Itoki AI model on 854 workers: meetings count, not length, hurts

Itoki, an office design company, began joint research with Matsuo Institute in 2025, starting from the idea that offices exist to raise productivity. The two identified two obstacles: output has no common definition across jobs and companies, and too many factors — HR systems, work tools, physical and mental state — are entangled. To sidestep subjective surveys, they turned to everyday behavioral logs.

The model has three layers: 70 behavioral features such as clock-in time, number of moves, meeting counts, and email and chat volumes; a condition layer covering mind, body, and relationships; and a self-rated performance layer. Using the VAR-LiNGAM causal method, they estimated both the direction and the time lag from one day's actions to the next day's state, and flagged a relationship as effective only when 70% or more of employees in a job type shared the same direction. The findings split into three: results confirming conventional wisdom, an unexpected discovery, and a refutation.

What this means hinges on scope and replication. The data covers one month at a single site, and the team itself notes that online and in-person meetings were not separated and that how often training data should be refreshed remains an open question. Because what moves condition differs by organization, expanding to other companies would require building a model per customer — so the value of the finding may rest less on the specific numbers and more on whether Itoki can turn this into a repeatable service.