Nobody learns to drive from a manual

A driving manual can tell you where the clutch is and when to change gear. It cannot teach you what the engine feels like just before you stall. You learn that by driving, often with someone beside you saying: listen to the engine, now try third.

Using AI at work has something in common with that. Knowing how to ask a question is a start. The harder skill is deciding what to ask it to do, what context it needs and whether the answer is good enough to use.

It is easy to find one useful way of working with AI and stop there. It is just as easy to trust it with a task it cannot reliably handle. Guided practice can help you notice both.

That is an easy argument for me to make. I sell AI training. So it is worth looking at what the research shows, and where my argument goes beyond it.

AI helps with some work and hinders other work

In a study of 5,172 customer support agents, access to an AI assistant increased the number of issues resolved per hour by 15 percent on average. Less experienced and lower-skilled workers gained the most.

A study involving 758 BCG consultants showed why that benefit cannot be assumed for every task. On work the model handled well, consultants using AI finished about 25 percent faster. On a task outside its capabilities, roughly 84 percent of consultants without AI reached the correct answer, compared with 71 percent using AI and 60 percent using AI with a prompt engineering overview.

A smaller METR trial found that 16 experienced developers took 19 percent longer when allowed to use early-2025 AI tools on their own projects. Afterwards, they still believed the tools had made them faster. METR's February 2026 follow-up suggested newer tools could be helping, but problems with who and which tasks entered the study made the size of that benefit uncertain.

These studies used different tools, people and tasks. They do not isolate one cause of success or failure. My takeaway is practical: test AI on the work you actually do, and measure the result.

The difficult part is recognising the limits

AI's limits are not always obvious. A model can produce a useful summary and then miss the exception that changes the decision. Two tasks that look similar to you may be very different for the tool.

Think of a customer email. AI might help you make the wording clearer. But if the reply promises a refund, someone still needs to check the policy and the customer's circumstances. A well-written answer can still be the wrong answer.

This is where I think guided practice earns its place. Working through a real task lets you see what information the model needs, where it makes assumptions and what you must verify yourself.

A course can introduce those habits. Applying them to your own work takes practice, because the consequences of a mistake depend on the job.

Training should improve judgment

The BCG result is uncomfortable for anyone selling training, including me. On the task outside the model's capabilities, the group given a prompt engineering overview had the lowest accuracy.

That does not prove training is harmful, or tell us exactly why that group did worse. It does show that a short overview was no guarantee of better judgment.

There is some encouraging evidence for more focused teaching. In a small experiment with 30 novices, people taught to define their requirements clearly improved their prompting more than people given conventional prompt engineering training.

That supports teaching people to be clearer about the result they need. It does not establish the financial return of AI coaching. I have no measured return to promise you, and these studies do not provide one for my programme.

Practice should start with a real task

Start with something your team does often enough to compare results. Choose a task where you can check the output before it affects anyone else. Then work through it together.

1. Define a good result. Write down what the output must contain, what it must avoid and how you will check it.
2. Try AI on a suitable part of the task. Give it the context it needs, within your organisation's data rules.
3. Review the answer together. Identify errors, missing information and assumptions. Adjust the request and try again.
4. Compare the full job. Count the time spent checking and correcting, as well as drafting. Keep the approach only if the result is worth it.

The comparison matters. The developers in METR's early study felt faster while taking longer. A useful session should leave you with a way to judge the work, including a way to notice when AI is adding effort.

A faster task needs to help the team

Suppose AI cuts the time needed to draft a report, but the reviewer then spends longer checking it. The writer has saved time. The team may not have.

That is why practice should include what happens after the answer appears. Who checks it? What evidence do they need? Can the next person use it without doing the work again?

Those questions connect an individual's use of AI to the company's process. A useful result has to survive the next step.

You may not need a coach

The support-agent study is a reminder that access itself can help. People improved when given a useful assistant within their existing work.

If your team is already getting reliable results, checking them well and learning from each other, you may not need outside help. A capable colleague can also provide guidance.

My case for coaching is narrower: it can help when people are unsure which tasks to try, keep accepting plausible mistakes, or cannot tell whether AI is saving effort. Whether that help is worth paying for should be judged by what changes in their work.

AI literacy also has a legal role in the EU

For organisations covered by the EU AI Act, Article 4 requires providers and deployers to take measures that support AI literacy among relevant staff and others using AI on their behalf. The amended text does not require a guaranteed level of literacy for each individual. The Commission also says a certificate is not required. Separate human-oversight training obligations apply to deployers of high-risk systems. What the duty actually asks for is set out in a separate article on Article 4.

That is a reason to take literacy seriously. The practical aim remains the same: help people understand the systems they use and make better decisions with them.

What I offer through LEAD AI

LEAD AI is practice on your own work, with me beside you. We work through the choices that matter: where AI can help, what to give it and how to check what comes back. Where it sits in the rest of a project is on the method page.

The studies are useful starting points. They describe particular people using particular tools at particular times. None can tell you what will happen in your team.

That is what the practice is for. Take a real task, try it carefully and find out.