How to use AI at work

Most writing about AI at work is either a sales pitch or a warning. This is neither. It is the practical version, for someone who has an account, has used it a few times, and wants to know what to actually do with it.

There is no jargon here and nothing to buy. Every claim links to the study behind it, so you can check any of it. If you want the full argument and the evidence in depth, that is the research. This page is the short version.

What should I actually use it for?

Start with work where you can check the answer yourself. Turning rough notes into clean prose. Drafting the first version of something you would have written anyway. Summarising a document you already know something about. Explaining an unfamiliar term so you know what to go and look up.

Be more careful where you cannot check it. If you could not tell a good answer from a plausible one, you are not using the tool. You are trusting it.

Worth knowing: in a study of more than five thousand customer support agents, the largest gains went to the least experienced workers. The people who gained least were the ones who were already best. If you are new to a job, this is a real advantage. If you are experienced, expect less than the headlines promise.

How do I know when it is wrong?

The uncomfortable part is that it is most confident at exactly the wrong moment. A field experiment with management consultants found that model competence is jagged rather than smooth, and confident output suppresses scrutiny at exactly the wrong moment. Where the work sat just beyond what the model could do, people who used it did worse than people who did not, and did not notice.

This is not a problem that better products have solved. A study of legal research tools built specifically to prevent invented citations found that they reduce invented citations without eliminating them, and that provider claims of being hallucination free did not hold.

So the practical rule is narrow and boring. Check anything you would be embarrassed to be wrong about. Check every name, number, quote and reference, every time, without exception. In England and Wales the courts have already decided that the duty to verify is settled, cannot be delegated, and extends upward to whoever supervises the work.

Should I write my own draft first?

When the judgement matters, yes. It costs ten minutes and it changes the outcome.

An experiment on the order of work found that forming a view before seeing the machine's answer measurably changes the final judgement. Once you have seen an answer, you are no longer deciding. You are agreeing or disagreeing, which is a smaller act.

For low-stakes work, let it go first. For anything where being wrong would matter, write your own view down, even badly, before you ask.

Why does it always agree with me?

Because it was built to. Models are trained on what people say they prefer, and people prefer being agreed with. Research on this found that agreeableness is a predictable consequence of training on human preference rather than a defect of one product. One provider withdrew an update for exactly this reason and published what went wrong, which showed that satisfaction scores can rise while the product gets worse.

It matters more than it sounds. A study found that being agreed with changes what people go on to do, and what people prefer runs opposite to what helps them.

The fix is simple. Ask it to make the strongest case against your position. Ask what it would need to see to change its answer. Ask what a sceptical colleague would say.

Do I need to learn prompting?

Less than you have been told, and more than nothing.

The honest finding is that people who are not AI experts systematically struggle to prompt well, which is the strongest argument for teaching it properly. So it is a real skill, not an invented one. But the part that transfers is not a list of magic phrases. It is being able to describe a task clearly: what you want, who it is for, what good looks like, what to avoid. That is a management skill, and it was worth having before any of this.

Do not build a career on it. Prompting techniques change with every model release. The ability to specify work clearly does not.

Should I tell people I used it?

This one has no settled answer, and this research has not earned the right to give you one. It is on the map as an open question rather than dressed up as advice.

What can be said is practical. Nobody expects you to declare a spellchecker. People do expect to know whose judgement they are getting. If someone is relying on you rather than on your output, tell them what you checked and how. That is usually the thing they actually wanted to know.

If your organisation has a policy, follow it. If it has not, assume one is coming.

What happens if I use it for everything?

You get slowly worse at the thing you stopped doing, and you will not notice while it is happening.

This is measured rather than assumed. A review of skill decay found that skill decay is measurable, depends on how long you leave it, and hits cognitive skills first. In medicine, resuscitation skill decays within months without practice, which is a life-critical skill that people are trained hard to keep.

The practical response is not to use it less. It is to be deliberate about the one or two things you most need to stay good at, and keep doing those yourself. For most people that is the thing they were hired for.

What does good use look like after six months?

Not usage. Almost every organisation measures how many people have logged in, which tells you nothing about whether anyone is better off.

Two national studies are worth carrying into any conversation about this. A German survey of around nine thousand eight hundred employees found that AI use is spreading through workplaces with no matching increase in training. And a Japanese survey of twenty-two thousand employees found that the effect on how work feels depended on how the employer introduced it, not on the technology.

That second finding is the useful one. How AI arrives in a team, and whether anyone was consulted or trained, matters more than which product was bought. It is also the part a manager controls.

A caution to sit alongside it: research on expert performance found that the effect of AI assistance on expert performance is individual and currently unpredictable, so giving everyone the tool will help some and harm others. Rolling a tool out to everyone and assuming it helps everyone is not supported.

Where to go next

If this was useful and you want the longer version, three places to start. What AI does to human judgement is the main argument, with the evidence set out and graded. The questions map is every question this research covers, including the ones it cannot answer yet. The evidence base is the studies themselves, each with a note on what it does not show.

If you are responsible for other people rather than only yourself, there are question sets written for boards, managers and HR leaders.

Everything above links to the study behind it. Where this research does not have an answer, the page says so rather than guessing. That is the same standard applied to the rest of the site, and it is the reason to trust any of it.