear and professional" gives the founder nothing to check.
A request you can copy
Below are the task the first answer was meant to do, the first answer, and the source material. List each claim in the answer that the sources don't support, quoting the sentence. Then list what [the reader: a customer, an investor, a landlord] would object to first. Don't rewrite the whole thing. Where you agree with a part, name the source that supports it.
Leave the source material in. A reviewer can't tell you whether an email matches a clause it has never seen.
How the Pingpong web review app runs it
The web review app sends your question through several models in order. Each later model receives the original question and every earlier answer, with instructions to assess the work so far. You get a final answer and can open each earlier pass to see where a model disagreed or added something.
Each reviewer reads the earlier answers, so a confident mistake in the first pass can carry forward. Agreement across models is not proof. Treat each flagged problem as something to check against your own documents.
When to use someone accountable instead
Medical, legal and tax questions where you need a person responsible for the answer belong with a licensed professional. A model review can help you write down the questions to bring to that person and the documents they will ask for.
For more on choosing a review, see when more than one model helps, sequential and parallel review compared and reviewing a decision between options.