None of the three companies bans the tool [21]. What they do instead is relocate authorship: Leapsome's guideline states that "'AI wrote it' isn't an answer if someone asks what you meant," and that "Every word should be something you back" [13]; Clay's requires employees to stand behind every idea and sentence in a document [17]. Drafting itself is unaffected; the shift shows up afterward, in the conversation once the document lands, because the writer can no longer treat the model as a co-signer.
The cost these rules are aimed at sits on the reading side of the transaction. Clay's policy puts it in accounting terms: extra time a reader spends decoding your meaning is a multiplicative cost on the team, while time spent making it clear first is a one-time cost every reader benefits from [18]. Run the BetterUp and Stanford figures through the same lens and the order of magnitude is visible: 460 of the 1,150 respondents reported an instance in a single month [1], and if each of those was one instance at close to two hours, a 100-person desk team is absorbing roughly 80 reader-hours a month, the equivalent of two full working weeks [2]. That figure is a floor rather than a precise estimate, and it does not appear in any productivity dashboard, because dashboards count documents produced.
Which is where the trade-off gets awkward for whoever wrote the adoption mandate. Jessica Zwaan of Leapsome theorises that workers overuse AI precisely because they have been told its use is an imperative, and worry that a short note will make them look lazy [9]; her formulation is that people associate productivity with volume when productivity is clarity [10]. A company that pushed usage as the metric last year is now issuing a second instruction that partly contradicts the first. You cannot both measure AI adoption by output volume and ask for brevity; one of those two numbers has to stop being a target.
Read narrowly, this is manners dressed up as governance, a memo about memos, read once and filed. What gives it weight is what the policy replaces. Sophie Alpert, the Clay engineer who wrote its rules, said her goal was to create space for people to push back without it seeming like a personal judgment [16]; before that, her question about confusing passages drew the reply that she should ignore them [19]. A written standard converts an accusation of bad writing into a citation, and that is a different social act.
What the record does not yet show is effect. The three policies cover about 780 employees between them [4], all at software or software-adjacent firms, and they are weeks old: Clay gave its roughly 100-strong engineering team the rules about seven weeks before the other 82 percent of the company [3], and Leapsome's went live in mid-August [11]. Nothing in the account reports a before-and-after measurement of document length, reader time, or slop incidence at any of them [5]. The viral LinkedIn post from Clay's chief executive, with nearly 10,000 reactions as of September 3 [20], measures recognition among managers, not results.
So the honest verdict for anyone deciding this quarter is narrow. The evidence supports the claim that verbal nudges failed [1] and that written standards give reviewers something to point at. Whether they work depends on a person in the review chain being willing to return a document and name the clause, and that is next quarter's question, not this one's.
