AI Impact on Revenue, Profits Strongest in IT

A new study by Torsten Slok, partner and chief economist at Apollo Global Management, finds evidence of artificial intelligence financial benefits in the information technology industry, but not elsewhere. “Revenue per worker is rising for the Mag 7 but is flat for the S&P 493,” he says. “Profit margins are rising for the Mag 7 but are flat for the S&P 493” as well. 

source: Apollo

 

source: Apollo

That might fuel concern that AI overinvestment is occurring. On the other hand, it seems also the case that AI investment is highly concentrated in relatively few IT firms. In other words, AI actually is not being deployed at scale outside IT. 

source: Apollo

So observers can take their pick of explanations: either AI is not generally producing measurable results or AI is not yet producing results because few firms outside IT have implemented at scale. 

At a high level, it also is likely that standard statistics miss some meaningful AI-related benefits. In other words, value is created, but not in ways that show up immediately, in every firm or industry, in the form of gross domestic product and productivity growth.

Type of benefit or measurement problem	Example involving AI	What conventional statistics may miss
Free or inexpensive services	A person uses a free chatbot instead of paying for tutoring, research, or basic professional advice.	Consumer value beyond the price paid.
Quality improvement	AI helps a physician explain a diagnosis more clearly or helps a lawyer produce a better analysis.	Better output when the number of services and their recorded prices barely change.
Time savings	A worker completes research in 20 minutes instead of two hours.	The value of saved time if it doesn't translate into recorded production, earnings, or leisure-related output.
New capabilities	A small business gains access to analytical capabilities previously affordable only to large firms.	New economic opportunities that have not yet generated recorded sales.
Intangible investment	A company spends months reorganizing workflows, developing proprietary data, and training employees.	Some of the investment and the resulting organizational capital.
Innovation spillovers	AI accelerates product development, scientific research, or software creation.	Future benefits that have not yet become market transactions or measurable output.
Unpaid or household activity	A consumer uses AI to plan a trip, learn a language, or manage household tasks.	Some of the resulting value because much household production lies outside GDP.

Measurement is itself a problem. The OECD, for example,  identifies intangible investment, cloud-computing accounting, and improvements in the quality and scope of services as important measurement challenges.

Also, value can be produced in economic welfare without showing up in GDP statistics. Such has been the case for many internet-supplied “no extra cost” services used by consumers, for example.

Economists call the difference between what a consumer would be willing to pay and what the consumer actually pays consumer surplus. So a2025 paper by Erik Brynjolfsson and coauthors, GDP-B: Accounting for the Value of New and Free Goods, develops an alternative approach to capturing benefits from free and new goods that standard GDP measures poorly. 

A 2026 Stanford Digital Economy Lab study, What Is Generative AI Worth?, used surveys and incentive-compatible choice experiments to estimate the value U.S. consumers place on access to generative AI tools. Its estimated aggregate consumer surplus rose from $116 billion in 2025 to $172 billion in 2026.

That is unlikely to satisfy skeptics insistent on visible AI impact on firm results. The International Labour Organization's May 2026 research brief, The Aggregation Paradox of AI , finds that strong task-level gains have not yet translated consistently into measurable gains at the firm, sector, or macroeconomic levels. Uneven adoption, workflow constraints, and the cost of organizational change help explain the gap. 

AI might produce large benefits for consumers and workers while capturing relatively little revenue for the companies providing it, as the Apollo analysis suggests is presently the case. 

Or, productivity gains might accrue primarily to a small number of firms, with limited diffusion to the rest of the economy, as the Apollo study also suggests. 

The absence of an immediate productivity boom is not strong evidence that AI is failing, or will fail. Past historical evidence suggests an important new technology takes some time to show benefits. 

But the existence of a historical productivity lag is not strong evidence that a productivity boom is inevitable, either.

Still, a delay in measured economic or firm benefits is not at all unusual. Past innovations such as steam power, electricity,  information technology and the internet have taken some time to generate clear economic value, because:

• Complementary investments in organizational processes, software, training, and new business models must be made

• Benefits may initially be concentrated in particular tasks, firms, and industries

• Some gains, particularly in consumer welfare and intangible assets, are imperfectly captured by conventional statistics.

Financial observers will still remain skeptical. But, as a historical matter, important new general-purpose technologies have almost-always, perhaps always, experienced a lag time between introduction and impact. AI is showing the same pattern. 

