A Chinese online health-insurance brokerage ran the AI lead routing play every operations deck recommends: put the voice agent on the low-intent leads, leave the high-intent leads to the humans. On the first three calls of a worker's day, the probability that a call was a high-intent lead rose by about 23 percentage points. Customer acquisition on those same calls fell by about 5 percentage points (Fu, Luo, Ramasubbu et al., 2026).
Same people. Better inputs. Worse output. The decline landed precisely where the improvement was largest, and both effects faded to nothing later in the workday โ the detail that rules out most of the comfortable explanations.
That pairing is the finding to sit with. It says the agent did not fail. The agent did exactly what it was deployed to do, and the humans downstream of it changed their behaviour in a way no routing business case prices.
What the study actually measured
The paper is "When Better Opportunities Backfire: Service Worker Responses to Parallel AI Deployment in Customer Acquisition," submitted to arXiv on 19 October 2026 by Xinyu Fu, Shulong Luo, Narayan Ramasubbu and three co-authors. It covers more than a half million customer outreach records surrounding a single AI voice-agent rollout, identified with a regression-discontinuity-in-time design โ comparing worker behaviour immediately before and after the switch rather than against a self-selected control group.
The useful design choice is that the authors trace effects across the workday rather than averaging them. That is what produces the signature:
- High-intent probability on the first three calls: up about 23 percentage points, fading to zero later in the day.
- Customer acquisition on the first three calls: down about 5 percentage points, flattening over the day.
- Workers reallocated their earliest leads about 8 percentage points more often, and shortened those calls by about 13%.
A calibrated decomposition in the paper finds the modest gain from better lead ordering is outweighed by a much larger negative behavioural response. Note the hedging: every figure in the source is stated as "about." Do not let anyone put a decimal on these in your internal deck.
Why AI lead routing moves conversion: reservation value, not skill
Here is where the finding stops being a curiosity and starts being an operating principle.
The authors frame it with sequential search theory. When a worker is holding a lead, the decision to keep working it is a comparison against what else is in the queue. Enrich the queue and you raise the worker's reservation value โ the quality bar the current lead now has to clear to be worth another five minutes. The lead in hand did not get worse. Its competition got better.
So the worker abandons it sooner. Hence the 8-point rise in reallocation and the 13% shorter calls. The behaviour is individually rational and collectively expensive: persistence is where a marginal conversion comes from in outbound sales, and persistence is the first thing an abundant queue taxes.
This is not deskilling. It is not morale in the vague sense that survey decks use. It is a specific, predictable response to a changed opportunity distribution, and the paper's real contribution to service operations is naming opportunity distribution as a state variable alongside the two every ops team already watches: workload and congestion.
Why your dashboard will not show you this
Three instruments will report the routing change as a success:
- Lead quality per human-handled call: up sharply.
- Agent-handled volume: new capacity, no incremental headcount.
- Average handle time: down, because the abandoned calls are short.
Conversion rate is the only one that moves against you, and in most mid-market reporting it is attributed to lead-source mix, seasonality, or the rep. The causal chain from "we routed the easy work away" to "the team stopped grinding the hard calls" crosses two dashboards owned by two different people.
The within-day fade makes this worse. An effect that appears on calls one through three and is gone by mid-afternoon survives almost any daily or weekly aggregation. Average it over a shift and you get a small, noisy conversion dip that looks like variance. Average it over a month and it disappears into the comp cycle. The signal is real and your reporting grain is wrong for it.
This is the same shape as the oversight-cost mistake
The pattern generalises past voice agents, and the adjacent literature is consistent enough to treat as a prior rather than a coincidence.
Production data from Perplexity's agentic product showed matched-task completion time falling from 269 minutes to 36 โ an 87% time and 94% cost reduction โ while the human follow-up work shifted upward into verification, extension, and tasks that crossed occupational boundaries (Perplexity / HBS, 2026). The work did not disappear; it relocated to a part of the org chart nobody had staffed for it.
McKinsey QuantumBlack put numbers on the same displacement from the cost side: for an agent performing a customer-service task in banking, token costs frequently represent just 20 to 25 percent of the variable run costs, while human oversight accounts for 70 to 75 percent (McKinsey QuantumBlack, 2026). Leaders optimise the visible quarter of the bill.
