You Ask ChatGPT Before Every Business Call and Trust Yourself Less Each Month. Here Is a Method That Fixes the Balance.

Why does leaning on ChatGPT feel fine until it suddenly is not?

Because it is right often enough that you stop checking. Then one day it is wrong while sounding exactly as sure as before, and the habit that would have caught it is gone.

The book follows Jonah, a one-person consultant whose clients pay him for judgment. He fed the tool every problem. First he checked its answers. Then he checked less. Then he started sending what it said, because it sounded surer than he felt. Eventually one of those answers, sent under his name, was wrong, and the client lost weeks and money before anyone spotted it.

If that drift sounds familiar, the problem is not that you use AI too much. It is that nothing in your routine decides which parts of a decision it gets.

Why “just double-check everything” does not work

The usual advice swings between two poles: trust it less, or read every word twice. Neither holds up on a busy week.

Rereading everything saves you no time, and it catches little, because a confident mistake reads as cleanly as a true sentence. Trusting it less, in general, just means you stop using a useful tool. What works is narrower: decide in advance what kind of help each decision gets, and point your checking at the few things that would hurt if they were wrong.

Researchers named the underlying pull decades ago. People working next to a confident automated aid follow it even when their own information says otherwise, and miss problems it fails to flag. The cause was ordinary trust, not carelessness.

Step 1: Put the decision in one of three lanes

Before you open the chat window, decide which lane this belongs in. The sorting below is our arrangement. Each line in it comes from the book.

When you are unsure, ask one question: if this is wrong, could someone hold me to it? If yes, it is at least the middle lane. Most of your day still sits in the first lane, and that is where AI saves the most time.

Step 2: Jot down your own answer first

On any middle-lane call, spend thirty seconds on your own read first. One line on a notepad: where do you think this lands? Then ask.

Now you are comparing two judgments instead of receiving one. If they agree, go ahead with more confidence than either gave you alone. If they point different ways, you have found the exact spot to look harder. You cannot do this afterwards, because once you read the answer it fills the space your own view would have used.

The reason this matters is the shape of AI’s mistakes. A study of 758 consultants at Boston Consulting Group mapped it clearly.

Inside the AI’s range, consultants finished about 12 percent more tasks in about 25 percent less time (Dell’Acqua et al., 2023).

On a task just outside that range, consultants using AI were about 19 percentage points less likely to get it right than those without it (Dell’Acqua et al., 2023).

Nothing on the screen marks that border. Your own first answer is the cheapest instrument you have for noticing you are near it.

Step 3: Keep it in the advisor’s seat

The same tool can brief you or decide for you, and the two look identical on the screen. The book gives a simple tell. An advisor leaves you with a choice still to make. A decider leaves you with nothing but a click.

So use prompts that keep the choice open:

• “Give me the strongest case against your recommendation.”
• “What would you need to know about this client to be confident, that I have not told you?”
• “Lay out three options without ranking them.”

In the book, Jonah asks the tool what to quote a long-time client. It says quote high, with a tidy justification. He knows something it does not: this client had walked away from another firm once because its number seemed greedy. He sends the high quote anyway, and the client goes quiet, then goes elsewhere. The tool missed the fact. Jonah made the choice.

Step 4: Read your doubt before you ask the machine to settle it

If you are opening ChatGPT to make an uneasy feeling stop, pause. The tool answers all doubt with the same confidence, so it erases a real warning as easily as a tired one.

Name the doubt in one sentence, then sort it. A warning is specific and gets sharper the closer you look, so go find the fact behind it. Fatigue is vague, late in the day and gone after sleep, so decide tomorrow. Fear gets worse when you picture people watching, so choose as if no one would ever know the outcome.

In a large field study, the quality of high-stakes professional decisions declined over the working day as decisions piled up (Hirshleifer, Levi, Lo and Teoh, NBER, 2018).

The line the book draws is exact: use AI to inform the doubt, never to silence it. Asking it to pull a vendor’s support record is informing. Asking whether you should feel okay is silencing.

Step 5: Run the four-line gate on anything costly

Before you act on an answer that could cost you money or a client, fill four lines. The book calls it the Judgment Gate. It takes two minutes.

1. The AI’s answer in one line. If you cannot shrink it to one line, you do not know what you are about to act on.
2. Inside or outside its range? General patterns it has read a million times are inside. Facts about this client, this market, this month are things it is inferring.
3. How you would know it is wrong. What would have to be true for this answer to fail? Does your own first read disagree?
4. Your independent check. The price against the actual sheet. The claim against the actual data. A source that is not the AI.

Another of Jonah’s clients wants to know whether to switch to three pricing tiers. The tool says yes and predicts about 15 percent more revenue. Line two flags that number, a confident figure resting on data the tool never saw. Line four sends him to the client’s churn records. People were leaving in their first thirty days over a setup problem, not over price.

Step 6: Write down who decided

The fix that research singles out is accountability. When people had to justify their decisions, automation bias dropped. You can build that pressure yourself with four short lines before a decision ships:

• The call, in one sentence.
• Who actually decided. Sometimes the honest answer is “the tool did and I approved it.”
• Who carries the result. Always you.
• Did you own it? Your reasons, written without looking at the screen.

Run it on easy decisions too. Easy is often a recommendation you did not notice you accepted. Then, once a week, make one real decision the long way, with the tool closed until you have your own answer.

In a 2025 survey of 666 people, frequent AI use was associated with lower critical-thinking scores, linked to handing thinking to the tool (Gerlich, Societies, 2025).

That is a correlation, not proof that AI causes the drop. It is still a good reason to keep your own judgment in regular use.

Fair doubts