AI Value Will Not be Captured by Firm or GDP Statistics

For all the legitimate worries expressed about firm-level artificial intelligence revenue magnitude, compared to capital investments, firm-level accounting and business model issues never capture the full value of the innovations, any more than such statistics can capture the full economic surplus created by earlier innovations such as the personal computer and the internet. 

It might in fact be the case that general-purpose technologies such as electricity create industry revenue that does not capture the value of electricity to society.

Or consider computing. The economic value of computation has exploded while the price of a unit of computation has collapsed. If computing enables a service used to cost $100 to fall to $1, measured consumer spending falls from $100 to $1. 

So our accounting metrics might suggest a falling output, even though consumers are $99 better off. 

That migration is one reason technological revolutions can produce enormous consumer welfare without creating a comparably enormous technology industry's share of GDP.

Consider Google Search. Users do not pay directly, but Google captures advertising revenue. But economic value is far greater. 

Researchers might spend only five minutes finding information that would previously have required two hours. AI chatbots vastly create more value than that, functionally allowing people to tap the expertise of many other experts inside and outside of a given domain. 

The point is that “surplus” or “value” are far greater than reported firm revenues and profits, across an entire economy. And that is an issue with nearly all digital goods, especially those available to users "for free.” 

if AI eventually  makes any product essentially free, conventional gross domestic product  could actually show less spending, even while society is receiving vastly more value from it.

That earlier analogies are search, Wikipedia, online maps, free email, open-source software and other digital goods people can use without additional charge. 

We are familiar with the "productivity paradox," where measurable output gains lag initial investment in information technology by years.

The issue for AI is not necessarily “productivity” metrics but impact on economic output. In a sense, productivity metrics are about inputs, while economic welfare or value is about outcomes. It will be easier to measure the former than the latter, if neither will be straightforward or easy. 

Economic value and market prices might not correspond very well. But that also is “typical” of many computing innovations. During the early phases, producers may capture a large fraction of the generated surplus, and we can measure that in revenue and profits. 

Later, an increasing fraction of value arguably shifts toward customers in the forms of lower prices, higher output, better quality or new products. 

That is why high social value and low producer margins can coexist. Or, to put it another way, an extraordinarily successful technology can become a much-less extraordinary business. 

As in the case of tax burdens, incidence (who ultimately pays)matters. In the case of the AI value chain, value will not be entirely captured by producers. 

Some of the value will be gained by all firms and entities able to use AI. Consumers might see lower prices and the ability to consume more. Some workers might see higher wages. New businesses, products or industries could emerge. 

The analogy might be the way the Internet changed and then created huge new types of products and activities. Search, instant messaging, video and audio streaming, online news sources, social media, location-based services and mapping as well as all online forms of media provide examples. 

Much of that output was previously either expensive or did not exist. And in many cases, usage is at “no additional charge” or at lower costs than prior products required. 

People also should gain more leisure time. All of that surplus will be created, but not captured at the firm level by producers. 

Technological abundance tends to do that. 

Why So Few Firms Can Point to AI-Driven Productivity Gains

Relatively few firms so far have been able to quantify artificial intelligence productivity gains. But that has been the case for computing in general and the internet: it takes time for innovations to transform business processes. In other words, associated and complementary intangible capital also must be created. 

Studying the productivity impact of computerization on 527 large U.S. firms over 1987-1994, professors Erik Brynjolfsson, MIT Sloan School of Management and Lorin Hitt, University of Pennsylvania, Wharton School found that computerization measured over a five-year to seven-year period found productivity and output contributions up to five times greater than over a one-year period. 

“The results suggest that the observed contribution of computerization is accompanied by relatively large and time-consuming investments in complementary inputs, such as organizational capital,” the researchers say.

We could note the same trend with regards to the internet: productivity did not improve quickly, as whole business processes had to be revised. 

