What to Use AI For at Work, and What to Keep (2026)

You are reading Chapter 3 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

Deciding what to use AI for at work comes down to one question, and it is not "can it do this". The answer to that is always yes. The question that sorts your list is whether you can check the answer faster than you could do the work yourself. If checking is cheaper, hand it over. If it isn't, keep it.

This guide gives you that sorting question, four kinds of work that pay off in your first week, four to keep for yourself, and a five-line rule for anything that doesn't obviously fall into either pile.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Ask an assistant whether it can do something and it will say yes, then produce something. Willingness is not capability, and a wrong answer arrives in the same confident tone as a right one. So the question tells you nothing about whether the output is any good.

Swap it for a question you can actually answer: can I check this faster than I could do it? That single test sorts almost every task on your list, because it prices in the part people forget - the reading, the verifying, the fixing.

The research points the same way. In a study of 758 consultants, people using an assistant on suitable tasks finished 12% more work, 25% faster, at higher quality. On tasks outside what it handles well, the same people got it wrong 19 percentage points more often than those working without it. The tool did not get worse. The sorting did.

If you have not opened an assistant yet, Chapter 2 gets you signed in in about ten minutes.

These four are cheap to check, because you already hold the material or you are the only person who has to accept the result.

A first version of something you were going to write anyway. You already know what it needs to say, so you can tell in seconds whether the draft is close.

A summary of material you supply. You have the source in front of you, so checking it is reading, not research.

Something you half-know, explained back to you until you know it. You will feel the difference between an explanation that lands and one that doesn't.

Your own content in a new form: notes into an email, a document into bullet points, long into short. The facts are yours already, so only the shape is new.

In one study of professional writing tasks, people using an assistant took 40% less time and their output was rated 18% higher in quality. Those were exactly this kind of task: short, self-contained, and easy to judge.

People assume the rule is difficulty: give it the easy jobs, keep the hard ones. That is not the line. The line is whether you could catch a mistake, and whether the work is yours to sign.

Facts you have no way to verify. Over 1,500 court decisions have now involved citations that did not exist, filed by people who could not check them and did not.

An assistant was not in the meeting and does not know your company. Local context is the one input it cannot get on its own.

Options are useful. The call is not something to hand over: a recommendation that reverses the moment you push back on it was never a decision.

The message where the point is that you wrote it. In a survey of desk workers, 40% said they received incomplete or unusable AI-generated work at least monthly, and lost around two hours each time fixing it.

Read from the top and stop at the first line that is true.

Then this is not the one. Anything that goes out unread is the wrong place to start.

Do it yourself. That is the sorting question, applied.

Paste them in first, then hand it over. The gap was the input, not the assistant.

Hand it over. A draft you edit beats a blank page you avoid.

Try it once. One attempt tells you more than a week of deciding.

Aim low on purpose. A ten-minute job you do every week is a better first target than the quarterly strategy document, and point the assistant at what you are slowest at rather than what you are best at. In a study of more than 5,000 support agents, the newest agents gained 34% while the most experienced gained almost nothing - the help lands where you are weakest.

Before: you try the assistant on whatever is in front of you, get a mixed result, and conclude it is either magic or useless depending on which task you happened to pick.

After: you sort first. Draft, compress, explain and reshape go over. Unverifiable facts, local context, the decision and anything you sign stay with you. Everything else runs through five lines, and you stop re-deciding it every time.

One thing carries into Chapter 4: a first answer that came back close but not right. Chapter 4 is about replying to it instead of starting again. If you want one-page references to keep beside you while you work, start with our free AI cheat sheets, and the rest of the syllabus lives in the AI Central Library.

Start with work where you already hold the material: a first draft of something you were going to write, a summary of a document you supply, an explanation of something you half-know, or your own content reshaped into another format. All four are quick to check, which is what makes them safe to start with.

Ask whether you can check the answer faster than you could do the work. If checking is cheaper, hand it over. If verifying would take longer than doing it yourself, keep it. "Can it do this" is the wrong question, because the answer is always yes.

Four things: facts you could not catch if they were wrong, anything that turns on your specific week, room or relationships, the decision itself rather than the options, and anything you are accountable for by name.

Difficulty is not the right axis. What matters is whether you can verify the result and whether the work is yours to sign. A hard task you can check is a better candidate than an easy one you cannot.

Point it at what you are slowest at. In a study of more than 5,000 support agents, the newest agents gained 34% in performance while the most experienced gained almost nothing, which is the pattern to expect: the help arrives where your own skill is thinnest.