McKinsey - The AI Value Curve: The Economics of Applying AI to Existing Work
Source: McKinsey & Company and QuantumBlack, The AI Value Curve: The Economics of Applying AI to Existing Work, October 2026
Key takeaways
The enterprise AI conversation is moving from adoption towards economics. Companies routinely measure users, copilots, tokens and use cases, but those metrics say little about whether AI is creating more economic value than it costs to deploy.
McKinsey estimates a $2.52 trillion annual global opportunity from applying generative AI to existing work. The estimate covers tasks where AI can currently automate or augment work at an effective cost below the way the work is performed today.
The relevant unit of analysis is the task, not the job. Jobs contain different combinations of activities, and only some may currently make economic sense to automate or augment. High AI potential therefore does not mean an occupation disappears.
The full cost of AI matters. McKinsey includes model and token costs alongside the human review, validation, correction and rework still required to produce an acceptable outcome. A technically automatable task is not necessarily an economically attractive one.
Some of the strongest AI economics sit in highly skilled, highly paid work. Legal, healthcare, management and technical roles can offer large per hour value advantages because AI can absorb routine preparation and analysis while releasing scarce professional capacity for judgement and higher value activity.
Large lower wage workforces are not automatically the best automation targets. The economics can be weaker where labour is relatively inexpensive or where AI addresses only a small portion of the work.
Financial services and technology are major value pools, but the opportunity extends well beyond them. McKinsey estimates approximately $530 billion of potential annual value in financial services, $360 billion in technology, media and telecom, $410 billion in education and $310 billion in healthcare and social assistance.
Scale can compensate for smaller per task savings. Healthcare, education and government can still generate large value pools because modest improvements applied across large service intensive workforces create meaningful economic gains.
AI economics are local. Labour costs vary substantially between countries while model costs are more globally consistent. The same AI workflow can therefore create positive value in one market and weak or negative economics in another.
AI economics are also highly sensitive to cost. McKinsey models a scenario where a 30% rise in effective AI cost reduces the US value pool by roughly $320 billion and the global opportunity from $2.52 trillion to $1.67 trillion.
Headline token prices are not the same as the effective cost of AI enabled work. Usage, model selection, context length, iteration, human review, rework and governance all affect the true cost of delivering an acceptable business outcome.
Companies need AI cost management alongside AI adoption. McKinsey argues that model routing, token governance, context management and cost per useful outcome are becoming a form of AI FinOps.
Productivity only becomes value when released capacity is deliberately used. If AI frees employees from existing work but the organisation has no plan for where their time goes next, the apparent productivity gain can disappear into lower utilisation.
The companies capturing the largest returns may not be those deploying AI most widely. Advantage comes from understanding where task economics already favour AI, prioritising the strongest business cases and redesigning work around those opportunities.
Executive summary
McKinsey and QuantumBlack’s The AI Value Curve asks a more demanding question than whether artificial intelligence can technically perform a task: does it make economic sense for AI to perform or support that task?
The distinction is increasingly important. Enterprises have spent the first phase of generative AI measuring adoption through metrics such as users, copilots deployed, tokens consumed and number of use cases. These measures help show whether technology is being used, but they do not establish whether AI is creating more economic value than the full cost of operating it.
McKinsey argues that this question becomes more pressing as AI spending scales. By 2030, the report cites an estimated $7 trillion of investment into AI infrastructure. That investment will require substantial recurring profit pools to justify it. New AI products and business models will create some of the return, but a large share will need to come from applying AI to existing enterprise work and improving productivity.
The AI value curve is designed to identify where those economics are already attractive. Rather than evaluating entire jobs or industries as automatable, McKinsey decomposes work into the tasks that make up occupations and compares the cost of performing those tasks today with the effective cost of completing them through an AI enabled workflow.
The effective AI cost goes beyond model pricing. It incorporates model and token consumption as well as the human effort still required to review, validate, correct or rework AI output. Only where an acceptable quality outcome can be produced at a lower effective cost does McKinsey count the resulting productivity opportunity as economically positive.