Study	Period / data	What it found	Relevance to the paradox
Brynjolfsson and Hitt, 1996, "Paradox Lost?"	Firm-level IT spending	Found substantial returns to information systems investment at the firm level despite weak aggregate evidence.	Early evidence that the "paradox" could be a measurement or aggregation problem. (PubsOnline)
Brynjolfsson and Hitt, 2000, "Beyond Computation"	Firm-level/case evidence	IT's value depended heavily on organizational transformation and intangible investments.	Probably the most important conceptual explanation for why technology's benefits arrive with a lag. (American Economic Association)
Brynjolfsson and Hitt, 2003, "Computing Productivity"	~600 U.S. firms, 1987–94	Returns to computers were 2–5 times greater over seven years than one year.	Strong evidence that complementary investments take years to generate their full payoff. (ResearchGate)
Oliner and Sichel, 2000	U.S. economy	IT accounted for roughly two-thirds of the acceleration in productivity growth between the first and second halves of the 1990s.	By the late 1990s, the productivity payoff of IT had become visible at the macro level. (Federal Reserve)
Stiroh, 2002	61 U.S. industries	Productivity acceleration was greatest in IT-producing and IT-intensive industries.	Shows diffusion beyond the technology-producing sector. (Federal Reserve Bank of New York)
Barua et al., 2004, "Net-Enabled Business Value"	>1,000 firms	Internet-enabled capabilities improved operational performance and ultimately financial performance; supplier/customer readiness mattered greatly.	Direct evidence that Internet adoption plus complementary organizational capabilities generated business value. (AIS eLibrary)
López Sánchez et al., 2006	464 Spanish firms	Both IT investment and workplace Internet use were associated with higher productivity.	Direct firm-level evidence of an Internet-productivity relationship. (ScienceDirect)
Bloom, Sadun and Van Reenen, 2007/2012	U.S. and European multinationals	U.S. firms obtained substantially greater productivity from IT, largely because of superior management practices.	Powerful evidence that management complements technology. (National Bureau of Economic Research)
Quirós Romero and Rodríguez Rodríguez, 2010	2,168 Spanish manufacturing firms, 2000–05	E-buying significantly improved firm efficiency.	Shows that specific Internet-enabled processes, rather than "Internet adoption" generally, mattered. (ScienceDirect)
Huang and Liu, Taiwan e-commerce study, 2013	Taiwanese manufacturing firms, 1999–2002	E-commerce and R&D both raised productivity; their combination was complementary, with network effects.	Internet value increased when combined with other forms of innovation. (ScienceDirect)
Najarzadeh, Rahimzadeh and Reed, 2014	108 countries, 1995–2010	Internet use had a statistically significant positive relationship with labor productivity.	Evidence that the Internet's productivity effects eventually appeared at the macro level. (ScienceDirect)

Beyond all that, some productivity enhancements are difficult to measure, especially when the capabilities do not have a price tag, and are usable without extra charge, such as search, email, navigation, maps or  social media. 

How do we capture the value of increases in consumer choice, reduced transaction costs, reduced search costs, easier price comparison, better product matching or enhanced convenience?

A corollary might be that the Internet and other general-purpose technologies such as electricity become less visible precisely as their economic importance increases. In other words, the technology becomes embedded in all products and services and becomes less visible as a result. 

So organizational performance enhancements are increasingly difficult to isolate from overall organizational prowess. 

In the case of AI, organizations might already be getting substantial value from AI through:

• employees completing tasks faster

• better-quality work

• broader scope of work

• fewer errors

• faster customer responses

• employees handling more work without additional hiring

• better and faster software development

• faster research and analysis

• improved sales and marketing.

But such improvements are tough to quantify; more qualitative than quantitative in terms of output. 

The upshot is that we should not be surprised when few organizations can point to quantitative output gains using accounting practices. First of all, it is too early for the big results to be proven. Also, some of the immediate gains are difficult to impossible to quantify in output, cash flow or profit figures. 

It will take time, even if investors are impatient.

AI ROI Metrics are Coming, Even if They are Essentially "Soft" Measures of Impact

We might as well be honest and predict that enterprises are going to develop all sorts of metrics that purportedly show the positive impact of their artificial intelligence investments, but that the metrics will quite probably be proxies that measure all sorts of things other than direct AI impact.

Still, some common metrics are a starting point:

• Cost per unit of output — Does AI reduce the labor or compute cost to produce a document, resolve a ticket, process a claim, underwrite a loan?

• Throughput / cycle time — How many units processed per hour, or how much time shaved off a workflow (e.g., code review, contract drafting, customer onboarding)?

• Error rates and rework costs — Does AI reduce defect rates, compliance exceptions, or manual correction loops?

• Headcount avoidance — the ability to scale output without proportional headcount growth. Often measured as "FTE equivalents automated."

Revenue-side metrics are less common, but might include:

• Conversion lift — Does AI-personalized outreach or recommendation improve sales conversion rates?

• Revenue per sales rep — If AI handles pipeline qualification or proposal drafting, does rep productivity improve?

• Customer retention / churn reduction — Does AI-assisted support or proactive intervention improve net revenue retention?

