The money AI saves in time, and where it disappears to

By Turing Post · July 2026.

This email is the seventh installment in the series "The Org Age of AI," co-written by Will Schenk (of TheFocus.AI) and Ksenia Se, dedicated to analyzing why generative artificial intelligence does not always translate into visible return on investment (ROI) for companies. The episode opens with a quote from economist Robert Solow: "You can see the computer age everywhere but in the productivity statistics," which serves as the throughline for the text's central argument.

The authors recall that the series began by posing a contradiction: AI is perceived as powerful, but the ROI still fails to appear clearly. After six previous episodes, they claim to be able to answer with more precision: AI can improve a specific task without improving the workflow around it, and it can improve that workflow without changing any outcome the company knows how to capture in its accounts. In other words, the technology can work exactly as promised and yet the value gets lost somewhere between the individual employee and the profit and loss statement (P&L).

The email also introduces the question of what an "AI-native enterprise" really is, following the authors' attendance at several industry conferences. According to their framing, it is not simply a company where a lot of people use AI, or where one or more agents complete impressive work: that is not enough to consider it AI-native. An AI-native enterprise would be one capable of turning artificial intelligence into repeatable organizational outcomes. The authors call this gap "the conversion gap," and argue that this is where most companies remain weak.

The bulk of the visible content draws on a real economic study: the work of economists Anders Humlum and Emilie Vestergaard, published in 2025 as an NBER working paper, which analyzed the use of generative AI in Denmark. The study combined survey responses from approximately 25,000 workers across some 7,000 workplaces with administrative records of earnings and hours worked, covering 11 occupations considered highly exposed to AI chatbots.

The workers who used these tools did report benefits: on average, they said they had saved around 2.8% of their total working time. Depending on the occupation, between 64% and 90% of respondents said they had saved at least some time, and most said they had redirected that time toward other tasks.

However, according to the study, those gains did not translate into detectable changes in recorded hours or earnings during the first two years after the launch of ChatGPT. The estimates were close to zero, with confidence intervals that ruled out average effects greater than roughly 2%. Employment and the wage bill also remained largely stable.

The authors are careful in interpreting this finding: they stress that the study does not prove AI created no business value, because it did not measure variables such as revenue, profits, customer satisfaction, output quality, avoided risk or speed of execution. It only shows a narrower pattern: workers reported time savings at the individual-task level, while administrative records showed no detectable change in hours worked or earnings.

From this they draw the episode's central idea: a task can become easier while the economy around it stays the same; something has to convert one into the other, and that conversion does not happen automatically.

To explain this phenomenon, the text returns to a model the authors call "the Capacity-to-Outcome Chain." They recall that the first episode of the series already argued that the bottleneck had shifted from the AI model's capability to organizational translation, describing three necessary transformations: tacit knowledge must become usable context, context must become bounded action, and human correction must become a feedback loop. According to the authors, these transformations make reliable AI work possible, but on their own they do not determine what the organization does with the resulting output.

The email lists the topics that will be developed next in the full episode: the capacity-to-outcome chain itself, six ways that capacity can be converted into value, a review of similar paradoxes from the past, the idea that the organization explains more than the individual, what "conversion discipline" consists of, and why the last link in the chain is authority. However, the detailed development of these points is left behind the paywall for Turing Post premium subscribers (they mention as notable subscribers people from companies such as Microsoft, Nvidia, Google, Hugging Face, OpenAI, a16z, as well as AI labs such as Ai2, MIT, Berkeley and government bodies), so the body of the email does not include the full development of those six points.

As alternative content, the email links to a YouTube video from Turing Post itself titled "Kimi K3 vs Inkling: Who Will Win Open Source AI?", dedicated to the week's most relevant open source model releases, aimed at those not interested in the newsletter's main topic.

Finally, the email links to the previous episode in the series, number 6, titled "The Flywheel: What Happens When Workflows Run Themselves," for those who want to follow the full line of argument from the beginning.