Using that framework, McKinsey estimates a $2.52 trillion annual global opportunity from applying generative AI to economically attractive portions of existing work.
The distribution of that value is particularly important. The report challenges the assumption that the largest AI returns necessarily come from automating the greatest volume of low wage tasks. Some of the strongest economics emerge from augmenting expensive professional expertise.
Legal professionals, healthcare practitioners, executives, managers and technical workers sit towards the higher value end of the curve. AI can absorb routine preparation, analysis, synthesis and administrative work, allowing highly paid people to redirect more of their time towards judgement, decision making and relationships.
The industry analysis leads to a similar conclusion. Financial services and technology have large opportunities, but healthcare, education and government also hold substantial potential. Their per task cost advantage may be smaller, but the scale of service intensive work means relatively modest productivity improvements can accumulate into very large value pools.
The economics also vary materially by geography. AI costs are comparatively global while labour costs are local. A workflow that produces an attractive return in the United States can therefore fall below the economic threshold in a market with lower wages.
Finally, the curve is dynamic. Model capabilities, inference prices, human review requirements and usage patterns are all changing rapidly. McKinsey demonstrates the sensitivity by modelling a 30% increase in AI costs. Under that scenario, approximately $320 billion of the US opportunity disappears and the global value pool falls by about one third.
The implication is that AI strategy needs to become more financially disciplined. Leaders need to evaluate work at task level, calculate the full cost of AI enabled delivery, prioritise applications with strong economics, manage AI operating costs and decide before deployment how released human capacity will be redirected.
The next stage of enterprise AI is therefore not simply about proving what the technology can do. It is about determining where using it creates economically sustainable value.
Key stats and quotable claims
McKinsey estimates an annual global economic opportunity of approximately $2.52 trillion from applying generative AI to economically attractive portions of existing work.
By 2030, approximately $7 trillion of capital is estimated to have been invested in AI infrastructure.
McKinsey says that investment would require roughly $760 billion in annual profit pools to justify it.
Financial services and insurance represent approximately $530 billion of the global AI value pool.
Education represents approximately $410 billion.
Technology, media and telecom represent approximately $360 billion.
Healthcare and social assistance represent approximately $310 billion.
The baseline US AI value pool is approximately $960 billion.
A 30% rise in effective AI costs reduces the modelled US opportunity by approximately $320 billion.
Under the same 30% cost increase, the global value pool falls from $2.52 trillion to $1.67 trillion, a decline of approximately 34%.
Enterprise large language model spending is cited as having tripled over a 12 month period by the end of 2025.
Legal occupations sit at the high end of McKinsey’s global occupation archetype value curve.
Healthcare practitioners and technical roles also show a high per hour AI value advantage.
Management and computer and mathematical occupations rank among the stronger economic opportunities.
Highly physical occupation groups generally show less generative AI value today because fewer working hours are addressable without robotics and labour costs can be lower.
McKinsey’s model includes human review, validation and rework in the effective cost of AI.
The analysis is built at the level of unique combinations of region, industry and 831 Standard Occupational Classification occupations.
McKinsey applies a 30% AI cost sensitivity to illustrate how changing economics can alter the addressable value pool.
Overall summary
The strongest idea in the report is that technical automation potential and economic automation potential are different things.
An AI model may be capable of doing a task, but if the model is expensive to run, needs repeated prompting or requires extensive human review, the workflow may cost more than having a person perform the task directly. Conversely, even partial automation can create significant value when it releases expensive professional capacity.
McKinsey therefore changes the unit of analysis from jobs to tasks.
A lawyer is not automatable or non automatable. The work contains research, drafting, preparation, review, client interaction, negotiation and judgement. Some activities may have strong AI economics while others continue to depend heavily on human expertise.
The same applies to executives, doctors, software developers, teachers and financial professionals.
This produces a counterintuitive result. The best economic opportunities do not necessarily sit where wages are lowest or where the largest number of routine tasks exist. Some of the strongest returns can come from removing relatively small portions of lower value activity from highly paid employees.
