The Economic Potential of Generative AI: Where the Trillions Actually Concentrate

Generative AI arrived as a horizontal technology, a capability that, in principle, touches every desk, every document, every line of code in the enterprise. That universality is exactly what makes it so easy to misread. When a tool appears useful everywhere, the instinct is to deploy it everywhere: a pilot in legal, a pilot in HR, a pilot in finance, a copilot license for everyone, a hundred flowers blooming. Two years into the enterprise wave, the results of that instinct are in, and they are sobering. The flowers bloom; the value does not.

The numbers that should anchor every executive conversation about generative AI come from McKinsey's June 2023 analysis, The economic potential of generative AI. Across the 63 use cases it examined, McKinsey estimated that generative AI could add the equivalent of $2.6 trillion to $4.4 trillion annually in global corporate value. For scale, the firm noted that the entire 2021 GDP of the United Kingdom was roughly $3.1 trillion. The technology's potential, in other words, is on the order of a large national economy, recurring every year.

But the figure that matters more than the headline is the distribution behind it. McKinsey found that about 75 percent of that value falls in just four business functions: customer operations, marketing and sales, software engineering, and research and development. Sixteen functions were studied. Four of them carry three-quarters of the prize. The rest (finance, HR, legal, procurement, strategy, supply chain) share what is left.

⤢ Expand

The macro shift: from a feature to a value distribution

The reason the concentration matters is that it inverts the usual logic of enterprise software adoption. Most platforms create value in rough proportion to seats deployed: more users, more value, more or less linearly. Generative AI does not behave that way. Its value is a function of two things at once: how much of a function's work is language-shaped and pattern-shaped (and therefore automatable or augmentable), and how directly that work converts into revenue or cost.

Customer operations score high on both. McKinsey estimated generative AI could lift productivity in customer care at a value of 30 to 45 percent of current function costs, because a contact center is, in essence, a machine for producing and consuming language at scale, and every minute saved is a measurable cost removed. Software engineering scores similarly: McKinsey put the direct productivity impact at 20 to 45 percent of current annual spend on the function, driven by faster initial code drafts, refactoring, and root-cause analysis. Marketing and sales convert creative and analytical work directly into pipeline; the firm estimated the marketing productivity uplift alone at roughly $463 billion annually, equivalent to 5 to 15 percent of total marketing spend. R&D rounds out the four, where generative design and simulation compress discovery cycles.

What these four share is not that they are "knowledge work"; almost every function is knowledge work. It is that they combine high language-density with a short, legible path from output to money. That combination is rare. It is precisely why the value clusters rather than spreads.

$2.6–4.4T

Annual value potential across 63 use cases (McKinsey, 2023)

~75%

Of that value in four functions (McKinsey, 2023)

30–45%

Customer-care productivity value vs. function cost (McKinsey, 2023)

~2/3

CEOs ranking AI a top-3 priority (BCG AI Radar, 2025)

The forces: why "everywhere" feels right and pays badly

If the data is this clear, why do so many organizations still spread their generative-AI effort thin? Three forces conspire.

The first is demand symmetry. Because the technology is genuinely useful in every function, every function lobbies for it with equal conviction. The legal team's request for contract summarization looks, on a slide, just as compelling as the contact center's request for agent assist. But "useful" and "valuable at scale" are different claims. A tool can save a lawyer ten minutes a day and still move no needle the CFO can see; the same tool, applied to a function where the work is the cost base, can move points of margin. Symmetric demand produces symmetric funding, and asymmetric returns.

The second is pilot economics. Pilots are cheap, visible, and politically rewarding. A portfolio of twenty pilots signals momentum to the board. But pilots do not concentrate; they fragment. Each one consumes scarce integration, data-engineering, and change-management capacity (the genuinely constrained resources) and spreads that capacity across functions where most of it will never compound into production value. The organization ends up rich in proofs of concept and poor in deployed capability.

The third is measurement drift. When value is hard to attribute, activity becomes the proxy. Seats provisioned, prompts run, meetings "AI-assisted": these get counted because the real number, dollars of margin or revenue moved, is harder to isolate. Counting activity quietly rewards breadth over depth, because breadth is easier to show. The concentration thesis is, at bottom, an argument for measuring the thing that is hard to measure.

The mistake is not under-investing in generative AI. It is investing everywhere at once, and capturing the prize nowhere.

The decisions: chase the concentration

Treating the value distribution as a strategy document, rather than an interesting chart, changes four decisions.

Where to place capacity. Your binding constraint is not licenses. It is the small number of people who can wire a model into a real workflow, govern its outputs, and manage the change. That capacity should flow, disproportionately, to the functions in the upper-right of the value map: high value intensity and high implementation feasibility. A single deeply instrumented deployment in customer operations or software engineering will, on McKinsey's own ranges, out-return a dozen shallow pilots elsewhere. Concentration is not caution; it is how the math works.

How to sequence. The four high-value functions are not equally easy to enter. Software engineering and customer operations tend to offer the fastest path because the workflows are digital, the outputs are checkable, and the baselines are already measured. Marketing and sales follow. R&D is the highest ceiling and the longest build. Sequencing for early, attributable wins funds the harder, higher-ceiling work later, and, just as important, builds the organizational muscle to govern AI in production before you bet the discovery pipeline on it.

What to measure. Replace activity metrics with function-native value metrics from day one: cost-per-contact and resolution time in customer operations; cycle time and change-failure rate in engineering; pipeline and content-throughput in marketing. If a deployment cannot be tied to one of these, it is a pilot, not a program, and it should be labeled as such so it competes honestly for capacity.

Where the data lives. Concentrating value in customer operations, engineering, and R&D means concentrating sensitive data (customer records, proprietary code, research IP) in the path of a model. The functions with the highest value are, not coincidentally, the ones with the most to protect. That argues for running the highest-value workloads on infrastructure you control: a private machine, walled off, EU-resident where regulation demands it, where your data stays where you can see it. The concentration thesis and the sovereignty thesis point the same direction: the work worth doing is the work worth protecting.

Why this is urgent now, not later

The ambition gap has closed. In BCG's AI Radar 2025, two-thirds of CEOs reported ranking AI among their top three strategic priorities, and 82 percent said they were more optimistic about it than a year earlier. The scarce input is no longer executive will or capital appetite. It is focus, the discipline to point a now-funded, now-mandated effort at the four functions that hold the value, rather than letting it diffuse across sixteen because every function asked nicely.

That discipline is what separates the organizations now reporting measurable returns from those still cataloguing pilots. The technology did not change between the two groups. The allocation did.

Where to start

Start by drawing your own version of the map. For each of the four high-value functions, name one workflow where the work is the cost base, the output is checkable, and the baseline is already measured. Pick the single highest-intensity, highest-feasibility candidate. Instrument the baseline before you deploy, so the value is provable rather than asserted. Then concentrate your scarce integration and governance capacity there until it compounds into production, and only then move to the next.

The reflex to do everything is understandable; the technology really is useful everywhere. But usefulness is not the same as value, and value, on the best available evidence, is not spread evenly. It is concentrated. The strategic move is to concentrate with it.