Every AI drafting-rollout memo I read says the same thing. Faster emails, clearer emails, same voice as before. The measure of success is the acceptance rate.
The acceptance rate is measuring the wrong signal.
The Osaka desk
A senior comms director at a global asset manager showed me two folders last month. Same team in Osaka. One folder from March, one from July. Copilot in Outlook went live between them.
She pulled twenty emails from each folder and lined them up side by side. Nobody in Osaka had complained. The acceptance rate was healthy. The adoption dashboards were on the green side.
She scrolled slowly. The March folder had softeners, small apologies, polite request forms. The July folder read like the team had been translated into English by someone who did not speak Japanese. Still fluent. Still on-brand for the firm. Also, quietly, no longer the team's voice.
She asked me if I had ever seen this measured. I had not.
What Sakai and colleagues just measured
A preregistered experiment from Indiana University Bloomington and the University of Washington ran 175 full-time employed adults through three email-writing tasks. The sample was 89 American and 86 Japanese, each writing in their own language. The team coded 771 workplace emails for politeness markers, apologies and polite request forms.
The drift is bidirectional and asymmetric. When the AI draft matched the writer's cultural register, the writer's email barely moved. When the draft disagreed with the register, the writer's email lurched toward the draft by roughly an order of magnitude more.
Japanese participants under a misaligned draft: |M| = 3.70 markers per 100 words. Under an aligned draft: |M| = 0.23. American participants under a misaligned draft: |M| = 4.90. Under an aligned draft: |M| = 0.53. Culture-by-draft interaction β = 1.50, p < .001.
Then the part that stopped me. Japanese participants retained more of the culturally-aligned draft (0.904 vs 0.869, p < .001). Americans showed no draft-type difference at all (p = .689). One "acceptance rate" number can hide opposite reliance dynamics under the same headline.
Every off-the-shelf drafting assistant defaults to one communication register. The writers whose native register disagrees with that default are the ones being reshaped hardest. The metric on the rollout dashboard cannot see it. Before the next wave, I would run the small internal version of Sakai's study on a representative sample of writers. One task without the assistant, one task with it. Then measure the shift in the specific markers your firm cares about. The Osaka voice was worth something to that firm. Nobody priced what it would cost to lose it.