The pressure to invest in AI is real. So is the expectation of a return.
Enterprise leaders face two demands at once. They must respond to the opportunities created by AI, and they must explain what those investments are delivering.
In a January 2026 survey by Boston Consulting Group, 2,360 executives, including 640 CEOs, were asked about their AI plans. Respondents expected corporate AI spending to increase from around 0.8% of revenue in 2025 to 1.7% in 2026. Ninety-four percent said their organisations would continue investing even if AI did not deliver returns in 2026. These were expectations and intentions, not measured financial results. [1]
McKinsey's August 2026 survey found that 80% of respondents said AI had improved their own productivity, while 37% reported a positive contribution to their organisation's earnings before interest and taxes. Only 6% met McKinsey's definition of AI high performers: respondents who attributed at least 5% of those earnings to AI and described its impact as significant. All three figures reflect survey responses, not an independent assessment of the financial returns achieved by the companies concerned. [2]
The findings raise a practical question: What can an enterprise genuinely measure when it invests in AI, and what can it reasonably attribute to that investment?
This four part series examines the evidence behind different kinds of AI benefits, the calculations that can be made and the questions that remain unanswered.
Why AI changes the measurement challenge
A company can often define what it expects from a particular technology investment. If it replaces a manual process with software, for example, it can compare how much work is completed, how long it takes and what it costs. That does not guarantee a positive return, but it provides a starting point for measurement.
AI can be applied much more widely. The same assistant might support research, financial analysis, customer service, document preparation, software development and business decisions. Its contribution is not always confined to one process.
Agentic AI expands these possibilities. Rather than only helping an employee complete a task, an AI system can retrieve information, work with business applications and carry out several steps in a process, with suitable supervision. In some cases, it can perform work previously undertaken by employees or external providers.
This may create an identifiable financial benefit. A company might spend less on an outside service, avoid a planned hire or complete more paying work with existing resources. But an AI agent taking over tasks does not automatically mean that payroll or other expenditure falls. It also introduces costs of its own, including software, integration, supervision and correcting errors.
These are not problems unique to AI. They are familiar questions in evaluating business investments. AI makes them more prominent because it can change so many different activities at once.
Four questions behind an ROI claim
When a company reports a benefit from AI, four questions help establish what the claim means.
Did performance improve? Work may be completed faster, with fewer errors, or in greater volume.
How much of the change was caused by AI? Sales, costs and productivity may also change because of people, prices, demand or other process improvements. A fair comparison is needed to separate these effects as far as reasonably possible. For example, a business could compare the cost of handling similar transactions before and after introducing AI, while checking whether other changes also affected the result.
Did the improvement produce a financial benefit? Time saved is not automatically money saved. Higher sales are not automatically higher profit. On the other hand, avoiding a genuine future expense or earning additional profit from work made possible by AI can have financial value, even without reducing headcount.
What did it cost to achieve that benefit? A return on investment calculation must consider the extra costs caused by the AI investment, including initial implementation, ongoing operation and human oversight, over a stated period. Costs the business would have incurred anyway should not simply be charged to the AI project.
These questions apply the same basic financial discipline used for other technology investments. They do not demand a scientific experiment or an independent audit for every project. Reliable company records, a reasonable comparison and clearly stated assumptions may be sufficient for management to make a decision. The strength of the conclusion should match the strength of the evidence.
Three categories for examining enterprise AI returns
The following articles use three categories to organise specific applications and their evidence. They are not permanent labels for entire business activities. The same activity may fall into different categories in different organisations, depending on what has been measured and what can be established.
A. Where a financial return can be calculated
Category A covers applications where a financial benefit attributable to AI can be estimated from a credible comparison and the relevant costs can be identified. The result may be positive or negative. Forecast benefits must be identified as estimates, rather than presented as money already saved.
For example, an AI system might take over work previously purchased from an external provider. The business could compare the expenditure it would otherwise have incurred with its actual expenditure, then account for the costs of the AI process. That creates a basis for calculating a return.
