Generative AI’s $4.4 trillion is a scenario, not a cheque
McKinsey’s figure measured potential under assumptions, not a forecast. Breaking it into tasks, adoption, captured value and costs makes it useful.
McKinsey Global Institute estimated on June 14, 2023 that 63 generative AI use cases could deliver $2.6 trillion to $4.4 trillion in value each year. Its 68-page report compared the upper bound with the United Kingdom’s 2021 GDP. The scale was striking, but it was not a forecast of money that would automatically appear in the economy.
It was an estimate of potential: the annual value selected uses might generate if broadly implemented under the model’s assumptions. Between technical capability and an economic outcome stand adoption, integration, quality, investment, competition and distribution. Locating those layers makes any macroeconomic AI figure easier to read without confusing possibility with result.
What the range actually counts
McKinsey examined 63 use cases across 16 business functions. It estimated the value they might generate through measurable outcomes and projected those effects onto the structure of the 2022 global economy. The analysis excluded entirely new product or service categories the technology might create.
About 75% of estimated value sat in four functions: customer operations, marketing and sales, software engineering, and research and development. That did not mean three quarters of jobs in those areas would disappear. It meant the largest monetary share of the analyzed cases came from outcomes in those functions.
The $2.6 trillion to $4.4 trillion range covered those use cases. The report said the estimate would roughly double after including generative AI embedded in software used for other activities. In a different calculation, after removing overlaps, it combined use cases and labor productivity to reach $6.1 trillion to $7.9 trillion in potential. Mixing these figures as though they measure one thing creates incompatible headlines.
Potential, adoption and captured value
A potential figure normally sits at the end of a chain. First, the system must be capable of performing or assisting a task. Second, an organization must adopt it. Third, integration must not move the saving into errors, review or risk. Fourth, the benefit has to become additional production, lower cost, better quality or more revenue.
Every step reduces or delays the previous one. A tool may save ten minutes drafting and require fifteen minutes of checking. It may raise every competitor’s capacity, shifting the saving into lower prices rather than profit. It may release hours that an organization cannot productively redeploy. Social value, corporate income and worker pay are different quantities.
An operational reading therefore replaces “AI will create X” with an equation: task volume multiplied by adoption, net improvement per task and the share of value captured, minus implementation costs and harms. None of those terms should default to one. The macro range results from assumptions about them.
An automatable activity is not an eliminated job
The report estimated that existing technologies, including generative AI, had the technical potential to automate activities absorbing 60% to 70% of employees’ time. The unit was the activity, not the person. A job combines automatable tasks, assistible work, decisions, responsibility and relationships that can change in different directions.
The March 2023 preprint GPTs are GPTs made a related distinction explicit. It estimated that about 80% of the US workforce could have at least 10% of tasks exposed, while 19% could have at least half exposed. Exposure meant cutting task time by 50% at constant quality; the authors said it neither implied complete automation nor predicted an adoption timetable.
The statement “70% of work is automatable” erases four qualifiers: share of time, technical capability, a combination of technologies and a scenario. It also invites a one-to-one conversion from tasks to positions. The useful questions are which task bundle remains, who validates the output and how the occupation is redesigned.
Task evidence does not scale by itself
Experiments had already measured productivity in bounded settings. The April 2023 Generative AI at Work working paper studied a gradual assistant rollout among thousands of customer-support agents. It reported an average gain of roughly 14% in issues resolved per hour, with different effects by experience and skill.
That design offers something a demonstration cannot: it compares work outcomes around a real introduction of the tool. It still describes one company, workflow and metric. It does not prove every contact center will gain 14%, much less that the percentage applies to lawyers, doctors or programmers.
A separate preregistered experiment by Noy and Zhang, available from March 2023, assigned professional writing tasks to 444 participants. Access to ChatGPT reduced completion time and increased average quality under that protocol. It also shifted work toward idea generation and editing and away from rough drafting.
Both studies found gains, but in delimited tasks with their own measures. Moving from a trial to an economy requires weights: how many tasks are similar, how many organizations adopt, which costs emerge and how long the effect lasts. Applying an experimental percentage to all payroll is not measurement; it is another assumption.
How to reconstruct a large number
First identify the unit. Does the figure mean revenue, cost reduction, profit, productivity or welfare? McKinsey described potential economic value combining use-case outcomes. It did not claim $4.4 trillion in accounting profit or new spending on AI vendors.
Second, find the denominator and horizon. The figure was annual and global, based on the 2022 economy. Comparing it with a national GDP helped visualize scale, but did not make the measures equivalent. An estimated value flow can pass through firms, consumers and workers without appearing as a separate additional economy.
Third, separate scenario from forecast. A scenario asks what follows if conditions hold; a forecast assigns a likely path and date. McKinsey also modeled adoption and placed a midpoint for automating half of current activities around 2045, within a 2030-to-2060 range. That breadth displayed uncertainty, not a contractual schedule.
Fourth, inspect exclusions and overlaps. A case may save labor and raise revenue, but summing both without adjustment could double count value. A tool embedded in software may appear in the activities lens and the use-case lens. The report said it netted overlaps in the combined estimate; a summary should preserve that caution.
From a report to a company budget
An organization cannot decide with global trillions. It starts with one task and a baseline: volume, time, error rate, cost, quality and business outcome. It runs assisted and unassisted trials on the same cases, with blind reviewers where possible. Failures are recorded alongside averages.
Then it calculates net value per accepted output. Subtract licensing, compute, integration, security, review, training and maintenance. Include expected error costs and transition time. If the system produces faster while every item needs checking, the bottleneck may move without disappearing.
Finally, inspect distribution. Who saves time, who receives more control work and who owns the error? Does the saving improve quality, reduce prices, raise margins or remove hours? An aggregate figure cannot answer those questions, although each determines the effect on workers and customers.
The $4.4 trillion figure made generative AI’s possible scale visible in 2023. Its best use is not to justify every purchase, but to teach readers to decompose an estimate into task, evidence, adoption, capture and cost. Once those five pieces are written down, the number stops being an oracle and becomes a testable hypothesis.
This article was produced with artificial intelligence under human editorial oversight.
