Lower AI Use Targets Reduce Gaming and Raise Sales

Mandating AI use can solve an adoption problem while creating an incentive problem. A large medical-device company required roughly 5,000 employees to submit at least 200 monthly queries to an internal AI assistant, and first-time use rose after the policy. But the threshold also turned query volume into a target: employees concentrated activity at the cutoff, and 31% of observed queries were either repeated or off-task.

The paper studies the firm's response when it lowered the monthly requirement from 200 to 100 queries across branches at different times. Its central result is unusually clean for an AI-adoption study: the lower target reduced total query volume, but the contraction was concentrated in activity that appears least connected to work. Non-repeated, work-related querying changed little, while sales employees recorded higher monthly sales. The policy lesson is not that firms should abandon adoption targets; it is that a target calibrated to force initial experimentation can become counterproductive once use is established.

Core Contribution

The contribution is an empirical account of adaptive incentive design in workplace AI deployment. Much discussion of AI adoption treats usage as a simple outcome to maximize. The authors instead distinguish participation from productive use and show why a quota can make those measures diverge. A high minimum requirement induces employees to clear the metric, including through repeated prompts or queries unrelated to their jobs, rather than necessarily finding additional useful applications for the assistant.

That distinction matters because the firm did not simply observe low-quality use and issue generic guidance. It changed the rule itself. By exploiting staggered implementation of the 100-query target across branches, the study estimates the effect of recalibrating the requirement rather than merely comparing periods before and after a company-wide policy change. The design makes the paper more informative than a descriptive dashboard of AI usage, although it remains evidence from one organization and one internal tool.

Measuring Use and Gaming

The authors use assistant query logs to track adoption, query volume, timing, repetition, and whether a query is work-related. The classification scheme separates repeated and off-task activity from non-repeated, work-related queries; the appendix documents few-shot examples used for those categories. This allows the analysis to ask whether a lower quota suppresses useful engagement or mainly eliminates metric-chasing.

Figure 1 shows cumulative first-time use after the original mandate, establishing that the policy did succeed at getting employees into the product. Figure 2 then shows the distribution and within-month timing of queries under the 200-query target. The visual pattern is central to the argument: usage piles up around the requirement rather than varying only with apparent task demand.

The revised target therefore functions as a test of the marginal queries induced by the original rule. If those queries were mostly valuable work, cutting the requirement should materially reduce non-repeated, job-related use. If they were gaming, the lost volume should instead be dominated by repetitive or irrelevant prompts. The estimates favor the latter account.

Results and Analysis

After branches moved from a 200-query to a 100-query monthly target, total query volume fell by 30%. Repeated or off-task queries account for about 90% of that reduction. The estimated change in non-repeated, work-related queries was small and statistically insignificant. This is the paper's most consequential result: lowering the quota removed a large amount of recorded activity without detectable evidence that it displaced the query category the firm most plausibly wants to preserve.

For sales employees, monthly sales increased by 7% after the adjustment. The paper does not establish that every removed query directly caused a sale, but the combination is consistent with a practical mechanism: reducing quota-filling frees attention for work and may make remaining assistant use more purposeful. A 30% decline in headline usage would look like failure in a conventional adoption report. Here, the composition of that decline changes the interpretation.

The evidence supports a narrower conclusion than “less AI use is always better.” It supports revising an excessively high usage mandate in this setting, where the original threshold produced visible bunching and a substantial share of repeated or off-task prompts. Firms adopting internal AI tools should therefore monitor what sits at the margin of a usage requirement, not just whether employees meet it. The relevant operating metric is productive use, which is harder to measure but less vulnerable to being gamed.