The pitch always sounds right. Give AI the easy tier, let your humans focus on the valuable one. I've watched a dozen executives nod along. Six weeks later the aggregate conversion is flat, and nobody in the room can tell you why.
The AI did exactly what it was asked. The boring tier was doing work nobody wrote down.
The call she brought me
Head of inside sales at a B2B SaaS firm. Mid-October. Six weeks after her team rolled out an AI voice qualifier that handles the colder half of inbound.
She brings a dashboard. Aggregate conversion is flat. The ROI model promised a lift. She has already drafted a training plan for her reps.
"They need coaching on the warmer tier," she tells me. "It's a muscle they lost."
I ask her one thing. Pull the data by call-slot order. First three calls of the day, broken out separately.
She comes back the next morning quieter. The first three calls, which are now almost all warm leads, convert worse than they did before the rollout.
Her reps have the training. They are moving through the warm leads faster than they used to. Shorter calls. More hand-offs. The warmth is going and the number is going with it.
What Fu and co-authors just measured
A paper went up on arxiv this week doing exactly this experiment in the field. A Chinese online health-insurance brokerage deployed AI voice agents to handle its low-intent leads. Humans kept the high-intent ones. Over 500,000 outreach records surrounding the rollout. Regression-discontinuity-in-time design across the workday.
The AI did its job. The probability that a human's first three calls of the day were a high-intent lead rose about 23 percentage points. Exactly as designed.
Then the human customer acquisition rate on those improved leads fell about 5 percentage points over the same first three calls, flattening to zero later in the day.
The authors trace it to a worker response. Reps reallocate their earliest leads about 8 percentage points more often. They shorten the calls on them by about 13%.
A decomposition suggests the gain from better lead ordering is smaller than the loss from this behavioural response.
The boring tier was ballast: pacing, context-switching, a warm-up. Take it away and the humans speed through the smaller, denser pool. The focus never arrives.
My coaching client's dashboard aggregated across the day and the slot. The behavioural response stayed invisible until the quarter-end miss showed up.
If you shipped parallel AI into a service workflow and your conversion is "about the same," you have no idea what happened. Pull call-slot order, call duration by slot, and the reallocation rate on your reps' earliest leads. One of those three will tell you.