When Less Is More: Managing AI Adoption with Adaptive Incentive Design
A · Strong evidence University of Hong Kong economists use full query logs and a staggered branch-level policy change inside one firm to give causal, checkable evidence on what an AI usage quota produces.
Summary based on the full text·Original: about 50 minutes to read
In short
An arXiv working paper (26 September 2026) by Jie Gong, Jiayi Hou, Jin Li, Fei Pu and Xinjue Yao. A Chinese medical-device maker told about 5,000 non-production staff to send at least 200 queries a month to its internal assistant, with fines for missing it. First-time use jumped, and so did gaming: counts bunched at 200 and almost a third of queries were filler. When the target was lowered to 100, branch by branch at different times, volume fell 30%, nearly all of it filler, and sales rose 7%.
Key numbers
- 31 percentshare of queries that were repeated or off-task under the 200-queries-a-month mandate
- 30 percentdrop in monthly query volume after the target was cut from 200 to 100
- about 90 percentshare of that drop made up of repeated or off-task queries
- 7 percentincrease in monthly sales among sales staff after the lower target
- 14.8%staff who had used the AI assistant before the mandate
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Before the mandate only 14.8% of staff had used the tool; 31.2% used it for the first time in the first mandate month and 9.1% more the next.
- Under the 200-query target, among users 12% sent exactly 200 and 95% sent no more than 220. 31% of queries were repeated, personal or off-task, and activity surged near month end.
- After the cut to 100, low-quality queries fell from 31% to 18% and the month-end surge softened.
- Staggered estimates: monthly volume down 29 queries (30% of the mean); about 90% of the decline was repeated or off-task use; the change in non-repeated, work-related queries was small and not significant.
- Sales staff, the only group with individual output data, raised monthly sales by 7%, with flat pre-trends.
- Employees whose use had clustered near 200 reacted most; even pre-mandate users cut low-quality queries, so the quota induced gaming among willing users too.
How they got it
The firm (pseudonym MedTech Corp, about 9,000 staff in 46 branches) logged the full text and timestamp of every query to a DeepSeek-based assistant fine-tuned on internal documents. Logs were linked to personnel and sales records, March-December 2025. Headquarters lowered the target in July 2025 by memo; 19 branches switched at once, others later, 15 kept 200 through December. That staggered adoption across 42 branches supports difference-in-differences and event-study estimates. Query content was classified into work-related, repeated and off-task categories.
Limits
- One firm in China with a specific penalty regime (fine of RMB 200, about 2% of monthly pay).
- Branch timing was not random; it reflected how local HR read a memo, which may correlate with branch traits.
- Performance is observed only for sales staff.
- Query-quality classification is the authors' own scheme.
Why it matters for companies
Usage quotas produce usage, and a large share of it is noise that inflates adoption dashboards. This paper measures the size of that noise and shows that a lower target removed filler without cutting real work, while output rose. Firms tracking prompts or active seats as success should audit query content before setting or raising targets. A mandate can start adoption, but it needs a planned review point and a measure tied to work output.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Work and AI