The easiest AI metrics to collect are becoming the easiest ones to misuse.
Seats assigned. Active users. Messages sent. Tokens generated. These numbers can tell you whether people have access to AI and whether they are using it. They cannot tell you whether revenue increased, margin improved, risk fell, or a customer received a better outcome.
That distinction moved to the center of the AI ROI conversation this week. OpenAI’s new working paper analyzed more than 17 million messages across more than 1,500 organizations. It documents rapid, broad enterprise adoption. But in the paper’s public-company analysis, lagged revenue per employee did not have a statistically significant association with any of four measures of ChatGPT Enterprise usage intensity.
That is not evidence that ChatGPT creates no value. The analysis was not designed to test whether AI caused future revenue growth. It is evidence that usage telemetry does not answer the financial question leaders keep asking it to answer.
The next AI dashboard should not simply show that people used the tool. It should show what changed after they used it.
Here’s the week’s data—and what it means for leaders and professionals trying to make AI pay.
The numbers worth knowing this week
ChatGPT Enterprise output tokens grew roughly sevenfold from June 2025 to March 2026. OpenAI’s working paper found that about half of this growth came from organizations already using the product, while the rest came from new adopters. That is meaningful evidence of diffusion and deeper use. It is not, by itself, evidence of financial return. The paper explicitly says its task data do not measure downstream work products, productivity effects, or changes in organizational routines. (OpenAI)
None of four usage-intensity measures showed a statistically significant association with lagged revenue per employee. In OpenAI’s Table 2, the measures were messages per active week per employee, weekly active users per employee, output tokens per employee, and messages per weekly active user. Depending on the measure, the analysis included 396 to 482 ticker-year observations from U.S.-based public-company adopters. The regressions relate current usage to earlier financial characteristics; they do not estimate AI’s causal effect on future revenue. The right conclusion is narrow but important: messages, users, and tokens are not financial proof. (OpenAI)
“Frontier” firms generated 8.3 times as many output tokens per active user as typical firms in June, up from 2.6 times in January. In a separate Enterprise Signals analysis, OpenAI used token intensity to compare its heaviest-using enterprise customers with a middle group. OpenAI also states directly that tokens are an imperfect measure of business value: a short response can be valuable and a long response can add little. The frontier gap shows a widening difference in how deeply firms use AI—not a measured 8.3-times financial advantage. (OpenAI Enterprise Signals)
AI reportedly supports 30% of tasks at the average surveyed business, while leaders expect 21% ROI this year. SAP and Oxford Economics surveyed 2,600 director-level and higher respondents across 13 countries between March and May 2026. The average respondent organization reported $28 million in AI spending and expected 21% ROI this year, rising to 38% in two years. These are survey-reported estimates and expectations, not audited returns. They show how much value leaders believe AI should create—and how important it is to build measurement that can later confirm or challenge that belief. (SAP and Oxford Economics)
The pattern that matters
Enterprise AI now has plenty of adoption data. What many organizations still lack is the bridge from adoption to business performance.
A useful AI measurement system has three layers:
● Adoption: Who used the tool, how often, and for which approved use cases? ● Workflow performance: Did completion time, quality, rework, handoffs, or exception rates improve? ● Financial realization: Did the change increase revenue or margin, reduce a real cost, protect revenue, avoid a loss, or create capacity that was intentionally redeployed?
Each layer answers a different question. Problems begin when a company reports the first layer as proof of the third.
High activity can coexist with poor outcomes. A team can send thousands of prompts while correcting weak drafts, repeating failed steps, or using AI on work that was never a bottleneck. A low-volume workflow can create substantial value if it prevents an expensive error, improves a high-value decision, or helps a professional act at the right moment.
That is why every AI use case needs a measurement chain before it scales:
Usage → workflow change → business outcome → financial effect.
If a team cannot define the next link, it should not claim the one after it. More usage may be a healthy adoption signal. It is not a shortcut to an ROI calculation.
The EMDUIT take — for the professional putting AI to work
Consider an insurance agency using AI to prepare renewal summaries and draft client outreach.
The usage dashboard may show that every producer logged in, hundreds of summaries were created, and message volume doubled. That tells leadership the team adopted the system. The workflow dashboard asks better questions: How many summaries were approved without material correction? Did preparation time fall? Were missing documents or coverage questions identified earlier? Did the handoff to a licensed professional happen at the right point? Were compliance-sensitive recommendations kept under human review?
The financial layer then asks whether those improvements affected retention, premium preserved, service capacity, avoidable rework, or cost per completed renewal. Only then can the agency see whether adoption became operating value.
This is why EMDUIT focuses on the full operating loop around AI: the task, the behavior, the guardrails, the handoff, the measurement, and the next improvement. The goal is not to make the usage chart go up. It is to create a workflow whose business result can be observed, trusted, and improved.
Usage tells you the system was touched. ROI tells you the business changed.
Which AI dashboard in your organization still reports activity without showing the business result that followed?
If you are a licensed professional or business leader, share the workflow you are ready to measure differently.
Sources & methodology
Sources: OpenAI, How Organizations Use AI: Evidence from ChatGPT (working paper, updated August 11, 2026) | OpenAI, Enterprise Signals: What Frontier Firms Are Doing Differently (updated August 12, 2026) | SAP and Oxford Economics, The Value of AI 2026 (published July 15, 2026).
Methodology note: The OpenAI working paper’s financial analysis covers a selected subset of U.S.-based public-company adopters and relates current usage intensity to lagged financial characteristics; it does not estimate AI’s causal effect on future revenue. SAP/Oxford Economics findings are vendor-commissioned, survey-reported estimates and expectations; they are not presented as audited realized returns.