Answer

The definition of a drop doesn't change: an answered predictive call that was connected to nobody - no human agent and no AI bot ever joined. A call the bot handled end-to-end is a connected, answered call, exactly like a human conversation - it is never a drop.

Every predictive call is tagged at dial time as an AI lead or a human lead (by which pool it was dialed from), and that tag buckets the drops:

Value	Description
`answered calls`	All answered predictive calls (human + AI leads)
`human drop rate`	Human-lead drops / answered human-lead calls
`AI drop rate`	AI-lead drops / answered AI-lead calls
`total drop rate`	All drops / all answered calls

Key consequences:

AI-handled calls count in the denominator: a shift where bots successfully answer many calls lowers the total drop rate, it doesn't hide calls from it.
AI failures count as drops: an AI lead that answered but was never connected to a bot or a human is an AI drop and is included in the total drop rate used for pacing and compliance.
A failed bot-to-human transfer counts against the human side: if the bot asked for a transfer and no human connected before the lead hung up, the pacing engine counts it as a human-side drop, so the group slows down when transfers keep failing.
The total drop rate (human+AI together) is what drives the group's predictive ratio, over the group's usual drop-rate window.

On the Realtime page, each group card with active AI shows the drop rate split into Human and AI rows under the combined figure.