Microsoft provides a useful illustration of both the opportunity and the limits of published claims. In July 2026, it reported a 33% reduction in cost per transaction after using an AI toolkit as part of a transformation of roughly a quarter of its outsourced business processes. This is Microsoft's reported operating result, not a published calculation of net AI ROI. Its account does not provide enough detail about the original transaction costs, the precise basis of comparison or the costs of the new system for an outside reader to reproduce an ROI calculation. It also describes a wider process transformation, so the reduction cannot simply be attributed in full to AI alone. [3]
The distinction is important. A company can have enough internal information to calculate a return even when its public account does not disclose that information. We should report what the company says without claiming to have verified a calculation we cannot reproduce.
B. Where operational improvement is measured, but its financial return is not yet established
AI can help employees work faster, produce better documents or handle more tasks. Those are meaningful improvements, and they can often be measured directly.
Consider an employee who completes two days of research in two hours. The time saved is measurable. Multiplying those hours by the employee's wage, however, does not show that the company reduced its expenditure by that amount.
The released capacity may still have financial value. It might allow the company to serve additional paying customers or avoid hiring another employee when demand increases. If the additional profit or genuinely avoided expense can be demonstrated, the same application may qualify for Category A.
Category B means the operational improvement has been established, but the link to a financial return has not yet been shown. It does not mean that the return is necessarily zero or impossible to calculate.
C. Where AI's contribution to the result remains uncertain
Sometimes the business outcome is clear, but the available information does not show how much of it AI caused.
Consider a sales team using AI to research customers, prepare proposals and manage follow ups. Revenue might rise, but prices, market conditions, customer relationships and the salespeople themselves also affect the result. Without a credible comparison, the company cannot tell how much of the increase came from AI.
This is a question of evidence, not a claim that sales impact cannot be measured. In a working paper published in 2026, researchers examined seven AI applications on an online retail platform using randomly selected comparison groups. Their experiments, conducted in 2023 and 2024, found sales effects ranging from no detectable increase to 16.3%, depending on the application. That is evidence of AI's contribution to revenue in those specific tests. The paper does not establish the net ROI of the applications because it does not provide a full account of their costs and profit effects. Nor do the results establish what would happen when a company gives AI tools to an entire salesforce. [4]
The experiments show that better comparisons can answer questions that might otherwise remain uncertain. Similar efforts may be useful for customer service, sales and other business activities. Some long term strategic outcomes will be harder to assess, especially when decisions and market conditions change over several years.
Category C means that AI's contribution has not been established for the particular outcome being examined. It does not mean that no sound measurement method exists or could be developed.
What this means for enterprise leaders
The three categories are a way to examine the evidence, not a ranking of which AI applications are worth pursuing.
An operating improvement can be worthwhile even before its financial value is known. A financially measurable project can still lose money once its costs are considered. And a business result that cannot currently be attributed to AI may become measurable when better comparisons or records are available.
The important question is not simply whether someone has published a percentage or an ROI figure. It is what was measured, how much of the change can reasonably be credited to AI, which costs were included and which parts remain estimates or claims.
Where the next three articles go
Article 2: Measuring Financial Returns. We will examine candidate activities for Category A and show the calculation method for each one that can be supported: the formula, the information required, a reasonable comparison, the costs to include and the period being measured. Published company results will remain clearly identified as reported claims when the underlying figures are unavailable. We will not invent missing company data to produce an apparently verified ROI.
Article 3: Measuring Operational Improvements. We will examine Category B, including productivity, quality, speed and capacity, and show what additional evidence would be needed to establish a financial benefit.
Article 4: Understanding the Limits of Attribution. We will examine Category C, looking at where better records or comparisons could help and where the contribution of AI remains especially difficult to isolate.
The purpose of this series is not to produce an ROI figure for every use of AI. It is to distinguish what has been measured, what can be calculated, what has only been reported and what remains uncertain.
References
- Boston Consulting Group, As AI Investments Surge, CEOs Take the Lead, 15 January 2026.
- McKinsey, The State of AI in 2026: On the Road to ROI, 25 August 2026.
- Microsoft Inside Track, Streamlining Business Operations at Microsoft with an AI Toolkit, 23 July 2026.
- Fang, Yuan, Zhang, Donati and Sarvary, Generative AI and Sales Productivity: Field Experiments in Online Retail, CESifo Working Paper No. 12201, 2026. Earlier version: Generative AI and Firm Productivity: Field Experiments in Online Retail, arXiv:2510.12049. Experiments conducted in 2023 and 2024.
