Most people think the thing holding them back with AI is technical. They assume that to use it well you need to understand how it works under the hood. You don't. Some of the most capable AI users I know can't write a line of code. What they can do is get the best out of someone else.
That isn't a hunch. The Wharton professor Ethan Mollick, who has studied this as closely as anyone, argues that the people who are best at working with AI tend to be managers, teachers and parents, not engineers. His core rule is deceptively simple: treat AI like a person, but tell it what kind of person it is. The moment you stop thinking of it as a search box and start thinking of it as a colleague, everything changes.
So here is the path I take people down, from knowing nothing to being genuinely useful. Three levels, and one mindset that ties them together.
First, understand the "jagged frontier"
Before the levels, there's one idea you have to hold onto, because it explains nearly every bad experience people have with AI. The technology is not uniformly good or bad. It is brilliant at some things and confidently, fluently wrong about others, and the line between the two is invisible until you've crossed it.
Researchers from Harvard, Wharton, MIT and Boston Consulting Group ran a now-famous experiment with 758 consultants and named this the "jagged technological frontier." Inside the frontier, the results were dramatic: consultants using GPT-4 completed 12% more tasks, did them 25% faster, and produced work rated more than 40% higher in quality. But for tasks that fell outside the frontier, the very same tool made them around 19 percentage points more likely to get the answer wrong.
Same AI. Opposite outcomes. Learning roughly where that edge sits for your own work is the entire game, and the three levels below are really just a structured way of finding it.
Level one: ask it what it can do
Go to ChatGPT, or Claude, and ask the very first question we all asked Siri when the iPhone came out: "What can you do?" Then push the buttons. Try that. Try this. Ask it to do the next thing. Treat the first hour like a child with a new toy, not a professional with a deadline.
This sounds trivial, but most people never find the frontier because they only ever ask the one thing they walked in for. Spend an afternoon poking at the edges and you start to build an instinct for what it's strong at and where it bluffs. That instinct is worth more than any prompt template you'll ever copy.
Takeaway: before you trust it with anything important, play with it on things that don't matter. Curiosity is how you map the frontier.
Level two: treat it like a junior employee
Stop asking it Google-style questions and start giving it work. The most useful mental model is this: imagine a brilliant graduate who joined your company this morning. Top of their class, encyclopaedic knowledge, genuinely sharp, and absolutely no experience of you, your business, or the real world. What do you do with someone like that? You give them small things first, then steadily increasing responsibility.
So hand it real problems, not trivia. "Two of my staff are in conflict, here's the situation, what are my options legally and socially, and what's the most constructive way to handle it?" And then, crucially, debate it. Tell it where you think it's wrong and why. Make it defend its reasoning. This is exactly what Mollick means by telling it what kind of person to be: "you are an experienced HR director" puts it in a completely different headspace than a cold, context-free question.
The consultants who got the most out of AI in that big study had a name too. The researchers called them "cyborgs": people in constant back-and-forth with the model, blending their own judgement with its speed, rather than throwing it a task and copying whatever came back.
Takeaway: give it context and a role, hand it real work, and argue with it. Delegation plus debate beats one-shot questions every time.
Level three: step out of the chat box
The chat window is the training wheels. The real leap is letting AI act, not just talk. Tools like Claude Code and the new wave of coding agents run on your own computer, in the terminal, and can read your actual files, run commands and get things done, not just describe how you might do them. And typing into them is no harder than chatting on a website.
I pointed one at my own laptop, which had quietly filled up with a few hundred gigabytes of clutter over the years, and asked it to find out what was eating the space and tidy it up. Thirty gigabytes of junk gone, files I'd never have found by hand. No coding on my part, just a clear instruction and a tool allowed to act on it.
You don't need to be a developer for this. You need to be willing to step past the safe little chat box and let the thing do real work in the real environment.
Takeaway: the biggest gains come when AI can touch your actual work, your files, your tools, your systems, not just answer questions in a window.
The thread through all three: more responsibility
Notice the single move that connects every level: you keep handing it more. From "what can you do" to "solve this for me" to "go and do it." Capability follows trust, and trust grows as you learn where the frontier is. That progression never really stops.
We've all been promoted before we're ready
Here's the mindset that makes the whole thing click. We have all just been promoted into a job none of us trained for: managing AI.
Picture the best accountant in a firm. Brilliant with numbers, fast, accurate, the safest pair of hands in the building. One day the boss says, "here are five people, they report to you now." She has never managed anyone in her life. She's excellent at the work and a complete beginner at getting work out of other people. Those are different skills, and the second one is the one she now lives or dies by.
That is every one of us with AI right now. And it happened almost overnight, because the price collapsed. A model as capable as the original GPT-4 now costs roughly 60 times less than it did in 2023, and the price of a given level of quality has been more or less halving every couple of months. We have each been handed a researcher, a designer, a junior lawyer and a tireless assistant, all with a top-tier education, for the price of a couple of coffees a month.
And managing someone more capable than you is genuinely hard. Most of us have only ever trained people below us, teaching them to do the job we already know how to do. Almost nobody has had to direct someone faster and more knowledgeable than themselves. That isn't a technical problem, it's a management one: setting clear direction, giving good context, checking the work, and knowing what to keep for yourself. Which, conveniently, is the exact "centaur" split the researchers found worked best: decide deliberately what the human does and what the AI does, and play each to its strengths.
So stop asking "what can this tool do?" Ask "I've got this brilliant, inexperienced employee, how do I manage them?" Answer that, and the rest takes care of itself.
The takeaways
- It's a people skill, not a coding skill. The best AI users manage, teach and delegate well. If you can get the best out of a person, you can get the best out of AI.
- Respect the jagged frontier. AI is superb inside its range and confidently wrong outside it. Learn where that edge sits for your work before you rely on it.
- Climb the three levels: explore it like a toy, delegate to it and debate it like a junior hire, then let it act on your real files and tools.
- Keep handing it more responsibility. Capability follows trust, and trust grows as you learn its limits.
- Manage, don't just prompt. You've been promoted to running a brilliant, inexperienced team member. Treat the job like management, because that's what it is.
Further reading: Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (Harvard / Wharton / MIT / BCG); Ethan Mollick, Co-Intelligence: Living and Working with AI; Epoch AI, LLM inference price trends.