Someone Still Has to Build the First Rung

Starfleet has an academy for a reason. You don't hand an ensign a tricorder and assume you've trained an officer.

Yet “give them an AI account” is being asked to carry an extraordinary amount of weight in the discussion about preparing people for work. Access matters. So does what happens after the welcome email.

The OECD's July 8 paper, Skills in the AI Age, treats that preparation seriously. It distinguishes exposure to AI from the likelihood of losing work to it, and describes adoption barriers that include cost, infrastructure, and shortages of skills. Smaller businesses face particular difficulties. Its recommendations include broad AI literacy and support for training throughout working life.

Useful. Now somebody has to turn that into a Tuesday afternoon in an actual workplace, with a beginner who needs help and a manager who already has too much to do.

What is the beginner supposed to learn?

Last month's jobs report left me thinking about the demand for senior judgment in junior roles. We can't keep treating experience as something that appears elsewhere, ready for us to hire when we need it.

Consider a new analyst preparing a purchasing recommendation for a department. An agent can produce a comparison of vendors, a plausible cost estimate, and a polished recommendation. The analyst can tidy it up, send it along, and look productive. Perhaps the recommendation is even right.

What has the analyst learned?

I'd rather give them the problem before giving them the agent's answer. Ask what they need to know about the department. Who will use the product? What does it need to work with? What would make the cheapest option an expensive mistake? Have them write a short plan, then let the agent help with the research.

Now there's something to compare. Perhaps the model found a useful alternative. Perhaps it recommended a product that won't meet a requirement the department forgot to mention. Perhaps the attractive price assumes a contract nobody wants to sign. A beginner needs practice noticing those possibilities and asking somebody about them.

That's where the mentor earns their time. “Why do you trust this price?” is a better teaching question than “Did you use the approved template?” Let the analyst follow the uncertainty back to a source, revise the recommendation, and explain the change. The uncomfortable moment when a neat answer falls apart can teach more than another hour of prompting tips.

Then follow the recommendation into use. Did the department buy it? Did the people doing the work get what they needed? If it caused a problem, let the learner see the problem and help fix it. Too many training exercises end the moment the document is submitted.

Pay for the person doing the teaching

I don't miss the idea that junior people should spend years on drudgery because their predecessors did. There was no educational magic in moving figures between spreadsheets. Some repetitive work did teach people how the organization operated; some was simply a bad use of their time. We can be glad to lose the latter while finding better ways to teach the former.

The obstacle is often quite ordinary: nobody has made room for the supervision. A manager is told to produce more with fewer people, then to mentor the new hire in whatever spare time remains. The new hire gets an account, a collection of training links, and instructions to ask questions. Everyone is surprised when the arrangement disappoints.

If a company expects AI to save money, it can make an explicit decision about how much of that saving goes toward developing people. It can also discover that a proposed saving was overstated once the time needed to check and teach is included. That's useful to know before the budget is approved.

On July 22, Anthropic published the agenda for its $200 million Economic Futures Research Fund. Among the things it wants to test are apprenticeship and mentorship approaches for a world in which AI absorbs junior tasks. I'd like to see those studies compare actual workplace arrangements: paid supervision, access to tools alone, and conventional training. Follow the participants long enough to find out who can take on more responsibility.

A completion certificate won't answer that question. Neither will a participant saying the course was enjoyable. Can they investigate an unfamiliar problem? Can they recognize when the model is outside its depth? Will a colleague trust them to handle something without checking every step? That's the progress an employer and a learner should both care about.

Actionable Steps

Start with one paid junior role and one project small enough to supervise properly. Name the mentor, reserve time in their calendar, and decide which decisions the learner will make. Use the tools, review the reasoning together, and come back to the result after someone has lived with it.

Make participation possible. An unpaid placement rules out people who need wages. Evening-only training is a poor fit for someone with care responsibilities or a second job. Those are design decisions about who gets an opportunity, whether or not the program's brochure mentions them.

For educators, I'd put more weight on a student explaining how their answer changed. Let them use AI, then ask them to defend a conclusion and deal with a new constraint. They will need subject knowledge to do that. A person who knows nothing about a field has very little basis for challenging a confident answer about it.

We keep describing judgment as the human skill that will survive. Fine. Then we need to pay for the time and mistakes it takes to develop it. Otherwise we're congratulating ourselves on identifying the answer while declining to teach it to anyone.

LLAP. 🖖