When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition
A · Strong evidence Operations and economics researchers from Georgia State, Pittsburgh, Renmin, PolyU and Sydney use more than half a million outreach records and a regression-discontinuity design with extensive robustness checks to measure how humans respond when AI takes the easy-to-lose work.
Summary based on the full text·Original: about 2 hours to read
In short
A working paper (September 2026 version, on arXiv October 2026) by Xinyu Fu, Shulong Luo, Narayan Ramasubbu, Peng Shen, Xincheng Wang and Yucheng Wang. A Chinese online health-insurance broker expanded an AI voice agent to take most low-intent ("cold") leads, leaving human agents a richer pool of high-intent ("hot") leads. Human conversion fell exactly where the pool improved most, early in the day. Workers who expected better leads ahead persisted less with the lead in hand, and that response wiped out the expected gain.
Key numbers
- below 25% to above 60%jump in the share of cold leads handled by the AI voice agent at the 28 August 2024 expansion
- 23.4 percentage pointsrise in the chance that a human agent's first three calls of the day are high-intent leads
- 4.5 percentage pointsfall in human customer acquisition on those same first three calls
- 8.1 percentage pointsrise in agents returning their earliest leads to the shared pool (pre-period mean about 25%)
- about thirty timessize of the negative behavioural response relative to the mechanical gain, under the authors' calibration
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- On 28 August 2024 the AI's share of cold leads jumped from about 25% to above 60%; its share of hot leads stayed below 10% in daytime hours.
- The chance that a human-handled call was a hot lead rose 23.4 pp at call positions 1-3, 15.7 at 4-6, and was about zero at the end of the shift.
- Human customer acquisition fell 4.5 pp on the first three calls and 2.5 pp on the next three, was unchanged mid-day, and rose 3.4 pp at the end of the shift.
- Agents returned their earliest leads to the shared queue 8.1 pp more often (a rise of about 32% on a 25% base) and cut early call duration by roughly 13%.
- Net day-level performance effect: close to zero. The behavioural response is about thirty times the mechanical gain from better lead ordering.
- In the calibrated model, restoring the pre-expansion within-day rhythm of hot leads would raise conversions per call by about 13.5% (95% interval 10.5 to 16.6%).
- Slower pace, learning and customer-side reactions were tested and do not fit the pattern.
How they got it
Regression discontinuity in time around the expansion date, using 409,622 human-handled and 157,468 AI-handled outreach records (about 13,000 agent-days) over roughly ±60 days, with a 10-day donut. Outcomes are residualised on agent, calendar-date and hour fixed effects and estimated by within-day call position. The theory is sequential search: better expected future opportunities raise the reservation value of the current one. Robustness covers alternative bandwidths and bins, an event study and an earlier, weaker rollout date.
Limits
- One firm, one country, one product; the routing rule and incentive scheme shape the result.
- One co-author is from Waterdrop Research Institute; the paper does not state the firm's identity or data relationship in the abstract.
- The thirty-times figure depends on a calibrated decomposition, not a direct experiment.
- The window covers about two months after the change, so longer-term adjustment is not observed.
Why it matters for companies
Most business cases for routing work to AI assume the remaining humans keep working as before on a better mix. This study measures the humans left in the system and finds they change how they work, enough to cancel the gain. Contact centres, sales teams and shared-service desks that split queues between AI and staff should watch persistence, hand-back and call-length metrics by time of day, not just averages, and treat the sequencing of the human queue as a design lever alongside volume.
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