The check is the job

The equation that decides everything

Time saved equals time to write, minus time to brief, minus time to check, minus time to fix.

Every argument about whether AI helps at work is an argument about the third term, and almost nobody measures it. People measure how fast the draft arrived. The draft arriving in nine seconds is not the outcome. The outcome is a finished thing you are willing to put your name on, and the distance between those two is checking.

Generating is easy, verifying is not

There is a useful asymmetry underneath this. For some tasks, checking an answer is enormously cheaper than producing one. For others it is exactly as expensive, and for a few it is more expensive.

Four bands, and you can place any task you do:

Free to check. The answer verifies itself. A spreadsheet formula either returns the number you know is right for row 12, or it does not. A regular expression either matches your test strings, or it does not. Code either runs, or it does not. These tasks are where AI is at its most useful, and it is no coincidence that programmers noticed first.

Cheap to check. A few seconds each. Is this the right client name? Does this total match the invoice? Is that section number real? Cheap, but only if you do it — and the cheapness is what makes people skip it, because a check that takes three seconds feels like it cannot matter.

Expensive to check. A claim about the law in your jurisdiction, a clinical dosage, a statistic attributed to a report, a translation into a language you do not read. Verifying any of these takes about as long as producing it properly would have. If your task lives here and you did not supply the source material, the tool has not saved you time. It has moved your work from writing to auditing, and auditing somebody else's plausible prose is slower and less pleasant than writing your own.

Impossible to check. What was actually decided in a meeting you did not attend and which was not recorded. What a client meant. Whether a person will be a good employee. No amount of care recovers a fact that is not available to you, and a fluent answer here is worse than no answer, because it removes the feeling that you are missing something.

The rule that follows

Any task where checking costs more than doing is not a candidate, however good the draft looks.

This is the single most useful sentence in the course, and it is unpopular because the draft always looks good. Fluency is produced by the same machinery whether the content is right or wrong, so quality of prose carries no information about quality of content. You cannot use your reaction to the writing as evidence about the writing.

Where the facts came from decides the cost

Notice that the cost of checking is not a property of the task. It is a property of where the facts came from.

Ask for a summary of a policy document you attached, and every claim is checkable against a page you have in front of you: cheap. Ask for a summary of "current data protection requirements for a clinic in your country", and every claim requires research you have not done: expensive.

The same request, the same model, the same output length — and a difference of an hour in what it costs you to be responsible for the result. This is why the strongest single habit in professional use is to supply the source rather than ask from memory. It does not make the model more truthful. It moves your checking from the expensive band to the cheap one.

The gap you cannot close by being careful

Here is the honest limitation, and it is the one most often glossed over.

You cannot check what you could not have written. If you do not know the field, you can confirm that a citation exists and that the words are spelled correctly. You cannot confirm that the argument is sound, that the standard cited is the current one, or that the crucial exception has been omitted. Omission is invisible by definition: nothing on the page tells you what is not on the page.

This means AI is least safe exactly where it feels most valuable — at the edge of your competence, where it produces work you could not have produced. A tax specialist checking AI-drafted tax analysis is doing real verification. The same specialist checking AI-drafted employment advice is reading it for plausibility and calling that a check.

A workable rule: use it inside your competence to go faster, and outside your competence only to orient yourself before asking somebody who knows. Orientation is a legitimate and useful mode. Presenting the orientation as an answer is not.

Measure it once, properly

Take one task you already do. Do it your usual way and note the minutes. Do the next three with AI and note the minutes — including the checking and the fixing, timed honestly, from the moment the draft appears to the moment you would send it.

Most people who do this find two things. First, the number is smaller than they expected and still positive: real, modest, worth having. Second, one or two tasks on their list come out negative, and those are always tasks where they had been enjoying the draft and paying for it later.

A later lesson turns this into a two-week trial you can put in front of a manager. For now, the point is smaller and harder: the stopwatch has to run through the checking, or you are measuring the wrong thing.