OpenAI EU Agent Report: How Should Companies Calculate ROI After a Surge in Usage?

As companies adopt AI, it’s becoming increasingly easy to achieve an impressive growth curve: more accounts, more tasks, and higher token consumption. But what procurement leaders really need to ask is whether delivery times have improved, whether rework has decreased, and whether the new costs are offset by the benefits.OpenAIThe EU Agent Usage Report, released on October 14, 2026, provides insight into changes in usage and serves as a reminder for companies to reevaluate their measurement methods.

First, understand the statistical scope of the report

据Introduction to the Official OpenAI ReportIn its EU ChatGPT Enterprise sample, Codex and ChatGPT Work already account for more than half of the output tokens; since May 2026, the number of weekly active users of these two products among EU small and medium-sized enterprises has increased more than eightfold. The report classifies businesses with no more than 200 associated users as small and medium-sized enterprises.

These data come from OpenAI customers and cannot be directly extrapolated to reflect AI adoption across all EU companies. Output tokens reflect the volume of content generated, while weekly active users reflect usage; neither metric directly indicates profit, labor-hour savings, or delivery quality. Therefore, interpreting the aforementioned growth as year-over-year productivity growth would go beyond what the data can support.

The same increase in dosage may lead to different results.

Here’s a managerial scenario: Two teams both double their token consumption—one completes more qualified orders, while the other spends most of its computing power on repeated attempts and correcting answers. On the procurement ledger, both teams appear equally active; yet their operational results may differ significantly. Ranking teams based solely on consumption may also create misguided incentives, leading employees to break tasks down into smaller pieces and generate longer model outputs.

Nor should a decrease in usage be interpreted as a failure of the approach. After a team has organized templates, cached shared materials, and eliminated unnecessary steps, it may be able to complete the same work with fewer calls. Therefore, while usage intensity is useful for identifying process changes, evaluating effectiveness also requires standardized task definitions, acceptance criteria, and cost parameters.

Adjust the denominator of the ROI

Companies often track subscription fees and API billing, but overlook the time spent on system integration, data organization, managing permissions, and training employees. During the pilot phase, the additional checks performed by project leads are often not accounted for separately; once the system is rolled out, the auditing process may become a new bottleneck. A reasonable cost analysis should cover preparation, execution, auditing, rework, and maintenance—not just the model invocation phase.

Duplicate counting of benefits should also be avoided. If the same person saves two hours—and uses that time to process more orders, the additional output should be recorded; if used to reduce overtime, the actual reduction in expenses should be recorded; and if it simply results in a more relaxed schedule, it should be accurately recorded as freed-up time. These two hours should not be counted twice—once as a wage cost and once as additional revenue.

For tasks that do not generate direct sales revenue, it is more practical to first evaluate the unit cost of acceptable delivery. For example, to complete a set of approved supplier documentation, how many man-minutes and how much modeling cost are actually required, and how much time is needed to correct errors if they occur? This metric can be used to compare performance against the original process as well as across different tools.

What Evidence Should Be Preserved for a Small Pilot Project?

This article recommends starting with tasks that are repetitive, have clear acceptance criteria, and can be corrected if errors occur, and ensuring that verifiable records are kept for the pilot program:

• Select tasks of similar difficulty, while retaining both the original workflow and the agent workflow, to avoid selecting only samples that are likely to succeed.
• Clearly define the acceptance criteria in advance, distinguishing between "passed on the first attempt," "passed after manual correction," and "ultimately failed."
• Record the total time elapsed from when the task is received until the results are accepted, including waiting time, verification, and rework.
• Record the number of minutes spent on manual intervention separately, as well as any additional costs incurred after a task fails.
• Separate short-term training costs from ongoing operating costs, and specify the statistical period and sample size.

Consider a purely hypothetical scenario: In the original process, it takes 60 minutes to deliver a qualified document. If an agent reduces the time needed for the first draft to 10 minutes, but review and revision take 45 minutes, the actual time saved is only 5 minutes. If, in another process, the first draft takes 20 minutes and review takes only 10 minutes, the latter might actually be more worth promoting. Records of the delivery chain can reveal these differences.

Procurement Competition Will Move Toward Verifiable Delivery

From a corporate procurement perspective, products that provide a complete task history, clear billing, and exportable acceptance data are more likely to be included in long-term budgets. Modeling capabilities are crucial, and the team’s ability to accurately understand what the product has accomplished and where it has failed also influences renewal decisions.

This report is useful for raising the question: Are there new opportunities for automation in our business? The next step should be to focus on a few specific processes, establish a realistic baseline, and then decide whether to scale up investment. Ultimately, what companies need is a steady increase in qualified output, accompanied by a corresponding cost structure.

Source:OpenAI: The EU’s Next Chapter in AI Is Agentic(October 14, 2026). The evaluation framework and hypothetical scenarios described in this paper are based on independent analyses.