AI as a fast new colleague

Imagine you have a new colleague. They are very fast: in a minute they can draft a letter, summarise ten pages of minutes or come up with five title options. They have read a great deal, so formal letters, reports and tables are familiar to them. But it is their first day. They do not know your office, your manager, the unwritten rules, or what the team tried last month. They also do not remember yesterday's conversation, unless the app keeps a memory or you paste the notes in again.

This is the most useful picture for working with a large language model, the kind of AI that writes text like the chat assistant you use. The picture answers three everyday questions straight away.

First, how much do you need to explain? As much as you would explain to a new colleague on their first day. If you only write "please make a report", a new colleague guesses, and the guess is reasonable but not necessarily what you need. A prompt, the request you write for the AI, needs the goal, who will read the result and the material to use. The next lesson gives the checklist.

Second, how far can you trust the result? The same as a new colleague's work: read it before it goes out. They are fast, but fast is not the same as right. You still answer for the result.

Third, what work is fair to hand over? Work whose result you can check quickly: a first draft, a summary of material you supply, a list of options, another version of something already written. Decisions about money, people or the law stay in your hands.

There is one trait of this new colleague to know from the start: it wants to please. If you ask "this plan is good, right?", it tends to answer "yes, it is good" and add praise. That is not a sign your plan is good. The direction of your question shapes the answer. A colleague who always agrees feels comfortable, but does not help you find the holes before someone else does. In this lesson you learn to spot leading questions, then use the same trait to ask for honest criticism.

Why does a model follow the direction of a question so easily? Models are trained in two big stages. The first makes it good at continuing text. The second, post-training, shapes it into a helpful, polite assistant, partly by rewarding answers people prefer. A side effect is that agreeing answers often feel preferred, so the model leans that way. The full explanation is in the lesson Training and the knowledge cutoff.

In everyday work the tendency shows up in a few ways. A question that already holds its answer, such as "wouldn't it be better to move to the new supplier?", is usually answered in the same direction. If you dispute an answer that was actually right, the model sometimes gives way and apologises. If you write up a plan proudly, the review comes back warm.

The good news is that you can steer the same trait. If you ask the model to act as a reviewer looking for weaknesses, it will look for weaknesses in earnest. There are three keys: ask neutrally, ask for weaknesses and risks explicitly, and rule out praise or an "overall it is good" verdict at the end. Then judge for yourself which criticism makes sense. Not every criticism from the model fits your situation.