Dialogue scenes on Avowed and The Outer Worlds 2 play back through Sequencer, and a studio tool generates them: the characters, the timing, the shell of the scene. What it cannot do is make the performance read. That part was hand work, and all of it happened inside the Unreal Editor.
Conversations are authored with an in-house Obsidian tool and generated into Sequencer, which is where the whole scene lives - the characters, the shots and every animation on them. Nothing is assembled in a DCC and imported; the sequence in the editor is the deliverable, so the editing happens there too.
The body pass is assembly rather than authoring. The project already holds a library of animation, and the job is choosing from it and stacking pieces in the sequence - closer to building with blocks than to keyframing - until the motion reads and, more importantly, until it agrees with what the character is saying.
That last part is the judgement the pass lives or dies on. A gesture that looks fine on its own can undercut the intent of a line, and the same clip can land or fail depending on what it follows. Working directly in Sequencer is what makes that judgement cheap: every change is watched against the real dialogue at the real timing, immediately.
Facial animation starts from a tool that generates lip sync from the dialogue. That gets the mouth roughly onto the words, which removes a large amount of tedium, but a generated track is only phonetically right - it does not know which word carries the line.
The polish is done by hand in the curve table: correcting shapes that read wrong and pulling the performance toward the delivery rather than the phonemes. It is the same relationship as the body pass - the tool covers the volume, the hand pass makes it act.