The pressure to invest in AI is real. So is the expectation of a return.
Enterprise leaders face two demands at once. They must respond to the opportunities created by AI, and they must explain what those investments are delivering.
In a January 2026 survey by Boston Consulting Group, 2,360 executives, including 640 CEOs, were asked about their AI plans. Respondents expected corporate AI spending to increase from around 0.8% of revenue in 2025 to 1.7% in 2026. Ninety-four percent said their organisations would continue investing even if AI did not deliver returns in 2026. These were expectations and intentions, not measured financial results. [1]
McKinsey's August 2026 survey found that 80% of respondents said AI had improved their own productivity, while 37% reported a positive contribution to their organisation's earnings before interest and taxes. Only 6% met McKinsey's definition of AI high performers: respondents who attributed at least 5% of those earnings to AI and described its impact as significant. All three figures reflect survey responses, not an independent assessment of the financial returns achieved by the companies concerned. [2]
The findings raise a practical question: What can an enterprise genuinely measure when it invests in AI, and what can it reasonably attribute to that investment?
This four part series examines the evidence behind different kinds of AI benefits, the calculations that can be made and the questions that remain unanswered.
Why AI changes the measurement challenge
A company can often define what it expects from a particular technology investment. If it replaces a manual process with software, for example, it can compare how much work is completed, how long it takes and what it costs. That does not guarantee a positive return, but it provides a starting point for measurement.
AI can be applied much more widely. The same assistant might support research, financial analysis, customer service, document preparation, software development and business decisions. Its contribution is not always confined to one process.
Agentic AI expands these possibilities. Rather than only helping an employee complete a task, an AI system can retrieve information, work with business applications and carry out several steps in a process, with suitable supervision. In some cases, it can perform work previously undertaken by employees or external providers.
This may create an identifiable financial benefit. A company might spend less on an outside service, avoid a planned hire or complete more paying work with existing resources. But an AI agent taking over tasks does not automatically mean that payroll or other expenditure falls. It also introduces costs of its own, including software, integration, supervision and correcting errors.
These are not problems unique to AI. They are familiar questions in evaluating business investments. AI makes them more prominent because it can change so many different activities at once.
Four questions behind an ROI claim
When a company reports a benefit from AI, four questions help establish what the claim means.
Did performance improve? Work may be completed faster, with fewer errors, or in greater volume.
How much of the change was caused by AI? Sales, costs and productivity may also change because of people, prices, demand or other process improvements. A fair comparison is needed to separate these effects as far as reasonably possible. For example, a business could compare the cost of handling similar transactions before and after introducing AI, while checking whether other changes also affected the result.
Did the improvement produce a financial benefit? Time saved is not automatically money saved. Higher sales are not automatically higher profit. On the other hand, avoiding a genuine future expense or earning additional profit from work made possible by AI can have financial value, even without reducing headcount.
What did it cost to achieve that benefit? A return on investment calculation must consider the extra costs caused by the AI investment, including initial implementation, ongoing operation and human oversight, over a stated period. Costs the business would have incurred anyway should not simply be charged to the AI project.
These questions apply the same basic financial discipline used for other technology investments. They do not demand a scientific experiment or an independent audit for every project. Reliable company records, a reasonable comparison and clearly stated assumptions may be sufficient for management to make a decision. The strength of the conclusion should match the strength of the evidence.
Three categories for examining enterprise AI returns
The following articles use three categories to organise specific applications and their evidence. They are not permanent labels for entire business activities. The same activity may fall into different categories in different organisations, depending on what has been measured and what can be established.
A. Where a financial return can be calculated
Category A covers applications where a financial benefit attributable to AI can be estimated from a credible comparison and the relevant costs can be identified. The result may be positive or negative. Forecast benefits must be identified as estimates, rather than presented as money already saved.
For example, an AI system might take over work previously purchased from an external provider. The business could compare the expenditure it would otherwise have incurred with its actual expenditure, then account for the costs of the AI process. That creates a basis for calculating a return.
