There's a genre of article that's become impossible to avoid this year. You know the one. It lines up a stack of AI tools — a writer here, a scheduler there, a research assistant, a design bot, a bookkeeper — adds up their monthly subscriptions, and announces that for less than you spend on coffee you can now replace a $50,000 hire. TechRadar ran the latest one in May: eleven tools, one imaginary salaried employee, made redundant before they were ever hired.

The maths always looks unanswerable. And if you run a small business, the temptation it plants is real. You glance at your payroll, you find the line that most resembles the job the list describes, and you start to wonder whether you could let it go.

I'd think very hard before you did. Not because the tools don't work; most of them do, and rather better than the people selling them manage to convey. It's because the list has performed a sleight of hand. It has priced a job as though a job were a pile of tasks. It isn't. A salaried role is three different things stacked on top of each other, and only one of them is actually for sale.

What a salary actually buys

Start with tasks — the visible, describable work. Draft the email, build the deck, chase the invoice, schedule the posts, pull the numbers into a report. This is the part the lists can see, because it's the part you could write down. It's also, not by coincidence, the part AI has become genuinely good at — the describable, repeatable work that used to be scarce and now isn't. If a task can be specified clearly enough to appear in a job advert, there's a fair chance it can be specified clearly enough for a model to have a proper go at it. On tasks, the $50k lists are broadly right.

Next comes judgement — knowing which task to do, in what order, and spotting when the sensible-looking answer is the wrong one for your particular situation. Here the tools are much weaker, for reasons I've dug into before when writing about the judgement gap: AI is a confident adviser and a poor decider, because it can't see the context that actually determines the right call. It doesn't know which customer has been wobbling for a month, or which member of staff is about to hand in their notice, or that the supplier quoting you is the one you'll need a favour from in November. It produces a plausible answer from what's in front of it. Whether that plausible answer is a good one is a separate skill, and it stays with the human.

Last comes accountability, and nobody puts it on the list because you can't buy it in monthly instalments. When the campaign goes out with the wrong price on it, when the client is furious, when a decision got made on top of numbers that turned out to be wrong, someone has to own the outcome: carry the consequence, learn from it, and be trusted a little differently the next time. A tool can't do that. You can't sack it, you can't dock its bonus, you can't sit across a table from it and agree how you'll both make sure it doesn't happen again. When you replace the salary, the accountability doesn't vanish with it. It transfers straight back to you.

An old mistake, running backwards

The owner who reads the tool list and reaches for the payroll is making the mirror image of a very old mistake.

Picture the same owner in 2019, hiring a head of marketing off an impressive CV and a good interview. They treated the role as a box to be filled: here is the job description, here is a person who matches it, done. A year and fifty-odd thousand pounds later, they discovered that filling the box and solving the problem were never the same thing. The description was a shadow the real work cast — and they'd hired the shadow.

Firing a salary because a stack of tools matches the job description is that error played in reverse. You're still mistaking the description for the work. In 2019 it cost you a bad hire. In 2026 it costs you the judgement and the accountability that the job description never mentioned, because job descriptions never do. They list the tasks. The tasks were always the easy part to name and, it turns out, the easy part to automate.

A better question than the lists ask

These articles are built to answer one question: what can these tools do? The honest reply is "more than you'd think." A more useful question, though, is narrower and much more your own: which £50,000 of work in my business is judgement-light, repeatable, and reversible if it fails?

Three tests, and the work has to pass all three. Judgement-light means the right answer doesn't lean much on context the tool can't see — the output is more or less the same regardless of what's going on in the corners of your business. Repeatable means it happens often enough, and similarly enough, that a machine's consistency is an asset rather than a liability. Reversible means that if it goes wrong you'll find out quickly and cheaply, and you can put it right before it compounds.

I won't pretend the line is always obvious. Plenty of work sits in the grey middle, and I don't always call it correctly in my own business — I've handed things to a tool that turned out to need a person's read, and kept people on work a model would have done just as well. But three rough tests you apply honestly beat a subscription price you don't interrogate at all.

Work that passes all three — first-draft copy, social scheduling, transcription, tidying data, the monthly report nobody enjoys writing — you should hand to the stack, and hand over enthusiastically. This is exactly what the tools are for, and keeping a person on it out of habit is its own kind of waste.

Work that fails the tests is a different animal. It looks like tasks from the outside, which is what fools the lists, but it's really a person carrying judgement and consequence, with the visible tasks coming off as a by-product.

Take whoever does your invoicing. On paper it's tasks: raise the invoice, send the reminder, reconcile the payment — all automatable, all on the list. But the good one does more than invoicing. They decide which slow-paying client gets a gentle nudge and which gets a firm one; they notice that the customer who's suddenly late is the same one who mentioned a cash-flow wobble last month; they know when a debt is better walked down the corridor than chased by email. Automate the raising and sending, by all means. But the day you let that person go, the judgement doesn't leave with the tasks. It lands on you, between everything else, usually at the moment you can least afford to get it wrong.

Even the showcase companies can't tell the difference

This confusion travels a long way up the food chain. Earlier this year Amazon set targets pushing most of its developers to use AI tools every week and tracked the usage on internal leaderboards. Some employees, feeling the pressure, started running pointless jobs through the AI purely to move up the board — a habit that got the nickname "tokenmaxxing." The metric was measuring activity, not value, so people gave it activity. Amazon reportedly retired the leaderboard once it worked out what it was actually rewarding.

The $50k lists make the very same error, just from the other end. They count the tasks a role performs and call that its value. Activity is easy to count; judgement and accountability aren't, so they get left off the spreadsheet — and once they're off the spreadsheet, they quietly drop out of the decision too. If a company Amazon's size can mistake activity for value on a leaderboard, the same trick is far easier to fall for on a payroll line, where the judgement and the accountability were never written down in the first place.

None of this is an argument against the tools. Buy them, and buy them with your eyes open. Build the stack. Hand it every scrap of work that's judgement-light, repeatable and reversible, and you'll free up real money and real hours. But keep the salary for the work that's none of those things, because that work never was a task; it only looked like one from the outside. The businesses that come out of this era ahead will be the ones that looked at each salaried role, worked out which part was tasks and which part was judgement, and then automated the first while protecting the second.