A chief executive spending less time synthesising briefing documents can devote more capacity to strategic decisions. A healthcare professional can spend less time on documentation and more time applying clinical judgement. A financial manager can reduce preparation and analysis work and spend more time on decisions and stakeholder interaction.
The second important idea is that scale and value per hour need to be considered together.
Financial services generates large value because both the underlying wage structure and the nature of the work are favourable to AI. Education has a smaller per hour advantage but a very large volume of addressable work. Healthcare combines a substantial workforce with expensive professional expertise.
There is therefore no single metric that tells leaders where to deploy AI.
The third idea is that AI economics are not stable.
Model costs can decline, but consumption can rise even faster. More capable models may require more compute. Data centre capacity, energy and specialised hardware can become more expensive. At the same time, organisations may get better at designing workflows that require less human correction.
The economic threshold can consequently move in either direction.
This changes how leaders should think about AI business cases. They should not assume that a use case which makes economic sense today will automatically improve forever. Nor should they assume a currently unattractive use case will remain unattractive as models improve.
The AI portfolio therefore needs to be continually repriced.
Finally, the report makes an important workforce point. Identifying where AI releases capacity is only the first half of the economic equation. The organisation also needs to decide what happens to the capacity it releases.
If an employee saves five hours each week and simply becomes less utilised, the theoretical productivity gain may never appear in financial performance.
Value capture therefore requires redeployment before deployment: identifying the higher value activities released employees will perform, what capabilities they will require and how the organisation will measure whether the released capacity actually moved into productive use.
Deep dive
From AI adoption to AI economics
The first phase of enterprise generative AI has largely been a story of capability and adoption. Companies wanted to know whether models could perform useful work, whether employees would use them and whether the technology could be integrated into existing systems.
The next phase is fundamentally different. As deployment scales, leaders need to know whether the technology creates value after accounting for everything required to turn an AI output into an acceptable business outcome. Adoption tells an organisation that people are using AI. It does not tell leadership whether the economics work.
Enterprises frequently track numbers such as active users, copilots deployed and tokens consumed. These measures can show usage and technical progress, but they cannot establish whether a use case is economically preferable to the existing way of doing the work.
This distinction becomes increasingly important as AI operating costs grow. The relevant question is no longer simply whether AI can perform an activity. Leaders need to determine whether AI can perform or support the activity at sufficient quality and at a sustainable effective cost, while allowing employees to redirect their time towards more valuable work.
What the AI value curve measures
The AI value curve applies an economic lens to generative AI automation and augmentation of existing work. It builds on earlier research into automation potential but adds labour economics and the full cost of an AI enabled workflow.
The analysis works at task level because every occupation contains a mix of different activities. Some may be well suited to AI, some may require significant human judgement and others may not be meaningfully addressable by generative AI at all.
For each economically addressable activity, the curve compares the fully loaded cost of performing the work today with the effective cost of performing it through AI.
The human cost includes wages, benefits and employer overhead. The AI side includes token and model costs as well as the human review, validation, correction and rework required to bring output up to the level at which the organisation would accept a human deliverable.
The distinction around acceptable quality matters. The benchmark is not whether human and machine outputs are identical. It is whether the AI enabled workflow meets the acceptance standard relevant to the specific task, which can vary materially between a routine document and a safety sensitive clinical workflow.
Where the combined AI and human workflow creates an acceptable outcome at a lower effective cost than the current method, the difference represents an AI enabled value advantage.
The job is not the unit of automation
The report repeatedly cautions against interpreting economically automatable work as automatable occupations.
A job is a portfolio of activities.
Some of those activities may move almost entirely to AI. Others may be accelerated while a person remains accountable for the result. Some will continue to depend heavily on human judgement, interpersonal interaction or physical execution.
The more useful workforce question is therefore not simply whether a role is exposed to AI. It is how the mix of work inside the role changes and whether the resulting human AI combination improves the economics of producing the outcome.
This also explains why apparently small changes can create substantial value. Removing a limited number of routine activities from an expensive specialist can release capacity that is disproportionately valuable.
