The Path to Agency Runs Through Education

There is a fork in the road that almost every company deploying AI hits, and almost nobody recognises it at the time. The fork is between access and agency.

Access is what most companies are doing. Hand the team a license. Give them a chat window. Give them a button that says "draft with AI". Move on. Call it a rollout.

Agency is something else entirely. Agency is when the user understands what the tool is, what it can and cannot do, where its blind spots live, and how to think about its output. Agency is the difference between someone who uses AI and someone who is meaningfully empowered by it.

Access is cheap and fast. Agency is slower and more expensive. The companies that win the next decade will be the ones that paid for agency anyway.

Here is what happens in an access-only rollout, and I have watched this play out at small scale and large scale across very different sectors.

Week one: enthusiasm. People generate things. The novelty is delightful. Output goes up. A small number of power users emerge spontaneously.

Month two: stratification. The power users are pulling away. Everybody else is using the tool for the most superficial possible tasks — rewording emails, summarising things they could have read. The deep work is going untouched because the deep work requires understanding what AI is good and bad at, and most users do not have that understanding.

Month six: ambient mistrust. The tool has produced enough plausible-sounding but wrong output that users have learned, the wrong way, to trust it for everything or to trust it for nothing. Both are bad. The middle position — trust it for these things, verify it for those things, never trust it for these other things — requires education that was never delivered.

Agency with AI is built from four ingredients, and they're all teachable.

One. A working mental model of what the tool is. Not a technical understanding — most users don't need to know what a transformer is. A functional understanding: the model has no memory across conversations, it has no real-time knowledge by default, it produces fluent prose whether it knows the topic or not, it will confidently invent specifics if it doesn't have them. These are the four things every user needs in their bones before they touch the tool for serious work.

Two. A sense of where it's strong and where it's weak. Strong: drafting, restructuring, summarising patterns it's seen many times, generating options. Weak: anything requiring exact recall of obscure facts, anything requiring novel reasoning at the edge of human capability, anything requiring lived context. The boundary is not stable — it moves every six months — but the shape of the boundary is what matters.

Three. A discipline around verification. What gets checked, by whom, against what reference? This is where most teams fail catastrophically. The output is checked once, gets approved, and the user concludes that future outputs of the same shape don't need checking. The slippage compounds. Quality decays. Trust evaporates.

Four. A vocabulary for collaboration. How do you ask for what you want? How do you correct the model when it's wrong? How do you steer it toward your standard? This is the prompting layer, and while it's the most visible part of AI literacy, it's actually the least important of the four. The other three matter more.

Building the four ingredients above takes real time. Not a forty-five minute training session. We're talking about weeks of progressive exposure, with structured practice, with feedback, with the chance to see good and bad output side by side and develop a sense for the difference.

Most companies will not pay this. They'll calculate the loaded cost of taking the team off-task for that long, conclude it's prohibitive, and revert to the access-only model. They'll get the access-only outcomes I described above, and they'll wonder, eighteen months later, why their AI deployment hasn't moved the metrics.

The companies that do pay for it find something interesting. The team's baseline capability changes. They use the AI better, yes — but they also make better decisions in general. They've internalised a set of skills (checking sources, articulating intent, recognising fluent-sounding nonsense) that travel beyond the tool. The investment in AI literacy ends up being an investment in thinking literacy. The return is much larger than the AI return alone.

Don't run a one-shot training event. They don't work. People forget within three days and revert to whatever habit was easiest.

Do this instead. Build a small program — six to eight weeks — that runs alongside people's actual work. Every week, a short concept (forty minutes), a structured exercise on a real task from the user's own job (an hour), and a debrief in pairs about what worked and what didn't (thirty minutes). That's two and a half hours a week for six weeks. The math is fine, and the outcomes are extraordinary.

By the end of six weeks, you don't have users who can use the AI tool. You have users who can think about AI — which means they can adopt the next tool, and the one after that, without you having to retrain them. You have built durable capability rather than fragile dependence. The difference shows up in every metric you care about, and it shows up forever.