Essay
The public debate on Artificial Intelligence has settled into two lazy camps. One says the machines are coming for your job. The other says relax, the tools just make everyone a bit more productive. Both make the same mistake: they treat the technology as the variable and the human being as a constant. It runs the other way round.
These systems are amplifiers. They take whatever judgement, taste, clarity and rigour you bring to them and multiply it. If there is not much there to begin with, they multiply that instead — just faster, and now in more languages.
Why the same tool produces opposite results
Two people open the same model. One knows exactly what problem they are solving, what a good answer looks like, and which part of the work actually creates the value. They use the machine to compress the mechanical middle and spend their own time on the decision. Their output jumps by an order of magnitude.
The other has no real fix on what good looks like, so they take the first plausible answer and ship it. The work comes out faster, longer, better formatted, and utterly undifferentiated. Nothing became more productive here. It just became more efficiently average.
The machine does not supply the standard. You do.
The skill that got expensive
Production, variation and iteration are close to free now, and that collapse in cost has repriced everything sitting next to it. Making things stopped being the valuable part. Knowing what deserves to exist, and telling a coherent answer apart from a correct one, is where the value moved to.
Most organisations are walking straight into this trap because these systems are fluent by design, and fluency reads as competence — especially to someone without real depth in the subject. So teams wave through outputs they would have sent back to a junior colleague for another draft, simply because the machine said it with more confidence than the junior ever would have dared.
- Would I have accepted this from a person on my team?
- Do I understand it well enough to defend it in a room where someone disagrees?
- Is this a real answer, or a very well-written average of the internet?
Delegate the work, never the judgement
There is a clean line between tasks and decisions. These tools belong on the task side: drafting, summarising, translating, structuring, exploring options, doing in ten minutes research that used to swallow a week. Keep them off the decision side: what we refuse to do, what we are willing to be wrong about, what we tell a customer when the news is bad, who earns a promotion.
Leaders who let that line blur lose something painfully slow to rebuild. Judgement is a muscle, built by making calls, living with what follows, and adjusting. Outsource enough of that struggle and in two years you have a senior team with excellent tooling and no instinct — exactly the profile the market stops paying for.
Amplification applies to organisations too
A company with a clear strategy, honest data and people who are actually allowed to disagree will use these systems to move a great deal faster. A company with a confused strategy, political reporting lines and a culture of covering yourself will use the same systems to generate more documents defending decisions nobody wants to own. The technology does not fix the operating culture underneath it. It exposes it.
That is why serious deployments end up as leadership problems rather than technology ones. The tools work fine. What is in question is whether the organisation underneath is worth multiplying at all.
The objection: won’t everyone just have the same tools?
The obvious rebuttal is that access levels the field — if every competitor can reach the same model, the advantage should cancel out, the way calculators never made one accountant better than another. That only holds if the tool itself were the scarce input. It never was. What was scarce was always the standard applied before the prompt gets written, and the judgement applied after the answer comes back, and neither of those changes because everyone now has the same access.
What actually happens when a whole industry gets the same amplifier is that the gap between operators widens instead of closing. The average performer produces more average work, faster, and the market floods with it. The rare operator with real judgement stands out more starkly against that flood, not less, because the baseline of “adequate” has risen and adequate has stopped earning anyone’s attention. Shared access does not equalise outcomes. It raises the price of being genuinely good and collapses the price of being merely fast.
The second-order cost: atrophy you cannot see happening
The first-order risk of over-delegating judgement is a bad decision shipped this quarter. The second-order risk is worse and much harder to catch, because it says nothing on this quarter’s numbers at all. It shows up in the capability of the people making decisions three years from now.
Judgement is built the way every other skill is built: through repeated, unassisted attempts that sometimes fail, with the person who made the call carrying the consequence and adjusting. Every time that struggle gets outsourced to a system that hands over a fluent answer on demand, the rep is lost. Nobody notices in month one. The organisation notices in year three, when a genuinely novel problem shows up with no precedent for the machine to remix, and the room discovers nobody there has made an unassisted call in a long time.
It is a harder problem than a bad quarter precisely because it stays invisible on every dashboard that measures output, and output is exactly what keeps looking healthy while the capability underneath quietly erodes.
What this means for how you hire and promote
Most hiring processes still screen for fluency — can this person produce a clean document, a tidy analysis, a well-structured plan. That used to be a fair proxy for competence, because producing those things well was genuinely hard. It is not hard anymore. Fluent output is now the cheapest thing in the building, and screening for it mostly tells you whether a candidate knows how to prompt.
What should replace it looks closer to an interrogation than an interview: hand the candidate a plausible-looking answer, including one a machine produced, and ask them to find what is wrong with it. Their ability to spot the flaw, defend the correction and explain the trade-off tells you far more about the judgement you are actually trying to hire, and it is the one thing the amplifier cannot fake on their behalf.
The same logic should govern promotion. Promoting someone because their output volume went up is rewarding the wrong variable. Promote the person whose filter improved — whose rejection rate on their own drafts went up, who can tell you precisely why the ninth version was wrong and the tenth was not.
There is a quieter version of this test for leaders themselves, not just for the people they hire. If most of what leaves your desk in a week is a fluent draft you approved rather than a call you actually made, the amplifier is running through you without you supplying anything worth amplifying. That is a comfortable way to stay busy, and a fast way to become replaceable, because a comfortable, fluent approval is precisely what these systems are best at generating without you.
The organisational habit worth building now
Individual discipline matters but it does not scale by itself. What scales is a small number of explicit rules that survive staff turnover: which categories of decision always require a named human sign-off no matter how confident the machine’s draft looked, which reviews are mandatory before a customer-facing output ships, and how junior people still get hard, unassisted first attempts so the muscle described above actually gets built instead of quietly skipped.
None of that requires slowing down. It requires deciding, once, where speed is free and where it costs you, and holding that line while everyone around you is moving faster on both.
The practical position
None of this is a case for caution. The people getting the most out of these systems are rarely the careful ones — they are the ones with a standard high enough to reject nine outputs and recognise the tenth. Volume of use matters far less than the quality of the filter behind it, and the filter is entirely human.
Use these systems hard. Refusing them on principle is not integrity, it is a handicap you have chosen to carry. But treat every output as a draft from an extremely fast collaborator with no stake in the outcome, no memory of your customers and no career risk if it turns out wrong.
And invest in the thing being multiplied. Deepen the domain knowledge. Sharpen the taste. Practise making decisions under uncertainty. The multiplier is generous and completely indifferent — it will scale excellence and mediocrity with exactly the same enthusiasm.