A $2.52 trillion global opportunity
Using this task level approach, McKinsey estimates approximately $2.52 trillion in annual global economic opportunity from applying generative AI where the cost of the AI enabled workflow is below the cost of performing the work today.
The figure should not be interpreted as guaranteed realised value or as a forecast of workforce reductions. It is a modelled economic opportunity associated with the portion of work where current AI economics appear favourable.
Realisation will depend on implementation quality, adoption, organisational constraints, regulation, workflow redesign, labour arrangements and what companies actually do with the capacity released.
Nevertheless, the curve gives leaders a directional map of where the economics appear strongest today.
The distribution is as important as the total.
Large opportunities arise in two broad situations. The first is high volumes of routine or support work where modest savings accumulate across a large workforce. The second is highly paid professional work where releasing even a relatively small number of hours has substantial economic value.
What the industry curve shows
McKinsey’s industry value curve combines two dimensions.
The height of a sector on the curve represents the average cost advantage of AI over the existing human workflow after accounting for the remaining human effort required to review output.
The width represents the amount of economically automatable work available.
The total economic value is therefore driven by the combination of how much work can economically shift and how valuable each shifted hour is.
Financial services and insurance has the largest modelled opportunity at approximately $530 billion. Technology, media and telecom represents another $360 billion.
Those sectors combine relatively expensive knowledge work with substantial portions of activity that AI can support.
But the chart also challenges the assumption that AI value is concentrated only in technology intensive sectors.
Education represents approximately $410 billion and healthcare and social assistance approximately $310 billion.
Their value emerges through a different combination of economics. Per task savings may be less dramatic than in financial services, but enormous service intensive workforces mean modest improvements can accumulate into a very large value pool.
Administrative burden can decline, professional scope can expand and skilled employees can spend more time on work requiring human expertise.
Healthcare, education and government matter more than the headline narrative suggests
The sector analysis is important because public discussion around enterprise AI often focuses heavily on technology and financial services.
McKinsey’s curve suggests that this is too narrow.
Healthcare has a large workforce containing both administrative work and scarce professional expertise. If AI reduces routine preparation, documentation or analysis, clinicians can devote more capacity to work where clinical judgement matters most.
Education follows a similar pattern. A relatively moderate hourly AI advantage can still generate a large value pool because the workforce is large and there is a meaningful volume of activity around preparation, administration and knowledge work.
Government citizen services also show meaningful potential even though the per hour advantage is not among the highest.
The implication is that leaders should not rank industries purely by how technically advanced they appear. Workforce scale and the economic value of released professional capacity can create attractive opportunities in sectors that receive less attention in the AI investment narrative.
The occupation curve changes the automation story
McKinsey then looks at the same economics by occupation archetype.
The results challenge a common assumption that AI creates the greatest value by replacing relatively low wage administrative work.
Legal occupations sit towards the top of the global curve, followed by healthcare practitioners and technical professionals, management and computer and mathematical occupations.
Business and financial operations also represent a significant opportunity because a meaningful portion of the activities inside these jobs can be supported by AI.
By contrast, administrative support, protective services and personal care currently show a smaller per hour advantage. Many physical occupations offer still less value from generative AI because the technology cannot perform the physical portion of the job without robotics.
The strongest current generative AI economics therefore concentrate in cognitive work where labour is expensive and where preparation, synthesis, drafting or analysis can be accelerated.
Expensive expertise can create better economics than cheap labour
One of the clearest conclusions from the occupation analysis is that AI can generate equal or greater value by augmenting highly paid professionals than by automating larger quantities of inexpensive labour.
The reason is straightforward.
An hour released from a highly paid specialist is economically worth more than an hour released from a lower wage employee, provided the technology can perform the relevant task at acceptable quality and reasonable cost.
Legal work illustrates the principle particularly clearly. Research, first drafts, synthesis and document preparation can consume significant amounts of professional time even though the highest value contribution of a lawyer is often judgement, negotiation and advice.
The same pattern can apply to management, healthcare and technical roles.
AI does not need to replace the professional to generate substantial value. It needs to remove enough lower value activity that the professional can spend materially more time applying the expertise for which the organisation is paying.
