‘AI Agents Are Like Junior Employees’ — What That Line Really Means for You

Every few months a new phrase enters the conversation about work, and right now the phrase is “AI agents are like junior employees.” It sounds friendly, almost generous. But when someone tells you that AI agents replace jobs the way a new hire helps out — a keen assistant who never sleeps, never complains, never asks for a raise — it is worth slowing down and asking what that comparison is actually doing. Because a metaphor is never neutral. It tells you where to look, and just as importantly, where not to. And this one has been chosen with some care.

What an AI agent actually is

Start with the plain mechanics, because the marketing tends to skip them. A chatbot answers. You type a question, it produces text, and the exchange ends there. An AI agent is aimed at something more ambitious: it is a system built to act, not merely respond. Give it a goal — “reconcile these invoices,” “draft and send the follow-up emails,” “pull the data, build the report, flag the anomalies” — and it breaks that goal into steps, takes each step, checks the result, and moves to the next. It can call other software, browse, run code, and loop back when something fails.

The key words are multi-step and semi-autonomous. A calculator does one thing when you press a button. An agent is increasingly designed to string many actions together with far less human hand-holding in between. That is the genuine shift. It is also why the technology is being pointed not at single tasks but at whole sequences of routine knowledge work — the bundles of small, repeatable jobs that make up a real role.

I want to be careful here, because this is a field thick with hype. These systems are improving, but they are not the tireless, flawless colleagues the pitch decks imply. They make confident mistakes, they lose the thread on long or unusual tasks, and they still need supervision for anything that matters. “Aimed at” is doing honest work in these sentences: the ambition is real, the reliability is uneven, and the gap between demo and daily reality remains wide. Keep both facts in view. Overstating what agents can do is how people get talked into bad decisions — and so is pretending nothing is changing.

The junior employee, by design

So why “junior employee”? Why not “automation,” which is what it plainly is? Because “automation” makes you think of the thing it replaces, and “junior employee” makes you think of the thing it helps. The framing is reassuring precisely because it is built to be reassuring. It invites you to picture the agent sitting beside you, taking the drudgery off your plate, freeing you up for the interesting work. Who could object to that?

Notice the sleight of hand. A junior employee is someone you mentor, someone who is learning, someone who will one day be your peer. The relationship implies a future and a mutual stake. Software has none of that. It will not grow into a colleague, it will not organise with you, it will not one day ask for the corner office. The metaphor borrows the warmth of a human relationship to describe a purchase of a tool — and in doing so it quietly reframes a substitution as a gift.

“Junior employee” makes you picture the agent beside you, taking the drudgery off your plate — which is exactly why the phrase was chosen over the plainer word: automation.

This is a pattern I keep returning to, because it is how technology gets captured again and again: a genuinely useful capability is wrapped in language that hides who is actually being served. The spinning jenny was sold as relief from tedious labour too. The relief was real. So was the wage collapse for the weavers who lost their bargaining power. Both things were true at once, and the cheerful framing made sure everyone looked at the first and not the second.

What it means for jobs — and what it doesn’t

Here is where the honest, un-hyped account has to hold two things together. AI agents are not about to delete every office job. But they do put real pressure on a specific and large category of work: bundles of routine, rules-based knowledge tasks. Processing claims. Drafting standard documents. Triaging tickets. Moving data between systems and checking it as it goes. The more a role is made of predictable, repeatable steps, the more exposed it is — which is a longer version of the question I explore in will AI take my job.

But — and this is the part the metaphor works hardest to obscure — exposure is not destiny. Whether an agent augments a worker or substitutes for one is almost never decided by the technology itself. It is decided by an employer. The same tool that could let a team of five do the work of eight, and spend the freed time on harder problems, could instead let two people do the work of five while the other three are let go. Nothing in the code chooses between those futures. A person with a budget does.

That is why the “augmentation versus substitution” debate, argued in the abstract, tends to go nowhere. It treats the outcome as a property of the machine when it is really a property of a decision — a decision about how to distribute the gains. And decisions have interests behind them. Which brings us to the question the friendly framing is built to keep you from asking.