Microsoft provides a useful illustration of both the opportunity and the limits of published claims. In July 2026, it reported a 33% reduction in cost per transaction after using an AI toolkit as part of a transformation of roughly a quarter of its outsourced business processes. This is Microsoft's reported operating result, not a published calculation of net AI ROI. Its account does not provide enough detail about the original transaction costs, the precise basis of comparison or the costs of the new system for an outside reader to reproduce an ROI calculation. It also describes a wider process transformation, so the reduction cannot simply be attributed in full to AI alone. [3]
The distinction is important. A company can have enough internal information to calculate a return even when its public account does not disclose that information. We should report what the company says without claiming to have verified a calculation we cannot reproduce.
B. Where operational improvement is measured, but its financial return is not yet established
AI can help employees work faster, produce better documents or handle more tasks. Those are meaningful improvements, and they can often be measured directly.
Consider an employee who completes two days of research in two hours. The time saved is measurable. Multiplying those hours by the employee's wage, however, does not show that the company reduced its expenditure by that amount.
The released capacity may still have financial value. It might allow the company to serve additional paying customers or avoid hiring another employee when demand increases. If the additional profit or genuinely avoided expense can be demonstrated, the same application may qualify for Category A.
Category B means the operational improvement has been established, but the link to a financial return has not yet been shown. It does not mean that the return is necessarily zero or impossible to calculate.
C. Where AI's contribution to the result remains uncertain
Sometimes the business outcome is clear, but the available information does not show how much of it AI caused.
Consider a sales team using AI to research customers, prepare proposals and manage follow ups. Revenue might rise, but prices, market conditions, customer relationships and the salespeople themselves also affect the result. Without a credible comparison, the company cannot tell how much of the increase came from AI.
This is a question of evidence, not a claim that sales impact cannot be measured. In a working paper published in 2026, researchers examined seven AI applications on an online retail platform using randomly selected comparison groups. Their experiments, conducted in 2023 and 2024, found sales effects ranging from no detectable increase to 16.3%, depending on the application. That is evidence of AI's contribution to revenue in those specific tests. The paper does not establish the net ROI of the applications because it does not provide a full account of their costs and profit effects. Nor do the results establish what would happen when a company gives AI tools to an entire salesforce. [4]
The experiments show that better comparisons can answer questions that might otherwise remain uncertain. Similar efforts may be useful for customer service, sales and other business activities. Some long term strategic outcomes will be harder to assess, especially when decisions and market conditions change over several years.
Category C means that AI's contribution has not been established for the particular outcome being examined. It does not mean that no sound measurement method exists or could be developed.
What this means for enterprise leaders
The three categories are a way to examine the evidence, not a ranking of which AI applications are worth pursuing.
An operating improvement can be worthwhile even before its financial value is known. A financially measurable project can still lose money once its costs are considered. And a business result that cannot currently be attributed to AI may become measurable when better comparisons or records are available.
The important question is not simply whether someone has published a percentage or an ROI figure. It is what was measured, how much of the change can reasonably be credited to AI, which costs were included and which parts remain estimates or claims.
Where the next three articles go
Article 2: Measuring Financial Returns. We will examine candidate activities for Category A and show the calculation method for each one that can be supported: the formula, the information required, a reasonable comparison, the costs to include and the period being measured. Published company results will remain clearly identified as reported claims when the underlying figures are unavailable. We will not invent missing company data to produce an apparently verified ROI.
Article 3: Measuring Operational Improvements. We will examine Category B, including productivity, quality, speed and capacity, and show what additional evidence would be needed to establish a financial benefit.
Article 4: Understanding the Limits of Attribution. We will examine Category C, looking at where better records or comparisons could help and where the contribution of AI remains especially difficult to isolate.
The purpose of this series is not to produce an ROI figure for every use of AI. It is to distinguish what has been measured, what can be calculated, what has only been reported and what remains uncertain.
References
- Boston Consulting Group, As AI Investments Surge, CEOs Take the Lead, 15 January 2026.
- McKinsey, The State of AI in 2026: On the Road to ROI, 25 August 2026.
- Microsoft Inside Track, Streamlining Business Operations at Microsoft with an AI Toolkit, 23 July 2026.
- Fang, Yuan, Zhang, Donati and Sarvary, Generative AI and Sales Productivity: Field Experiments in Online Retail, CESifo Working Paper No. 12201, 2026. Earlier version: Generative AI and Firm Productivity: Field Experiments in Online Retail, arXiv:2510.12049. Experiments conducted in 2023 and 2024.
Continue reading