The executive example
The report makes the point particularly clearly through senior executives.
Chief executives sit near the high value end of the occupation type curve.
That does not imply that CEOs are likely to be automated. It means that portions of executive work can be accelerated economically.
Meeting preparation, managing information, synthesising briefing documents, reviewing board material and scanning competitive developments can consume large amounts of senior management time.
AI can potentially reduce that workload.
The actual economic benefit then depends on what happens to the released time. If leaders use it for higher quality strategic decisions, complex judgement, stakeholder engagement or relationship building, AI can increase the effective return on expensive executive capacity.
This illustrates a broader principle running through the report: the economic value of augmentation lies partly in what humans do with the capacity AI returns to them.
Physical work follows different economics
The curve looks materially different for jobs dominated by physical execution.
Waiters, assemblers, quality inspectors, automotive mechanics and HVAC technicians sit towards the lower end of the current generative AI curve.
There are two reasons.
First, a smaller portion of the work can be performed through generative AI without corresponding advances in robotics or other physical automation.
Second, lower labour costs narrow the economic gap between the existing workflow and an AI enabled alternative.
This does not mean those jobs are permanently outside AI transformation.
AI capabilities and robotics will continue to evolve, and falling effective costs can move more activities above the economic threshold.
The curve should therefore be read as a picture of current economics rather than a permanent boundary around what technology can eventually address.
AI economics are local
McKinsey’s geographical analysis introduces another important point.
The cost of AI can be broadly similar between markets because organisations may use the same underlying models and infrastructure.
Labour costs are not.
The value differential between human labour and an AI enabled workflow therefore changes considerably by country.
This is particularly visible in the comparison between US and global averages. Higher US wages create a much wider potential cost differential across several occupations and industries.
A use case that comfortably generates value under US assumptions may become marginal in a country where the equivalent worker is significantly cheaper.
Software development illustrates the effect. It remains attractive globally, but the per hour value difference is dramatically larger in the United States because the underlying labour is more expensive.
Sales and administrative support can similarly sit above the economic threshold in the United States while offering weaker or potentially negative economics elsewhere.
The implication is important for global organisations.
A global AI strategy cannot assume that the same business case applies uniformly across every country.
Local economics do not determine workforce strategy by themselves
The report is careful not to turn the geographic curves into a prescription for workforce reductions.
A positive AI cost differential does not mean demand for the occupation necessarily falls.
A profession can have substantial AI augmentation potential while demand for the underlying human expertise continues to rise.
The curves also do not incorporate every practical constraint organisations face.
Employment law, collective bargaining, regulation, change management, implementation cost and workforce readiness can materially affect whether an economic opportunity can actually be captured.
The geographic analysis is therefore best used to evaluate use cases locally rather than to determine a broad workforce strategy solely from wage differences.
AI cost can move the curve very quickly
The value curve is not static.
Both AI capability and effective cost are changing rapidly.
A task that is marginal today may become attractive because a new model performs it more reliably, fewer iterations are required or human review falls.
At the same time, falling model prices do not guarantee falling enterprise AI expenditure.
Usage can grow faster than unit costs decline. New use cases increase consumption. More capable models may require greater compute. Data centre power and specialised hardware can also become more expensive.
The cost that matters is therefore not the advertised price of a token.
It is the total cost of delivering an acceptable useful outcome through the AI enabled workflow.
The human component of AI cost can change as well
Human review represents a particularly important part of the economics.
Early AI deployments can require substantial checking, correction and escalation.
As organisations gain experience and models become more reliable, that burden may decline, improving the business case.
But the reverse is also possible.
Complex or high consequence uses may require more governance and validation than initially expected. Organisations may discover hidden rework or risk costs only after pilots reach production.
Leaders therefore need to understand not only how model prices move but how the human workload around the model changes.
The goal is not to eliminate human judgement indiscriminately.
It is to place human effort where it contributes enough value to justify its cost.
A 30% cost change materially alters the opportunity
McKinsey stress tests the value curve using a 30% increase in AI cost.
The scenario is illustrative rather than predictive, but the effect is large.
