After several years of experimentation, the return on investment of artificial intelligence has become a central concern for executive management. Companies can no longer simply accumulate proofs of concept, generative assistants, and demonstrators. They must now answer a more demanding question: what economic value does artificial intelligence actually produce?

A study published on July 15, 2026, by SAP, in partnership with Oxford Economics, provides new insights into this debate. Conducted among 2,600 executives in thirteen countries, it indicates that the surveyed companies allocate an average of $28 million to AI. They estimate their return on investment at 21% in 2026, compared to 16% in 2025, and anticipate a rate of 38% over the next two years. The expected financial value would thus increase from $6.3 million to $15.9 million.

These figures seem to confirm that artificial intelligence is gradually leaving innovation labs to enter operations. However, they must be interpreted with caution. The study measures declared returns and executive expectations. It does not constitute an independent consolidation of the accounting results produced by each project.

The nuance is essential. An expectation of profitability is not yet realized value. A technical capability is not yet an economic result.

The real challenge is to understand how a company converts a technological improvement into measurable performance. Artificial intelligence can speed up a task, produce an analysis, or automate an action. However, it does not, by itself, create economic value. This depends on strategic choices, data quality, work organization, decision-making processes, and the ability of teams to use the time or resources thus freed up.

From Experimentation to Execution

According to the SAP study, approximately 30% of tasks performed in the average company are now supported by some form of artificial intelligence. This proportion could reach 48% over the next two years. At the same time, the share of investments qualified as strategic has almost doubled to reach 17%. Yet, 41% of organizations continue to deploy AI in the form of fragmented initiatives.

This situation reveals a common contradiction. Companies are using AI more, but they do not always have a common strategy to select use cases, coordinate investments, and measure results.

An internal assistant can provide quality summaries. An agent can speed up the processing of customer requests. A predictive model can improve anomaly detection. These results attest to the proper functioning of a technology. They do not automatically constitute a return on investment.

To produce economic value, technical capability must lead to an observable improvement in a business process. This can translate into reduced processing time, lower unit costs, decreased error rates, or increased revenue, improved quality, or better risk management.

Two companies can use the same model and obtain very different results. One can integrate the tool into a clearly defined process, train teams, organize result validation, and measure performance evolution. The other can make the same tool available to employees without modifying responsibilities or working methods.

The difference then does not come from the model. It comes from the organization.

Time Saved Is Not Yet ROI

A large part of the benefits associated with artificial intelligence is currently expressed in hours saved.

An employee can prepare a document in thirty minutes instead of two hours. A developer can produce a first version of code more quickly. A customer service department can automate part of its responses. An analyst can review a document set in a few minutes.

These gains are real. However, they constitute a productivity improvement, not necessarily a financial saving.

The time freed up only becomes economic value when it is reassigned to a useful activity. It can allow processing more cases, responding to more customers, accelerating a project, reducing the use of service providers, or improving service quality.

When this reallocation is neither organized nor measured, the gain remains theoretical.

A company should therefore not announce a return on investment by simply multiplying the number of hours saved by the average hourly cost of employees. This calculation assumes that every hour saved is immediately converted into cost reduction or additional activity. In practice, this conversion depends on management choices, available workload, and the ability of teams to absorb new assignments.

To measure value creation, several stages should be distinguished.

  • The first corresponds to technical capability. AI accelerates, automates, or assists a task.
  • The second concerns operational gain. Delays decrease, volume processed increases, or the number of errors declines.
  • The third assumes a transformation of company operations. Roles, processes, or decision circuits are adapted.
  • The last corresponds to realized economic value. Revenues increase, costs decrease, margins improve, or a risk is reduced.

Many organizations reach the first two stages. Far fewer manage to sustainably convert these improvements into financial performance.

Agentic AI Raises Expectations

Agentic artificial intelligence occupies an important place in the expectations described by the SAP study.

Unlike a generative assistant, which primarily produces a response, summary, or content, an agent can execute a sequence of actions. It can search for information, consult different systems, interpret a situation, propose a decision, trigger an operation, and verify the result.

The surveyed companies anticipate an average return of $17.6 million linked to agentic AI over the next two years, more than four times the estimate made the previous year. Furthermore, 83% of executives attribute moderate to very high transformation potential to agents. Yet, only 3% of organizations declare themselves fully prepared for their deployment.

The gap between ambitions and the level of preparation constitutes a warning signal. An agent does not simply generate text. It can act on a process, access data, communicate with applications, and trigger operational, economic, or legal consequences. The more its autonomy increases, the more the company must clarify the distribution of responsibilities between people, agents, and information systems. The performance of an agent therefore cannot be evaluated solely by its execution speed or by the number of automated tasks. It must also integrate its reliability, error rate, level of autonomy, decision quality, traceability, and the cost of its supervision.

The study indicates that 38% of companies do not yet have a systematic human supervision process for their agentic workflows. About 37% have not implemented appropriate controls regarding rights and access, while only 44% maintain a registry of agents present in their organization.

An agent can reduce an operational cost, but it can also amplify an error on a large scale. It can streamline a process, but also introduce a compliance, cybersecurity, or loss of control risk. The economic potential of agentic AI therefore remains inseparable from the quality of its integration into the company.

Data Conditions Value Creation

The SAP study identifies data quality as one of the main obstacles to return on investment. Nearly 73% of surveyed organizations report difficulties related to incomplete data. In parallel, 79% declare suffering from rework, delays, or task accumulations caused by insufficient quality AI results.

These results recall a reality often obscured by technological discourse: artificial intelligence does not automatically compensate for a company’s informational weaknesses.