• Time-to-market — Does AI-accelerated R&D or software development compress product cycles in ways that generate earlier revenue?

Other operational outcomes also sometimes are quantified:

• Accuracy or precision rates on specific tasks (e.g., document classification, anomaly detection in fraud)

• Audit findings or compliance exceptions reduced

• Model risk KPIs — false positive/negative rates in detection systems

• Employee time recaptured — hours per week freed from low-value tasks, redirected to higher-value work

• Employee satisfaction / retention — particularly in roles prone to burnout from repetitive work

• Decision quality — harder to measure, but some firms track downstream outcomes of AI-assisted decisions against historical baselines

As rational as all that sounds, the metrics are “soft.” The attribution problem is severe, as AI is almost never the sole variable changing in a deployment. 

AI might be deployed while other changes also are occurring:

• Process redesign — Most AI deployments force workflow reengineering. Efficiency gains may be 60% process change and 40% AI

• Training and change management — The same tool deployed with weak adoption programs vs. strong ones produces dramatically different outcomes

• Data quality improvements — Organizations often clean and structure data as a precondition to AI deployment; that alone drives gains

• Personnel changes — New hires, role restructuring, or management changes co-occur with AI rollouts

• Macroeconomic or market tailwinds — Revenue gains during an AI deployment may reflect market growth, not AI impact.

• Hawthorne effects — Measuring a team's performance changes behavior regardless of the tool.

The point is that it can be almost impossible to isolate the impact of AI cleanly. So most enterprise AI ROI figures are really "ROI of the initiative that included AI," not AI's marginal contribution.

The more interesting question might be "which specific processes have changed in ways we can measure, and do we understand why?"

Skeptics are correct to argue that attributing success purely to AI is often an oversimplification. But enterprises will have to try and do so, as investors will demand such “proof.”

So firms will supply such “proof” as best they can, even if the outcomes are not, strictly speaking, solely because of AI use. 

And that is not an unusual case. 

Research highlights that AI’s impact is heavily moderated by "complementary assets.” In other words, a firm’s  organizational structure, existing data quality and worker skill levels often do more to determine the outcome than the AI model itself.

Study Focus	Key Finding Regarding Attribution	Source
Productivity Paradox	AI adoption does not guarantee boosts; results are contingent on organizational structure and worker attributes.	Cho et al. (2026)
Social Penalty/Bias	Using AI for assistance causes observers to attribute success to the tool rather than the person, leading to negative competence assessments.	Reif (2025)
Supply Chain/Bias	In complex systems, responsibility is fragmented across vendors/platforms, making it nearly impossible to attribute specific outcomes to one source.	Sharma et al. (2026)
Task-based Impact	AI improves performance within its "capability frontier" but degrades it outside that range; attributing net gains requires granular task-level data.	Brynjolfsson et al. (2023)

The difficulty in quantifying the immediate return on investment for new technologies is a recurring theme in economic history.

During the 1970s and 1980s, despite massive corporate investment in information and communications technology, overall productivity growth in many industrialized nations remained stagnant. This led economists to question whether computers were truly providing the expected value.

Eventually, results were observed, but:

• Results lagged deployment: it took decades for firms to fully "reimagine" their organizational structures, business models, and workflows to leverage the new technology effective

• Value was indirect: better management, more efficient coordination or improved service quality, but correlation, not causation, remains a question. 

The measurable financial benefits of a transformative technology often became clear only after business processes were redesigned.

Technology	Scope of Impact	Key Findings	Source
ICT / General IT	U.S. Economy (1995–2000)	ICT accounted for 56% of labor productivity growth; added 1.18 percentage points to GDP growth.	Oliner & Sichel (2000)
Emerging Tech (AI/ML)	U.S. Public Firms (2009–2019)	Over a three-year period, “neither the mean nor the median abnormal ROE (expected performance) reaches statistical significance in the post-implementation period.” “The mean abnormal inventory turnover is −1.06, which is not significantly different from zero.” “Overall, our results…indicate no significant difference in performance between sample and control firms during the implementation period of emerging digital technologies.”	Li et al. (2024)
Internet / ICT	SME Growth (Global)	Web-savvy SMEs grew more than twice as fast as those with minimal web presence.	McKinsey (2011)

Still, in the meantime, we will see all sorts of metrics “demonstrating” AI impact. Enterprises making the investments have no choice but to try to do so, even if those metrics are “soft.”