Someone sent you something this month that nobody wrote.
A four-page reply to a two-line question. A strategy deck with a confident number on slide six that turned out to be invented.
It took them ninety seconds. It took you an hour.
Let AI produce the work, pass it on unchecked, and someone downstream has to find the mistakes and clean up the mess. You look productive. They pay.
Saved hours for whom?
I now distrust any claim that a tool “saved hours”. Saved them for whom? Nearly every productivity number is measured from one seat: the producer’s. The reader’s hour never appears.
For most of my working life, a long, polished document suggested someone had thought hard about it. That signal is now easy to fake. Send enough unchecked work, and people learn to read everything with your name on it twice.
I build with these tools and use them all day. One of my own AI agents took four paragraphs I had asked it to file and turned them into fifty minutes of work I then had to audit.
The most expensive thing these tools produce is a plausible answer nobody checked.
In a clinic, the checking is part of the work
Consider a hypothetical AI-generated patient summary. A medication copied incorrectly, a missing date or an unsupported interpretation can send the next person back through the whole record. A shorter drafting time says little about whether the summary is useful.
Before that summary is used in care, someone needs to check it against the source record, distinguish reported facts from interpretation, and resolve anything that could change a decision. Clinical judgement stays with the responsible clinician. A confident sentence is not evidence that this work has happened.
That makes the useful question more specific: did the tool reduce the total work of producing a reliable result, including the work it handed to someone else?
One test before you send it
Have you read it as carefully as you are asking them to?
If not, you are spending someone else’s time. In healthcare, you may also be asking them to carry a risk they did not agree to take.
Read next
By Ivan Vatchkov. Adapted from my original LinkedIn post. Published on Kalibra, 26 October 2026.