Two years ago the hottest skill in technology was "prompt engineering", the art of phrasing a request to an AI just so. Whole courses sprang up promising to teach the secret incantations. And then, quietly, the ground shifted. The models got better at understanding ordinary language, and the magic words stopped mattering nearly as much. A lot of people who invested in memorising clever prompts are now discovering they learned the spelling of a language that is already changing.
The skill that actually endures is not knowing what to type. It is judgement: knowing which work to hand to AI and which to keep, how to check what comes back, and when to trust the machine versus when to override it. That is not a trick you learn in an afternoon. It is a craft, and in 2026 it is quietly becoming one of the most valuable things a person on your team can have.
The good news is that it is learnable, and it does not require a technical background. Here is what "good at working with AI" actually looks like, and how to build it in a team of ordinary, busy people.
Think of AI as a brilliant, eager intern
The single most useful mental model we give clients is this: treat the AI as a talented intern who is fast, tireless, widely read, and occasionally, confidently wrong. Everything else follows from that picture.
You would not hand an intern a critical task with no context and then send their work straight to a client unread. Nor would you refuse to delegate anything to them out of fear. You would give them clear instructions, let them do a first draft of the heavy lifting, and then apply your own experience to check, correct and finish. Working well with AI is the same relationship. The people who get poor results are almost always the ones who either expect the intern to be a mind-reading oracle, or refuse to delegate to them at all.
Getting good at AI is less about learning to talk to computers and more about learning to delegate, the oldest management skill there is. Give context, set the standard, check the work. If you can run a good intern, you can run AI.
The four habits that separate the capable from the frustrated
Across the teams we've helped, the difference between people who get real value from AI and people who declare it "overrated" comes down to a handful of learnable habits.
- They give context, not just commands. The frustrated user types "write a proposal". The capable one says "write a proposal for a returning client in the logistics sector, warm but professional, emphasising our reliability, here's the last one we sent them". Context is the whole game, and it is a habit, not a talent.
- They treat the first answer as a draft, never the final word. The capable user reads what comes back with a critical eye, pushes back, asks for changes, and iterates. A conversation, not a vending machine. The frustrated user takes the first output, finds a flaw, and concludes the tool is useless.
- They know what they're checking for. This is the deepest skill. To catch an AI's confident mistake, you have to know enough about the subject to smell when something is off. AI does not replace your expertise, it raises the value of it, because your expertise is now the quality control on a much faster process.
- They know what never to delegate. The final judgement call, the sensitive client conversation, the decision that carries real risk, the thing that needs a human to own it. Capable users have a clear internal line, and they don't let the AI's fluency tempt them across it.
The literacy test that actually matters
Forget "can they write a clever prompt?" The real test of AI literacy is a different question: can this person tell when the AI is wrong? Someone who can spot the plausible-but-incorrect answer, the made-up figure, the subtly off tone, is genuinely AI-capable. Someone who can produce a beautiful prompt but trusts whatever comes back is a liability with good typing. Build your team toward the first kind of person, not the second.
Building it across a team without a training budget the size of a house
You do not need to send everyone on an expensive course, and frankly the expensive courses often teach the fast-expiring tricks anyway. AI capability spreads best the way most real skills do, through practice, sharing and a bit of structure.
Start with real work, not toy exercises. Pick genuine, low-risk tasks from people's actual weeks, the quote, the email, the summary, the tidy-up, and have them use AI on those. Skill sticks when it's attached to a job that mattered anyway.
Make sharing normal. The single cheapest accelerator is a channel or a ten-minute slot in an existing meeting where people show what worked and what flopped. One person's discovery becomes everyone's, and the flops are often more instructive than the wins.
Normalise the checking, loudly. The culture you want is one where "I checked the AI's numbers and it was wrong about two of them" is a badge of competence, not a confession. If checking is seen as distrust of a shiny tool, people will stop doing it, and that is precisely when AI becomes dangerous.
Give it a rhythm, not a one-off. Capability built in a single workshop evaporates. A little, regularly, beats a lot, once. This is exactly the thinking behind bite-sized, repeated learning, and it is why short, consistent practice outperforms the grand training day that everyone forgets by Friday.
Where this is heading
The gap opening up in every industry is not between people who use AI and people who don't, almost everyone will use it soon enough. It is between people who use it well, with judgement, and people who use it credulously, trusting whatever the confident machine hands them. The first group moves faster and makes fewer mistakes. The second moves faster and makes their mistakes faster too.
That difference is not about intelligence or technical background. It is about a set of habits and a certain healthy scepticism, both of which any team can build with a bit of intent. The businesses that invest in the judgement, not just the tools, will quietly pull ahead, and they'll do it with the people they already have.
