The fastest way to lose credibility at work with AI is not to use it badly. It is to use it well for six months and then be unable to explain how you reached a conclusion. Judgement is the thing you are paid for. It is also the thing these tools quietly erode if you let them.

This is not an argument against using them. It is an argument for using them in the places where they are strong and refusing to use them in the places where they are not.

What they are genuinely good at

Three things, reliably:

• First drafts of structured writing. Meeting summaries, standard emails, job descriptions, process documentation. Work where the shape is known and the effort is in the typing.
• Reformatting and extraction. Turning a messy list into a table, pulling dates out of a contract, converting notes into an agenda. Low judgement, high tedium.
• Being a first reader. Asking “what would someone push back on here” before you send something to a client. It is not always right, but it is fast and it is never too busy.

Notice what these have in common. In every case you can check the output in less time than it would have taken you to produce it. That is the test.

Where they cost you more than they save

The failure mode is not the obvious one. Everyone knows these tools invent facts. What catches people out is subtler: they produce work that is plausible at the level of structure but wrong at the level of substance, and structure is what we skim for.

A risk assessment generated in eight seconds will have the right headings, a sensible-looking matrix, and controls that sound reasonable. It will also be generic, because the model has never been in your building. The document passes a glance and fails an audit.

The rule that holds up: do not use these tools where you cannot personally verify the output. If you would not be able to defend a line of it in a meeting, you should not be sending it.

The judgement problem

Here is the thing nobody warns you about. Writing a first draft is not just producing text. It is how you find out what you think. When you skip that step often enough, you arrive at meetings with a position you have not actually reasoned your way to, and you cannot defend it under pressure, because you never built it.

You notice this the first time someone asks “why did you recommend that?” and you realise the honest answer is that it was in the draft.

The practical fix is a small one. Decide your position first, in two or three rough sentences of your own, then use the tool to develop and pressure-test it. The order matters more than the amount of time it takes.

What to never put in

Assume anything you paste may be retained, reviewed by a human, or used to improve a model, unless your organisation has a contract that says otherwise. That rules out:

• Personal data about colleagues, customers or candidates
• Unpublished financial figures
• Anything covered by an NDA
• Credentials, keys and internal system details

This is not paranoia. It is the same instinct you already apply to a group email.

Say when you used it

Not on every email. But when a document is going to be relied on by someone else, a single line is worth writing: “First draft assisted by AI, figures and recommendations checked by me.”

Two reasons. It sets an accurate expectation about where errors are likely, and it protects you when one turns up. Colleagues forgive a mistake in a tool-assisted draft. They do not forgive discovering you concealed how the work was made.

A test you can apply today

Before you use one of these tools for a task, ask: can I check this faster than I could do it?

If yes, use it, check it, move on. If no, the tool is not saving you time. It is transferring your effort from producing the work to auditing it, usually at a worse ratio, and adding a risk that was not there before.

Used that way, these tools are the best thing to happen to routine work in a decade. Used the other way, they are a very fast route to being confidently wrong in writing.