The US baseline value pool of approximately $960 billion loses around $320 billion.
Globally, the modelled opportunity falls from $2.52 trillion to approximately $1.67 trillion, a 34% decline.
Cost advantages narrow across industries. Strong use cases become less attractive and some marginal applications move into negative territory where using AI costs more than the current workflow.
The analysis demonstrates why AI investment cases cannot simply assume that costs will continually decline.
Hedging against AI cost reversal
McKinsey recommends designing AI operating models with enough flexibility to withstand changing economics.
Human AI workflows should allow work to shift between people and technology when the underlying economics change. Organisations can also negotiate pricing protection with providers where possible and avoid allowing critical expertise to disappear from roles sitting near the economic threshold.
Processes themselves should remain flexible enough to maintain service levels if AI usage becomes temporarily more expensive or needs to be reduced.
The broader principle is optionality.
An enterprise should not redesign work so rigidly around one AI cost assumption that it becomes difficult to respond when technology economics shift.
Transformation costs make the threshold harder still
The $2.52 trillion value pool does not include every investment required to capture the opportunity.
McKinsey’s curve focuses on the ongoing cost of performing tasks through AI, including tokens and remaining human work.
Enterprises still need to fund enablement.
Compute and cloud infrastructure, platforms, integration, training, data preparation and change management can all require material one off or transitional expenditure.
Including those investments would reduce the net near term value available.
This reinforces the importance of prioritisation.
A positive theoretical task differential does not automatically justify a transformation programme if the value pool is too small to recover the cost of implementation.
What leaders can do: map work at task level
McKinsey’s first recommendation is to map where economic opportunity exists inside the business rather than beginning with technology.
Leaders need to understand which activities consume employee time, which can be addressed digitally, which depend heavily on judgement and where AI changes cost or quality enough to matter.
The report suggests starting with a task level audit of major cost centres.
This is a materially different starting point from asking departments to generate lists of AI ideas.
The objective is to identify where the underlying economics are strongest before deciding which use cases deserve investment.
Prioritise value before scaling
The second recommendation is to size and prioritise the strongest opportunities before moving to scale.
A technically impressive demonstration can still represent a very small business opportunity.
Conversely, an unglamorous workflow can be economically significant if it occurs at large volume and has a strong cost differential.
McKinsey therefore argues for ranking AI investments according to the size of the economic opportunity rather than the novelty of the pilot.
This moves the portfolio conversation from technological possibility towards value creation.
Manage AI cost like an operating input
The third recommendation is to manage AI costs alongside adoption.
As expenditure becomes material, companies need to treat AI more like another factor of production.
Model routing, context length, token usage, governance and cost per useful output increasingly need active financial management.
Different tasks may not need the same model.
An organisation can potentially route routine work towards cheaper systems while reserving more expensive frontier models for activities where the incremental capability creates enough value.
This discipline is what McKinsey describes as a growing AI FinOps requirement.
Stress test every business case
The fourth recommendation follows directly from the cost analysis.
Business cases should incorporate both upside and downside scenarios.
What happens if effective AI costs rise by 30%? What happens if they fall by 30%? How sensitive is the project to human review requirements, model selection or usage volumes?
Understanding that sensitivity before scaling helps leadership distinguish robust use cases from those whose return depends on a narrow set of favourable assumptions.
AI economics should therefore be treated as dynamic rather than fixed at the point of approval.
Plan redeployment before deployment
The fifth recommendation may be the most important for translating productivity into enterprise value.
If AI releases a portion of an employee’s workload, leaders need an explicit view of what happens to that capacity.
Otherwise, apparent productivity gains can turn into lower utilisation or unmeasured slack.
The organisation should identify which higher value activities people will take on, establish explicit redeployment expectations and understand what reskilling or capability development is required.
The workforce question should therefore be addressed before deployment rather than after productivity is released.
An AI initiative is not economically complete merely because it saves time.
The organisation still needs to turn that time into something valuable.
Methodology
McKinsey’s AI value curve uses a bottom up model built at the level of individual combinations of region, industry and occupation.
