AI literacy at work

The gap in that one number is the whole story. Almost everyone has the tools. Almost no one can point to the money. A separate MIT report put it more bluntly: roughly 95% of generative-AI pilots delivered no measurable return, despite an estimated $30–40 billion in spending (MIT NANDA, The GenAI Divide: State of AI in Business 2025).

It’s tempting to read that as a tooling problem. Buy a better model. Roll out more licences. It isn’t. The companies stuck on the wrong side of that gap mostly have the same tools as the ones winning. What they’re missing is people who know how to use them with judgement, and a clear idea of what they were trying to earn in the first place. The distance between AI adoption and AI value is a literacy gap, not a tooling gap.

Here’s the uncomfortable part of the prompt-library approach: the tools are changing constantly. The interface you trained your team on this spring is already different. Next month it’ll be something else. If your AI literacy is a stack of memorized prompts, your literacy has a shelf life measured in weeks.

The gap between “learn AI” and “learn to think critically about AI” is small in wording and enormous in practice. One is a tutorial. The other is a stance: the ability to look at any AI output, in any tool, and know roughly how much to trust it and where to check. The stance transfers. The tutorial doesn’t.

What’s actually durable

Underneath every chatbot is the same basic machine: a system that predicts the next chunk of text by statistical likelihood, one piece at a time. It is a world-class guesser, not a librarian. It has no idea whether what it’s saying is true. That’s not a flaw someone will patch next year. It’s how the architecture works. Once that clicks, a lot of workplace behaviour changes. You stop treating the output as an answer and start treating it as a draft. You verify the things that matter. You keep your name on the final call.

None of that depends on which logo is in the corner of the screen.

There’s a study I keep coming back to, because it’s the cleanest version of the warning. Researchers at Wharton ran a field experiment with nearly 1,000 high-school maths students in Turkey. The group using a standard ChatGPT-style tutor scored 48% better than the control group while they had the tool in hand. Then the tool was taken away for the exam, and that same group scored 17% worse (PNAS, 2025; summarized by Knowledge at Wharton). The AI hadn’t taught them anything. It had done the reps for them.

If a robot did your reps at the gym, your muscles wouldn’t grow. The discomfort of working something out yourself isn’t a bug in learning. It is the learning. And that doesn’t stop being true the day you graduate and start drawing a salary.

At work the offloading is quieter, but the mechanism is identical. You let the model draft every client email, and slowly you lose the feel for how to frame a hard message. You let it summarize every report, and you stop reading closely enough to notice what’s missing. The KPMG 2025 global study (over 48,000 people across 47 countries) found that 66% of people rely on AI output without checking its accuracy, and 56% have made mistakes at work as a result (KPMG, Trust, attitudes and use of AI, 2025). That’s not a story about bad tools. It’s a story about atrophied judgement.

The fix isn’t to avoid AI. It’s to keep doing the part that keeps you sharp. Friction is valuable. Use AI to go faster on the work you already understand. Be careful about using it to skip the work that’s still teaching you something.

This is the practical core. Three habits do most of the work, and none of them are tied to a specific product.

Treat every output as a first draft, not a final answer

The people getting real value from AI use it to think faster, then apply their own judgement on top. They never let it have the last word. AI is a fast first draft. Real learning (and real accountability) is in the checking. Treat what comes out of the chat window as a starting point for your thinking, not a replacement for it.

Verify outside the chat window

Here’s the habit most people get wrong: they spot a suspicious claim and ask the AI “are you sure?” That is not a fact-check. The model can be confidently, consistently wrong, and asking it to grade its own homework usually just buys you a more confident version of the same mistake. Verification happens outside the chat:

• →Open the primary source. Confirm the document, the number, the quote actually exists.
• →Click the link. Models invent citations that look real and lead nowhere.
• →Mind the training cutoff. The model may be living in the past on anything recent.
• →For anything that matters (financial, legal, medical, regulatory) confirm with a qualified human who can be held accountable.

Don’t outsource your judgement, and don’t paste in sensitive data

In financial services this one isn’t optional. When you use a free chatbot, you are essentially speaking in a public square. Unless you’re a paying customer with the right privacy settings on, your inputs can be used to train the model. So never paste real client data, confidential financials, or anything you wouldn’t want on a billboard. When you need to test something against real numbers, build a fictional, what-if version instead.

Watch for the signs you’ve started outsourcing the thinking:

• →You're shipping AI output you couldn't defend if questioned line by line.
• →You've stopped noticing when a draft sounds right but says nothing.
• →You'd struggle to do the task at all if the tool went down tomorrow.
• →Your team's writing has flattened into one polished, point-less voice, the beige wallpaper of content.

If you own AI literacy for an organization (HR, L&D, a team lead handed the rollout) the temptation is to commission a course on the tool of the moment. Resist it. By the time it ships, it’s already dating.

Teach judgement, not keystrokes

The thing that transfers is the understand–use–question loop: enough of how the system works to know where it breaks, the practical habit of using it well, and the verification reflex to catch it when it’s wrong. You can’t question what you don’t understand, and you can’t use what you haven’t questioned. Build that, and your people can walk into any tool next year and stay safe.

Measure outcomes, not activity

Most organizations track AI usage, because usage is easy to count. That’s how you get tokenmaxxing: people firing off prompts and generating outputs no one reads, just to look AI-engaged on a dashboard (Human+AI). The moment a measure becomes a target, it stops being a good measure. Measure the outcome, not the activity. Did you serve customers faster? Sell more? Cut a real cost? Companies don’t make money by using AI. They make money by using AI for something.

Don’t make the rollout the source of the burnout

Piling on tools has a cost people underestimate. A March 2026 BCG study of close to 1,500 full-time workers found that those required to closely oversee AI (reviewing, correcting, interpreting its output) reported 12% more mental fatigue and 19% more information overload than colleagues with lighter AI workloads (BCG, When Using AI Leads to “Brain Fry”). More tools is not more capability. Teach fewer tools, deeper, with clear rules for what each one is actually for.

Close the recognition gap

There’s an equity problem hiding in most rollouts: research has found men praised and women penalized for the same AI use. The fix isn’t telling women to lean in harder. It’s building a system that rewards the right behaviour for everyone, actively encouraging people to experiment, and recognizing AI used for genuine innovation instead of treating it as a shortcut to be suspicious of.