When a model is connected to inconsistent, obsolete, or scattered data, it can accelerate the production of errors. When it does not have the necessary business context, it can provide plausible but not very usable answers. When it remains isolated from the company’s systems, its contribution to performance remains limited.

Data therefore does not only constitute a technical resource. It represents an economic infrastructure. A company wishing to measure the return on investment of its AI must integrate the expenses necessary to make its data accessible, documented, secure, and usable. These expenses can include cleaning, integration, normalization, access rights management, and the implementation of traceability mechanisms.

A calculation that excludes these elements risks overestimating the actual profitability of the project.

Calculating the Total Cost of Artificial Intelligence

The return on investment of AI is often overestimated because organizations underestimate the costs necessary for deployment.

They generally take into account licenses, access to models, and infrastructure expenses. They sometimes forget less visible costs: integration with existing systems, data preparation, training, control, cybersecurity, compliance, maintenance, and change management.

The calculation must also include the time spent by business experts on designing the system, validating results, and correcting errors.

In some projects, AI accelerates the production of a first version but simultaneously increases the verification workload. In others, data limitations force teams to manually rework part of the results. These costs must appear in the evaluation.

A simplified formula can be used:

AI ROI = (realized financial value − total cost of ownership)/total cost of ownership

The main difficulty does not lie in the formula, but in defining the realized financial value.

This can come from several sources:

  • effective reduction of expenses;
  • increased revenues;
  • improved margins;
  • accelerated sales cycle;
  • decreased error rate;
  • lower acquisition cost;
  • reduced incidents;
  • limitation of a financial or regulatory risk.

These indicators must be defined before deployment and compared to a baseline situation. Without a starting point, it becomes difficult to attribute an improvement to artificial intelligence rather than to a market evolution, a reorganization, or another investment.

It is also necessary to distinguish three forms of value. Realized value is already visible in the accounts or in operational indicators. Avoided value corresponds to a cost, error, or risk that did not occur. Potential value is based on a projection or capability that has not yet been converted into a result.

Confusing these categories leads to presenting anticipated benefits as acquired results.

Governance Also Has an Economic Function

The governance of artificial intelligence is still sometimes perceived as a constraint likely to slow down projects. However, it can directly contribute to their profitability.

According to the study, only 12% of companies consider their skills, processes, or frameworks fully ready to effectively frame AI. At the same time, 69% declare observing at least occasionally the use of unapproved tools, generally described as “shadow AI.”

This situation creates hidden costs: multiplication of subscriptions, duplication of solutions, exposure of sensitive information, lack of data tracking, and difficulty identifying responsibilities. Governance should therefore not be reduced to regulatory compliance. It also makes it possible to select investments, avoid redundancies, assign responsibilities, define supervision levels, and compare results between several projects. It provides the necessary rules to continue, correct, or stop an initiative based on the value produced.

It is not about systematically slowing down innovation. It is about determining which decisions can be automated, which must remain supervised, and under what circumstances human intervention becomes necessary.

Proportionate governance improves the quality of decisions and reduces the cost of errors. As such, it contributes to value creation.

Five Questions for Executive Management

Before increasing budgets allocated to artificial intelligence, an executive committee should examine five questions.

What economic problem do we want to solve?

The project must start from a clearly identified business difficulty or opportunity, not from an available technology.

Which indicator should evolve?

The expected result must be measurable: cost, time, revenue, quality, risk, customer satisfaction, or processing capacity.

What was the situation before deployment?

An initial baseline is essential to measure evolution and correctly attribute results.

What costs have we actually integrated?

The calculation must include data, integration, training, supervision, maintenance, and change management.

What change in operations is necessary?

The company must determine how work, decisions, skills, and responsibilities will evolve after deployment.

These questions shift the debate. They force executives to no longer consider artificial intelligence as a simple technological purchase. They make it a strategic and organizational decision.

From Use Case Catalog to Value Portfolio

Many companies have built use case catalogs. They must now build value portfolios. Each initiative should have a business owner, an economic indicator, a full cost, a baseline situation, a deployment trajectory, and stop criteria. Projects can then be grouped according to several purposes: operational efficiency, revenue growth, improved customer experience, risk management, or business model innovation. This approach makes it possible to compare very different initiatives and arbitrate investments according to their actual contribution to strategy. It also avoids concentrating all efforts on the productivity gains that are easiest to measure. Some applications produce more indirect value, such as improved decision quality, risk reduction, or accelerated organizational learning. These benefits can be strategic, but they must be documented with the same rigor.

The question is therefore no longer just: how many hours have we saved? It becomes: what economic result have we obtained, for what total cost, with what level of reliability, and in service of what strategic priority?

Organizations Will Make the Difference

The SAP study confirms that companies now expect higher returns on their investments in artificial intelligence. It also shows that their level of preparation remains uneven.

Usage is progressing faster than data quality, skills, processes, and control mechanisms. Ambitions related to agents sometimes exceed the actual capacity of organizations to integrate and supervise them.

The return on investment therefore does not depend solely on the power of models. It depends on the company’s ability to connect technology to its objectives, data, processes, and responsibilities.

Artificial intelligence creates new possibilities for action. It can accelerate, analyze, recommend, and automate. But it does not decide how the freed-up time should be used, which processes should evolve, or which economic priorities to pursue.

These decisions remain the responsibility of the organization.

Models will continue to progress. Costs will evolve. Technical capabilities will become more accessible. As tools spread, competitive advantage will shift toward the quality of their integration.

For executives, the question is therefore no longer how many tools have been deployed.

It is to determine what results have been obtained, for what total cost, and in service of what strategy.

It is on this condition that artificial intelligence can become a sustainable economic asset rather than a succession of costly experiments.