AI Lead Routing Reduces Human Customer Acquisition

AI deployment in contact centers is usually assessed through throughput, staffing, or the automation rate. This paper studies a different consequence: once an AI voice agent absorbs low-intent prospects, the remaining opportunities offered to human agents change in quality and timing. At a Chinese online health-insurance brokerage, the authors find that this enriched queue can reduce rather than improve human acquisition. The proposed mechanism is sequential search: better expected future opportunities raise workers’ reservation value, making them less willing to persist with a lead already in hand.

Core Contribution

The paper treats the within-day distribution of opportunities as an operational state variable alongside workload and congestion. Its central contribution is not simply that AI handles cold leads. It is that parallel deployment changes the alternatives facing human workers at each moment of the day, and those alternatives alter worker effort.

The setting separates leads into hot and cold classifications. After the expansion, cold leads are directed to AI agents, while human agents focus on hot leads; humans receive cold leads only when no hot lead is available, and hot-lead overflow goes to AI when all human agents are occupied. This creates a changing lead mix without removing human discretion over whether to continue pursuing, reallocate, or end a customer interaction. Figure 1 places the deployment within the operational timeline.

Technical Approach

The authors analyze more than half a million customer-outreach records around the August 28, 2024 expansion of AI voice agents. Their regression-discontinuity-in-time design compares outcomes around that rollout and traces treatment effects by call position within a worker’s day. This call-order structure matters: the model predicts that workers should react most when future high-quality alternatives are plentiful, rather than uniformly across all calls.

Figure 2 depicts the post-expansion allocation rule. It also clarifies why a simple comparison of AI and human conversion would miss the paper’s question. AI changes which leads arrive to humans and when, then workers respond to that altered queue. The empirical analysis therefore separates the direct ordering gain from behavior on the assigned lead.

The paper’s calibrated decomposition formalizes this distinction. Better ordering can improve the match between human agents and high-intent prospects, but a higher reservation value can induce workers to abandon or transfer leads more readily. The relevant operational outcome is the net of those effects, not the quality of the human queue alone.

Results and Analysis

The manipulation appears concentrated at the beginning of the workday. On the first three calls, the probability that a human agent receives a high-intent lead rises by about 23 percentage points after expansion; the effect fades to zero later in the day. Figure 3 shows the first-stage discontinuity in AI’s share of cold leads around the expansion date, supporting the interpretation that the rollout materially changed routing rather than merely coinciding with a gradual shift in demand.

Yet human customer acquisition moves in the opposite direction where the queue improves most. Acquisition falls by about 5 percentage points on the first three calls and then flattens across subsequent calls. The behavioral measures align with the reservation-value account: workers reallocate their earliest leads about 8 percentage points more often, persist less, and make calls about 13% shorter. These patterns are more informative than the aggregate acquisition decline alone because they identify a plausible worker-side channel rather than attributing the result entirely to lead composition.

The decomposition is the paper’s most consequential result. The authors report a modest positive effect from improved lead ordering, but a larger negative behavioral response that more than offsets it. For managers, the implication is narrow but material: reserving high-intent leads for people is not automatically value creating if agents can infer that better alternatives are waiting. Queue design and incentive design need to be considered together.

Limits of the Evidence

The evidence is persuasive for this brokerage’s rollout, but the design remains a single-firm, single-deployment study. The paper observes behavior around a time cutoff, so its interpretation depends on the absence of other discontinuous changes at expansion. It also measures acquisition behavior in a health-insurance sales workflow with a particular hot/cold classification and routing policy; the size, or even direction, of the response may differ where leads cannot be reallocated, compensation is structured differently, or AI handles more complex interactions. The results establish a credible operational mechanism in this setting, not a universal effect of voice-agent deployment.