McKinsey Did the Math: Generative AI Is Worth $2.6 to $4.4 Trillion a Year — but Whose World Does It Really Move?

A little while ago, McKinsey put out a sixty-eight-page report.

When I read that number, I did a double-take.

$2.6 to $4.4 trillion. A year.

To give you a feel for it: the UK's entire GDP in 2021 was about $3.1 trillion. In other words, generative AI alone could stuff another entire UK economy into the global economy every single year.

My first reaction: isn't that number a little too sunny?

It was only after reading the whole thing that I realized McKinsey hadn't pulled it out of thin air. They broke down sixty-three specific use cases, spread across sixteen business functions, then mapped those onto 850 occupations and more than 2,100 "detailed work activities" — and added it up, line by line.

This is a tally worth walking through with you.

What Does "Step-Change" Actually Mean?

Let's get one thing straight first: generative AI is not the same thing as the "AI" we've been shouting about for the past decade.

The "AI" you used to hear about was mostly discriminative — feed it a pile of data, and it would classify, predict, and optimize for you. Powerful, sure, but it only knew how to crunch numbers.

Generative AI is different. It can write, draw, hold a conversation, and generate code from a plain-language description. It has started doing a job that machines used to be completely unable to do: understand natural language.

Why does this matter?

Because nearly a quarter of all work activity worldwide requires more than a middling command of human language. Think about it — meetings, emails, customer service, sales calls, research, onboarding the new hire… none of these were jobs machines could touch before.

Now, they can.

McKinsey calls this wave a "step-change." Stairs, not a ramp.

75% of the Money Sits in Four Places

So where does that $2.6 to $4.4 trillion actually land?

McKinsey's answer is counterintuitive: it is heavily concentrated in four functions. Customer operations, marketing & sales, software engineering, and R&D.

Just those four, between them, swallow 75% of the total value.

Why these four?

Because they share one thing in common — they are naturally language-heavy, creation-heavy work. Customer service is talking, marketing is writing, programmers write code, R&D writes proposals. And that is exactly what generative AI is best at.

Let me give you a concrete example.

One company, with five thousand customer service agents, put generative AI to work. Afterward, issues resolved per hour rose 14%, average handling time per issue dropped 9%, and both agent attrition and the rate of customers asking to speak to a supervisor fell 25%.

The more interesting part is the next line: the biggest gains went to novice agents; the top agents barely moved, and even dipped occasionally.

Why?

Because the AI was teaching novices the "talking tricks" that veteran agents had accumulated. Novices were pulled up to the veterans' level. The veterans were already there — nothing left to pull them up to.

You see, this is a completely different logic from every automation story that came before. The machines of the past replaced repetitive, low-end work. This time, the machine is teaching people.

This Time, It's Coming for the High Earners

At this point, I have to pause.

For decades, every automation story has been the same version: the machine ate the low-skilled worker's lunch. Assembly-line workers, cashiers, loaders — the lower your education, the lower your wage, the more replaceable you were.

Generative AI flips that story on its face.

What it actually wants to touch most is the highly educated, well-paid knowledge worker.

Why?

Because its core capability is processing natural language and specialized knowledge. And that is precisely what lawyers, doctors, analysts, professors, engineers, and product managers — these "knowledge workers" — do every day.

McKinsey's own data: in 2017 they estimated a certain number for the automation potential of "applying specialized knowledge." When they recalculated in 2023, that potential had jumped thirty-four percentage points; "managing and developing people" leapt from 16% to 49%.

Translated into plain English: the "use your brain, lead people" jobs we used to think machines could never, ever do — the machine has now gotten its foot in the door.

What's it like?

Past automation was a leg-sweep — it cut the legs out from under the people standing low. Generative AI is a palm dropping from the sky — it presses down on the spire.

The Timeline Just Moved Forward a Decade

So when does all this happen?

McKinsey gives a window: half of today's work activities being automated will occur between 2030 and 2060, with a midpoint of 2045.

Sounds far off?

No — that number itself is the news. Because McKinsey's 2017 forecast put the midpoint at 2053.

A full eight years earlier.

Why the pull-forward? Because they re-ran the numbers and discovered something: the year in which technology reaches "median human level" had been pulled dramatically earlier. Natural-language understanding, for instance, was originally estimated for 2027. On a second look, 2023 had already arrived.

