Most people I talk to about AI say roughly the same thing. “Yeah, we use it.” And they do. They use it to tidy up emails, summarise meeting notes, and occasionally ask it a question they could’ve Googled. Eighty-two per cent of teams now use AI at least once a week, according to recent workplace studies. Sounds impressive until you look at what they’re actually doing with it.

They’re riding a Ferrari like it’s a bicycle.

The access trap

We’ve solved the access problem. ChatGPT has hundreds of millions of users. Copilot’s baked into Office. Claude’s on every developer’s screen. The tools are there, they’re good, and they’re getting better fast. But access isn’t the same as ability, and confusing the two is where most organisations are quietly bleeding productivity.

Jeff Bullas made this point well recently: the real risk of AI isn’t misuse, it’s unskilled use. Not dangerous, not malicious. Just underwhelming. People default to the simplest possible interaction because nobody showed them what else is possible. Three-quarters of all ChatGPT conversations fall into just three categories: practical guidance, seeking information, and basic writing. That’s a search engine with better manners.

The gap between what these tools can do and what most people ask them to do is enormous. And it’s growing, because the tools are improving faster than the skills.

Piano scales and power chords

I think about this like learning piano. You can sit down and plonk out a melody with one finger. It works. You get a tune. But if you’ve never learned scales, chord progressions, or how to read a lead sheet, you’ll never play anything that surprises you. You’ll just keep plonking.

Most AI use in business right now is one-finger plonking. “Summarise this document.” “Write me an email.” “What does this acronym mean?” These are fine tasks. They save a few minutes here and there. About 3.5 hours a week on average, the research says. But they’re the equivalent of using a grand piano as a doorstop.

The people who’ve learned the scales are operating in a different league. They understand how to structure a multi-step prompt, how to feed context properly, how to use AI as a thinking partner instead of a search box. Not because they’re smarter, but because they’ve put in the practice. OpenAI’s own data shows that power users who’ve adopted features like custom instructions, projects, and reasoning models get qualitatively different outputs. They’re not using a better tool. They’re using the same tool better. Context engineering — the discipline of designing the information architecture around the question — is the most practical way to close that gap.

What the gap actually costs

IDC reckons AI skills shortages could cost the global economy $5.5 trillion by 2026. That’s a staggering number, but the local version is more interesting. In any given organisation, the person who knows how to use AI well is — conservatively — twice as productive on knowledge work as the person who doesn’t. Research from multiple studies suggests workers with advanced AI skills earn 56% more than peers in the same roles.

The gap isn’t evenly distributed either. Technical teams (IT, analytics, R&D) have pulled ahead. Frequent AI use among leaders has jumped from 17% to 44% since 2023. But in marketing, operations, sales, and finance, adoption remains patchy. These are exactly the functions where AI could make the biggest difference, and where the literacy gap hurts most.

The uncomfortable part? Employers aren’t fixing it. Only 26% of organisations now offer formal AI upskilling programmes, down from 35% a year ago. Forty-two per cent of employees say they’re expected to figure it out on their own. That’s not a training strategy. That’s a hope strategy.

Why self-teaching doesn’t scale

Most people learn AI tools through trial and error. They open ChatGPT, type something, get a response, and form an opinion based on that single interaction. If the response is mediocre (and it often is, because the prompt was mediocre) they conclude the tool’s overhyped and go back to doing things manually.

This is the self-teaching trap. Without structured guidance, people don’t discover what they don’t know. They never find out that AI can spot patterns in messy data, pressure-test a strategy through Socratic questioning, or do a first-pass market analysis that would take a human three days. They just keep asking it to fix their grammar.

I’ve watched this play out in real time. Someone on a team I was advising spent 20 minutes reformatting a spreadsheet manually. When I showed them how to upload it and ask Claude to restructure it with a specific schema — including error checking and a summary of anomalies — they were genuinely stunned. Not because the tool was new to them. They’d been using it for months. They just hadn’t thought to ask it to do anything beyond simple text tasks.

It’s like giving someone a professional kitchen and watching them make toast. The oven’s right there. The knives are sharp. But without someone showing them a recipe or two, toast is all you’ll get.

Show, don’t train

The organisations I’ve seen close this gap don’t do it with mandatory training courses or “prompt engineering” workshops. Those tend to produce a flurry of activity for a week and then nothing.

What works is embedding AI use into existing workflows with specific, contextual examples. The sales team preps for a call using the prospect’s last three earnings transcripts — fed to AI with the right questions, it produces a briefing in two minutes that used to take an hour. The marketing team turns a brief into five headline variants with reasoning about why each one works. Finance asks AI to find anomalies in a dataset and explain them in plain language.

The pattern is always the same: take a task the team already does, demonstrate how AI changes it from a 40-minute job to a 5-minute job, and let people experience the difference firsthand. That’s not training. That’s proof.

The other thing that works is giving people permission to experiment. In too many organisations, there’s an unspoken anxiety about looking stupid — about asking AI something “wrong” or wasting time on a tool that might not work. The best AI-literate teams I’ve seen treat experimentation as part of the job, not a distraction from it. They share prompts, compare outputs, and build a collective understanding of what works. It spreads faster than any training programme.

The 56% question

That stat about advanced AI users earning 56% more isn’t going away. If anything, the premium will grow. We’re at the point where AI literacy is splitting into two tracks — one group that treats these tools as a genuine extension of their thinking, and another that treats them as a slightly faster way to write emails.

Both groups have access to the same Ferrari. One’s cycling. The other’s driving. And the gap between them is compounding — because the people who use AI well get better at it faster. They learn what to ask, how to iterate, when to push back on a weak output. The people who use it badly just confirm their own low expectations.

The fix isn’t complicated, but it does require intent. Pick three workflows in your team that take too long. Spend an afternoon showing people how AI changes each one. Then get out of the way and let them practice.

Nobody’s going to wake up one morning and suddenly discover all the things they’ve been missing. AI literacy doesn’t arrive through osmosis. It arrives through deliberate, specific, show-don’t-tell moments that make people go “wait, it can do that?”

Create those moments. The tools won’t teach themselves. But given a nudge in the right direction, people learn fast. Faster than most leaders expect.

And then they stop cycling and start driving.