I have never actually seen an executive smash a pastry into his own forehead, but in fiction we are allowed certain freedoms.
Glenn was a middle manager in a large manufacturing company. Fairly paid, overworked, and running on the kind of fatigue that makes small problems feel like personal attacks. It had been a hard week. His brand-new puppy, left alone for only a few minutes in his vintage muscle car, had chewed through the back seat. This was not a good moment for Glenn.
The next morning, still carrying the emotional residue of what he had started calling the Great Interior Debacle of 2026, Glenn stopped at Donut Right Coffee Shop for a cup of Make Your Mind Right Java. While he waited, an email came in from his team. He scanned it quickly, saw the words “production deadline,” “problem,” and “did not miss,” and somehow assembled them into a crisis.
Already fuming, Glenn opened his AI assistant and typed:
Dear AI,
Your role is to act as a stern, hard-nosed manager.
The objective is to write a note showing my disappointment in the team for missing an important deadline.
The context is that my production team had an urgent deadline and missed it, probably because they were thinking about the Fourth of July.
Success will be when I see the realization on my employees’ faces that they have let me down.
The tone should be direct, aggravated, dissatisfied, and carry a slight “heads might roll” quality.
The output should be a strong, professional, but angry email I can send to the team.
AI, being helpful in the way a chainsaw can be helpful, gave Glenn exactly what he asked for.
The email was firm. It was clear. It was professionally disappointed. It may have used the word accountability at least twice.
Glenn sent it.
Then his coffee arrived.
With the first sip, some small part of his brain came back online. He reread the original email from his team. It said, “Glenn, we were able to avoid a problem and did not miss the production deadline.”
That is when the pastry met the forehead.
Coffee was not the problem.
The prompt was not even the problem. In a strange way, the prompt worked. Glenn told the machine what role to play, what outcome he wanted, what tone to use, and what would count as success. The machine followed the instructions with impressive obedience.
The problem was that Glenn had not paid attention.
This may be one of the quieter risks of artificial intelligence. We tend to worry about the tool becoming too powerful, too persuasive, or too autonomous. Those concerns are not meaningless, but the more common problem may be simpler. AI can make a bad assumption look organized. It can give our impatience a clean format. It can turn a mood into a message before judgment has had a chance to enter the room.
A solid prompt does not rescue a careless premise.
That seems worth remembering, especially when the tool feels efficient. AI can help draft, clarify, summarize, and structure. It can save time and sometimes reveal a better path through a thought. But it does not know whether I have read the original email correctly. It does not know whether I am reacting to reality or to the version of reality assembled by fatigue, irritation, and an under-caffeinated brain.
That part still belongs to me.
It’s clear the human in the loop is not just there to approve the final output. The human is there earlier, before the prompt, before the tone, before the clever instruction. The first responsibility may be much smaller and much harder.
Read the email.
Then drink the coffee.
Then decide whether anyone’s head needs to roll.
Leave a Reply