Follow the money, not the metaphor

If you want to understand any new technology at work, my advice is always the same: don’t argue about the metaphor, follow the money. So let’s do it here. Suppose the “junior employee” is real. Suppose an agent genuinely does a slice of your job — the routine slice — at a fraction of what it costs to employ a person to do it. That saving is not imaginary. It is the whole business case. So ask the only question that matters: where does the saving go?

Trace it. The company was paying a salary for that work. Now it pays a software subscription that costs far less. The difference — the gap between the old wage bill and the new licence fee — does not evaporate. It goes somewhere. It can go to the workers who remain, as higher pay or shorter hours. It can go to customers, as lower prices. It can go to shareholders, as profit. It can go to the software vendor, as rent. In practice, without anyone deliberately fighting for one of the first two, it flows overwhelmingly to the last two.

The productivity is generated at your desk. The savings are collected somewhere else entirely. That gap between where value is made and where it lands is the whole story.

This is the move the “junior employee” language exists to prevent. If you are picturing a helpful teammate, you are not picturing a transfer of income — from labour, which used to capture some of the value it produced, to capital, which now captures more of it. The productivity is generated at your desk, or where your desk used to be. The savings are collected on a balance sheet you will never see. The metaphor keeps your eyes on the helpful colleague so you don’t notice the accountant behind it.

And notice who is most exposed at the entry point. The routine, learnable tasks that agents target first are exactly the tasks juniors have always done — the rungs people used to climb on their way to expertise. Hand those to software and you don’t just cut a cost; you saw off the bottom of the ladder, which is the harm I trace in the vanishing entry-level job. The roles most in the firing line are, revealingly, the ones the metaphor names. It calls the agent a junior employee because the junior employee is who it is aimed at replacing.

Which work goes first

None of this lands evenly. The exposure concentrates where work is most codifiable, which is why the earliest pressure shows up in specific corners of the office rather than everywhere at once — the pattern I lay out in which office jobs AI replaces first. Broadly, the more your day is made of judgement, relationships, physical presence, and accountability for outcomes, the harder it is to hand to an agent. The more it is made of standardised information handling, the easier.

But even that line is not fixed by nature. It moves with the technology, and it moves with what organisations decide is “good enough.” A firm willing to accept more errors, or to push the checking onto customers, will automate more aggressively than one that isn’t. So the boundary between augmentation and substitution is not a scientific fact waiting to be discovered. It is a series of choices, made mostly by people whose incentive is to capture the saving rather than share it.

What workers and organisations could do differently

This is not a counsel of despair, and it is not a counsel of denial either. It is a counsel of clarity. If the outcome is decided by choices rather than by the machine, then the outcome can be contested — but only by people who can see what is actually being decided.

For workers, the first move is to stop competing with the agent and start doing the things it is worst at. Judgement under ambiguity. Responsibility that someone must actually own. The relationships and trust that no system can hold on your behalf. Learn to direct these tools rather than race them, because the person who supervises the agent captures more of its value than the person whose task it swallowed. And — unfashionable as it sounds — the bargaining power that comes from workers acting together is one of the few forces that has ever reliably turned a productivity gain into shared benefit rather than a private windfall. The gains from every past wave of automation were split the way they were split because of who could push back, not because of what the machines could do.

For organisations, there is a genuine choice available, and it is worth naming plainly. You can use agents to shrink the workforce and pocket the difference — the default, and the easy path. Or you can use the same tools to raise what each person is capable of, take on work you couldn’t before, and share enough of the saving that the people doing the work have a reason to make the whole thing succeed. The second path is not charity. It is how you keep the knowledge, the trust, and the loyalty that the routine-task view of work always undervalues right up until the moment it’s gone.

The technology is real, and some of it is genuinely useful. That was true of the power loom and the spreadsheet too. The question was never whether the tool worked. The question, every single time, is who takes the gains, who pays the cost, and who is in a position to fight back. “AI agents are like junior employees” is a lovely sentence for keeping you from asking it. Ask it anyway.