The Viral 'AI Costs More Than Employees' Graphic Forgets to Measure Output
A graphic doing the rounds claims AI is 21× more expensive than a human employee. It's comparing the wrong numbers — and quietly omits the one variable that decides whether the bill is good or bad: what those AI-augmented teams actually produce.
The Viral 'AI Costs More Than Employees' Graphic Forgets to Measure Output
A graphic has been doing the rounds. Heavy black bar, red type: "WE REPLACED YOU WITH AI. THEN GOT THE BILL." Beneath it, two figures. A human knowledge worker on the left: $65,000 a year. A four-person AI startup's token bill on the right: $113,000 a month. A big red arrow renders the verdict — 21× more expensive.
It's a genuinely good piece of design. It's also measuring the wrong thing.
I got into exactly this debate with a friend last week. His take was the honest one most people land on: the comparison is bad, and the bill is probably made up anyway. He's half right. But the more interesting problem isn't whether $113,000 is accurate — it's that the graphic puts two costs side by side and quietly omits the only variable that decides whether either number is good or bad: what got produced.
What's quietly missing: one salary counted yearly, one team's compute counted monthly. Different headcount, different time window — and not a single word about what either side actually produced.
First, the numbers don't even line up
Forget output for a moment and just look at the units. One side is a single employee's annual salary. The other is a four-person company's monthly compute. Those aren't comparable quantities, and putting them under one red arrow does most of the rhetorical work.
Annualise the AI bill and it's roughly $1.36M a year — for a team of four. Per head, that's about $28,000 a month, or some $340,000 a year of compute per person. That's a large, real number, and nobody serious is pretending tokens are free.
But the graphic asks "how much did this cost?" and stops there. The question it never asks is the one a business owner actually cares about: $1.36M of compute bought you what, exactly?
A cost figure on its own is meaningless. Spending is only expensive relative to what it produces — and the graphic shows you the bill while hiding the invoice it paid for.
The missing variable is output
Cost per seat is the wrong denominator. Businesses don't buy headcount or token allotments; they buy work that ships — tickets resolved, code merged, contracts reviewed, drafts written. The right denominator is units of output. And on that denominator, the picture doesn't just shift, it inverts.
This isn't a hopeful assertion. Over the last few years we've accumulated a genuinely solid body of controlled studies — randomised trials and field experiments, not vendor decks — measuring how much real work AI-augmented people get done. Here's what they actually found.
A consistent thread runs through all of it: the gains are largest for the least-experienced workers. AI doesn't just make good people slightly faster — it pulls newer and lower-scoring workers up the curve fastest, narrowing the gap between your best people and everyone else. A four-person team isn't producing four people's worth of work. It's producing meaningfully more.
But it's not free money — the frontier is jagged
Here's where I part ways with the breathless version of this argument, and with my friend's "AI does 4× to 8× everyone's output" instinct. Blanket multipliers are fantasy.
The same BCG and Harvard study that found a 40% quality jump also found its opposite. Dell'Acqua and colleagues described a "jagged technological frontier": a boundary where some tasks fall neatly inside AI's competence and some — though they look just as easy — fall outside it. On tasks beyond the frontier, consultants who leaned on AI did worse than those who didn't, confidently shipping wrong answers.
So the honest read isn't "21× cheaper" any more than it's "21× more expensive." It's this: on the right tasks, with the work designed around the tool, real measured gains run from roughly 15% to 55%, occasionally higher — and biggest for your junior staff. On the wrong tasks, the gain is zero or negative. Output is the variable that matters, and it's a variable, not a constant.
Redoing the math the graphic skipped
So let's actually finish the calculation the graphic abandoned. Same numbers — but with the denominator a business cares about put back in.
Sixteen knowledge workers at $65,000 is roughly $1.04M a year in salary alone — call it $1.3M–$1.45M fully loaded once you add benefits, software, space and management overhead. Suddenly a $1.36M compute bill for a four-person team that ships comparable output isn't a punchline. It's in the same ballpark — and it comes with four humans instead of sixteen to hire, manage and retain.
That's not a claim that AI is always cheaper. It's the point that you can't know either way from the bill. The graphic shows you one input cost and asks you to feel something. The only number that settles the question — output — is the one it leaves out.
So, is the bill "too expensive"?
It depends entirely on what those four people shipped, which is precisely what the graphic refuses to show you. If $113,000 a month of compute lets a four-person team deliver what used to take a team of twelve, it's one of the best deals in the building. If it's four people generating volume nobody asked for, it's a disaster at any price.
The token bill, on its own, tells you almost nothing. Measure output — per task, against a real baseline, on the work AI is genuinely suited to. That's the number worth panicking about, or celebrating. The arrow on the graphic is pointing at the wrong one.
Trying to work out whether AI is actually paying for itself in your business — not in theory, but in shipped output? Start a conversation. We'd rather measure it with you than sell you an arrow.