When AI goes wrong in hurried hands

Most AI mistakes at the office do not happen because the tool is bad, but because the person is in a hurry. With a deadline close, we tend to accept the first answer that looks tidy. There are four ways AI most often goes wrong, and all four slip easily past hurried eyes.

First, inventing. The model writes names, figures, dates, quotes or references that do not exist. This is called hallucination. The sentences are fluent and convincing, so there is no warning sign on the surface.

Second, agreeing. The model tends to follow the direction of your question. If you ask "This plan is good, right?", it is more likely to look for reasons to agree than to point out its holes.

Third, skipping. The model can quietly ignore part of the material or instructions: the last row of a table, the second attachment, or the condition "one page at most". The result looks complete, yet something is missing.

Fourth, going too far. This happens mostly with AI agents, AI that can carry out steps such as sending messages or changing files. An agent can do things nobody asked for, or report that everything is done when it is not.

These four mistakes share one thing: they all look like a correct result. That is why the speed of reading and checking, not the speed of getting an answer, decides whether AI saves you time or adds to your problems.

Hurry also makes us careless with what we paste. Wanting to be quick, people paste a whole customer file full of personal data, such as names, phone numbers or ID numbers, when the task needs only two columns. Protecting this data is covered in the module "Protect the data".

The fourth mistake deserves a closer look, because this is where people are most often fooled. A report is a sentence saying the work is done: "All invoices have been matched", "I checked the source", "The file has been updated". Evidence is something you can see for yourself without having to trust anyone: a matching table per invoice, a link to the source page with the quoted sentence, a file that really changed, or a message that is actually in the sent folder.

When AI writes "I have verified this", it is still a report. That sentence is produced the same way as any other, and can be just as wrong. So the habit is simple: whenever you receive a report that something is done, ask what evidence you can see for yourself, then look at at least part of it.

The calmest way to work with AI is to decide at the start which steps never run without human approval. This principle is often called human in the loop: AI may prepare, a person approves the important steps. The list is usually short and similar across offices: sending something outside the organisation, paying or moving money, deleting data, publishing something to the public, changing a system many people use, and making decisions about a person, such as hiring an applicant or imposing a sanction.

What they share: the step is hard to undo, or its consequences fall on someone else. Write this list before you start, not when the deadline is close. In a hurry, every exception feels reasonable.