The occupational analysis covers 831 Standard Occupational Classification roles.
For each combination, McKinsey compares the fully loaded hourly cost of labour, including wages, benefits and employer overhead, with the effective hourly cost of using generative AI.
AI costs incorporate token consumption alongside non token costs such as human review, validation, rework and risk related activity.
McKinsey then combines the cost differential with its estimate of the share of working hours made automatable specifically by generative AI.
Employment and productive working hours are used to calculate the economically addressable volume of work.
The resulting net value pool represents the modelled economic opportunity or loss associated with applying AI to that work.
The analysis excludes traditional automation technologies such as robotic process automation and interactive voice response and therefore focuses specifically on the additional economics associated with generative AI.
The model also holds some variables constant across countries while adjusting labour costs by region. The results should consequently be treated as a directional economic map rather than a company specific forecast.
Conclusion
McKinsey’s AI Value Curve represents an important shift in the enterprise AI discussion.
The first wave asked what AI could do.
The next wave asks where doing it actually makes economic sense.
The report makes clear that the two questions are not interchangeable.
Technical capability does not guarantee economic value. The effective cost of AI includes the model, its usage and the human work still needed to make the output trustworthy and usable.
This is why the task becomes the most useful unit of analysis.
Jobs are collections of activities with different levels of technical and economic exposure. AI can therefore create substantial value inside an occupation without replacing the occupation itself.
The report also challenges a common automation assumption.
The strongest economics do not necessarily come from removing the largest quantities of inexpensive human work.
In many cases, the greater opportunity is to release expensive professional capacity.
Legal, healthcare, management and technical roles sit prominently on the value curve because routine preparation and analysis consume time that could otherwise be used for judgement, decision making and expertise.
The industry analysis reinforces the same point.
Financial services and technology contain large opportunities, but healthcare, education and government also generate significant modelled value because relatively modest productivity improvements compound across large workforces.
The economics are neither global nor fixed.
Labour costs make the same AI workflow more attractive in some markets than others. Changing model prices, compute costs and human review requirements can move tasks above or below the economic threshold over time.
That creates a more demanding management discipline.
AI investments need to be continually repriced and stress tested rather than approved once and assumed to become more attractive automatically.
And the final economic step sits outside the model itself.
If AI releases human capacity, the organisation needs to deliberately move that capacity into higher value work.
Without redeployment, time saved is not necessarily value captured.
The organisations that generate the strongest returns may therefore not be those deploying AI across the largest number of employees or use cases.
They will be the organisations that understand the economics of the work itself, invest where the evidence is strongest, manage the full cost of AI and deliberately redesign how human capacity is used.
Key takeaways
Enterprise AI is moving from an adoption problem towards an economic allocation problem. Usage metrics can show whether AI is spreading, but they do not establish whether the resulting workflows create sustainable value.
The task rather than the occupation is the correct level for evaluating AI economics. Individual jobs combine work that can be automated, augmented or retained as predominantly human.
The full cost of AI includes more than tokens. Human review, validation, correction, governance and rework can materially change whether a use case remains economically attractive.
McKinsey estimates approximately $2.52 trillion of annual global opportunity from applying generative AI to existing work where current economics are favourable.
High wage professional work can create some of the strongest returns because AI releases scarce expertise from lower value preparation and analytical activity.
Financial services and technology contain major value pools, but healthcare and education demonstrate why workforce scale can make moderate productivity gains economically significant.
The same use case can have very different economics between countries because labour costs remain local while AI model economics are comparatively global.
AI costs also need to be continually stress tested. A 30% increase in McKinsey’s model reduces the global opportunity by roughly one third.
Leaders therefore need to manage AI spend with the same discipline applied to other operating inputs, including model routing, usage, human review and cost per useful outcome.
Most importantly, organisations need to plan redeployment before deployment. Productivity only becomes enterprise value when the human capacity released by AI is deliberately redirected towards higher value work.
The simplest way to capture McKinsey’s argument is: do not ask only where AI can do the work. Ask where AI can do the work economically, and what the organisation will do with the human capacity it releases.
