Product design question

Design a low-latency voice agent experience that feels genuinely human.

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Tests product design for the hardest usability bar in voice AI: making an interaction feel genuinely human, not just technically functional.

Define what breaks the human feeling today: response latency, robotic turn taking, and an inability to handle interruptions or overlapping speech naturally.
Prioritize latency first, since even a few hundred milliseconds of delay is what most reliably signals to a listener that they are talking to a machine.
Design for natural turn taking: allow the agent to detect and gracefully handle interruptions, pausing rather than talking over the user or restarting its sentence.
Add conversational texture: brief acknowledgments, natural pacing, and appropriate pauses, rather than a single unbroken generated response.
Handle uncertainty gracefully: when the agent is unsure, it should ask a clarifying question the way a human would, rather than guessing or stalling silently.
Confirm with the interviewer whether the target use case is customer support, where task completion matters most, or a more open ended companion style interaction, where texture matters more.

Names latency as the single biggest lever, correctly identifying it as the most common tell that breaks the illusion of a human conversation.
Designs explicitly for interruption handling, since real human conversation is full of overlapping speech that most voice agents handle poorly.
Adds natural conversational texture like brief acknowledgments and pacing, not just correct words, recognizing tone and rhythm matter as much as content.
Ties the design to a measurable proxy, like user reported naturalness scores or call length as an indicator of comfort, rather than treating human likeness as unmeasurable.

Focusing only on voice quality or accent while ignoring latency and turn taking, which matter more to the human feeling.
Designing a fixed, unbroken response style that ignores how real conversations involve interruption and overlap.
No way to measure naturalness, treating it as a purely subjective, untestable goal.

How would you handle a user who interrupts mid sentence with an unrelated question?
What latency target would you set as the bar for feeling natural?
How would you test whether a change actually made the agent feel more human?