More critically, the total volume of working hours that can be automated jumped from the old estimate of "about half" to "60% to 70%."

In other words, six to seven out of every ten hours that today's workers spend on the job can, in theory, be done by a machine.

"In theory," that is. Whether it actually gets done depends on something else.

The Dividend Doesn't Cash Itself

After McKinsey finished tallying that ledger, they added one very important sentence.

The gist: generative AI could add 0.1 to 0.6 percentage points a year to global labor productivity growth. But to realize that gain, there is a precondition.

The people squeezed out by it have to be able to move into other work at "at least the same level of output as in 2022."

That one line, I think, is the most valuable sentence in the entire report.

Why?

Because the trillions in front of it are only an "upper bound." It assumes everyone gets properly resettled, every skill gets retrained, every career transition goes smoothly.

But reality, as we all know, doesn't look like that.

A customer service rep with ten years under their belt has their edge leveled by AI — where do they go for their next job? A copywriter whose drafts the AI can already push to B-minus quality — where do they move? A programmer whose code drafting is now being done by Copilot, 56% faster — what should their value be redefined as?

These are questions McKinsey doesn't answer. They simply got the ledger right.

But once the tally is clear, the truly hard part has only just begun.

Four Industries — A Look at How It Moves In

Enough abstractions. Let me give you a few concrete industry slices, and you'll see how this thing actually enters a business.

Banking. Another $200 to $340 billion a year. Where from? Banking is heavily regulated, knowledge-heavy, and built on customer conversations — three traits that land squarely on generative AI's strengths. Risk reports, regulatory tracking, client coverage, internal knowledge-base search — these are exactly what it does best. Morgan Stanley is already using GPT-4 to give sixteen thousand wealth advisors "an on-call internal expert," pulling relevant insights out of a mountain of research reports in seconds; one European bank is using a similar system to chew through long ESG (environmental, social, and governance) documents — even digging out the charts and tables to answer questions.

Retail and consumer goods. $400 to $660 billion. The leverage point here is "personalization." Stitch Fix is already experimenting with DALL·E: describe the color, the fabric, the style, and the AI sketches plausible designs, which stylists then match to the closest real items in inventory. It turns "browse the whole rack" into "tell me what I want."

Pharma and medical products. $60 to $110 billion. This one is the most dramatic. Pharma typically pours 20% of revenue into R&D, and a new drug takes, on average, ten to fifteen years. One step in that process — "lead compound identification," the early hunt for a promising drug candidate — used to take traditional deep learning months. Generative AI compresses it to weeks. Behind that is money. But even more, it's lives.

You see, all three industries share one thing: their value comes from generative AI doing something machines previously could not do at all — not from doing the old things faster.

Acceleration isn't revolution. Doing new work is.

A Few Risks You Need to Know

Having talked up the upside, let me say the opposite.

At the end of the report, McKinsey lists a risk checklist. I think every item on it is worth remembering.

First, fairness. Models carry bias, because their training data carries bias. A virtual try-on app could distort how certain body types are shown.

Second, intellectual property. Where does the training data come from? Could what comes out trip over someone else's copyright? This is currently the most litigated area.

Third, privacy. Could information a user types in get "learned" into the model, then handed back out to someone else?

Explainability deserves a mention too. It gives you an answer — but why that answer? The model itself can't quite explain. Dig deeper and there's reliability: ask it the same question twice and it can give you two different answers — which one do you trust?

Security, as well. Hackers can launch a "prompt injection" — a crafted input that makes the model follow their instructions instead of yours. Then there's the environmental cost: training a single large model can emit roughly 315 tons of CO₂.

None of this is to scare you. But you have to think it through before rolling anything out.

How long a technology lasts is never about what it can do — it's about what it won't get wrong.

Back to That Number

That $2.6 to $4.4 trillion from the opening — reading it again now, it doesn't feel the same.

It's not an "oh wow, AI is amazing" number.

It's a choice: who gets to share in this trillion-dollar dividend? And who's left to catch the people it knocks down?

Toward the end of the report, McKinsey writes a line to the effect of: this report was, in its entirety, written by humans.

I laughed a little when I read that. It's almost as if they're saying — look, we're still the ones writing the reports. But next time — maybe not.

Generative AI has arrived. What's left is how we catch it.

The dividend goes to those who are ready. The shock lands on those who aren't.

That has always been the one unchanging rule of technological revolution.