And at the desk level, 42% of knowledge workers report spending more time verifying AI output than they save using it, with 52% regularly correcting AI-generated work produced by colleagues (Adaptavist, 2026). The saving and the cost land on different people, so a per-seat productivity view registers the gain and misses the offset entirely.
Three different research traditions, one structure: agent deployment moves a cost rather than removing it, and the move crosses an accounting boundary. The brokerage study is the version where the cost lands inside human motivation instead of human hours โ which makes it harder to see and cheaper to misattribute.
Where the same shape shows up in work that is not sales
This next part is inference, not finding โ the study covers outbound insurance calls and nothing else. But the conditions that produce the effect are stated clearly enough to check against your own queues, and they are not sales-specific. They are: a worker holds one unit of work at a time, the worker has discretion over when to abandon it, and the worker can see or feel what is waiting behind it.
That describes a surprising amount of mid-market operations. Support agents working a ticket queue they can re-prioritise. Finance staff clearing accounts-payable exceptions. Claims handlers. Credit reviewers. Anyone working a triage list where the easy items are about to be absorbed by an agent and the discretion over sequencing stays with the human.
Where the conditions do not hold, the finding should not worry you. Fixed-sequence work with no abandonment discretion, work where the queue is invisible to the worker, and genuinely non-substitutable work (one case, one owner, to completion) all break the mechanism. The diagnostic question is simply: after this routing change, can the person holding the hard item see something better waiting? If yes, you have the setup. If no, you do not.
The honest counter
I would not publish this as settled. Four limits matter:
It is one firm, one country, one kind of work. Outbound insurance sales is high-volume, commission-shaped, and unusually sensitive to reservation value because the queue is the compensation mechanism. A salaried operations team working a shared service backlog faces a weaker version of the same incentive.
It is a preprint. v1, submitted 19 October 2026, not yet peer reviewed.
RDiT is identification, not a controlled trial. A discontinuity in time is vulnerable to anything else that changed at rollout โ a script revision, a new comp plan, a lead-vendor switch. The within-day fade pattern is the strongest defence against that, since a confounder would have to explain why the effect appears only on the first three calls.
The net is a decomposition, not a P&L. The paper's conclusion that the behavioural response outweighs the ordering gain comes from a calibration exercise. It is a credible estimate of sign and rough magnitude, not an audited margin number.
What survives all four: routing changes the distribution of opportunity, and compensation and metrics built for a scarce queue misprice a rich one. That claim is directionally safe even if the 5 points are not your 5 points.
What to inspect before your next routing decision
The actionable unit is not the model. It is the queue and the seat.
Measure persistence, not just quality. Before the agent goes live, baseline attempts per lead, call duration, and reallocation rate โ split by position in the workday. If you only hold daily averages, the effect in this paper is invisible to you by construction, because it lives in the first three calls and dies by the afternoon.
Re-read the incentive against the new queue. A per-conversion commission on an abundant queue pays people to skim. If the agent is about to raise average lead quality, the comp question is whether anything still rewards the eleventh minute on a hard call. Persistence floors, attempt minimums, or a quality-weighted rather than count-weighted target are the usual repairs.
Decide the occupancy mode of the seat, not the task list. This is where most mid-market routing projects go wrong: they scope which tasks the agent absorbs and never restate what the human seat is now accountable for. An occupancy decision forces you to write down what the human does with an enriched queue; a task-split decision lets you leave it unstated and find out from the conversion rate. Scovai's AUGMENT surface takes that route explicitly โ each seat on the org chart is scored 0โ100% on how augmentable it is, and the mode is set per seat: person alone, person plus a Digital Employee, or the Digital Employee running the role end-to-end under a named human supervisor, with the ceiling set by the company rather than defaulting to full autonomy (Scovai AUGMENT).
Run the agent on a slice first and watch the humans, not the agent. The agent's own metrics will be fine. The instrumentation that earns its keep is on the workers whose queue just changed.
The decision this quarter
An AI lead routing decision is not a capacity decision with a motivation side-effect. It is a motivation decision wearing a capacity costume.
So stop asking which work your agent should take. Ask what your people are paid to do once the work you left them starts to look abundant โ and whether you will see the answer in your numbers before it shows up in your pipeline.