Designing Voice Agents That Ask Better Questions

A voice agent that asks bad questions wastes the caller's time and produces bad data. Good questions feel natural and capture what you need in fewer turns. The skill of asking the right question at the right moment is what separates polished agents from clumsy ones — and it's almost entirely a matter of prompt design.

TL;DR

• The best voice agent questions are short, single-purpose, and front-load the most important info.
• Don't ask for info you can infer or look up.
• Don't ask multiple questions at once.
• Confirm-back on critical info; don't confirm trivial info.

What makes a good voice question

Five properties:

Short. One clause; under 10 words ideally.

Single-purpose. One piece of info per question. Don't ask "what's your name and account number?" — ask each separately.

Open or closed appropriately. Open questions ("how can I help?") for discovery. Closed questions ("can you confirm Tuesday at 3?") for confirmation.

Front-loaded. The most important word should come early. "Date — what date works for you?" beats "Could you possibly tell me what date might work for you?"

Naturally phrased. Sounds like a person, not a form. "What time?" beats "Please specify the time of your preferred appointment."

Common bad question patterns

A few that show up in production voice agents:

Compound questions. "Can you confirm your name, the date you'd like, and the type of appointment?" Caller forgets one; agent has to re-ask.

Buried instructions. "I need to confirm a few details before I can book. What's the name on the account, your phone number, and the appointment time?" By the time the caller is parsing the question, they've forgotten what was asked.

Over-confirmed trivia. "So your name is Sarah, with an S? OK, and that's S-A-R-A-H, correct?" Annoying.

Under-confirmed critical info. "Got it, see you Tuesday." When you didn't actually confirm Tuesday meant next Tuesday.

Form-style language. "Please provide your phone number." Sounds robotic.

When to ask vs when to look up

Default: don't ask if you can look up.

If the caller's phone number is the caller ID, don't ask for their phone number. If their account is identifiable from the phone number, don't ask for their account number. Look up first; confirm.

Pattern:

1. Use whatever signal you have (caller ID, prior call summary).
2. Ask one identifying question if needed.
3. Confirm what you found.
4. Move to the actual goal.

This cuts 1–2 turns off most calls.

When to ask vs infer

Sometimes you can infer what the caller wants from context. Default: don't infer; ask.

The agent should infer when:

• The intent is overwhelmingly clear (caller said "I want to book").
• The inference is reversible (the agent can correct course quickly).

The agent should ask when:

• Multiple reasonable intents exist.
• The action is consequential (booking, charging, cancelling).
• The cost of inferring wrong is high.

Confirm-back rules

Confirm critical info; skip trivial confirms.

Confirm:

• Appointment date and time.
• Phone numbers, account numbers, addresses.
• Any commitment ("you want me to charge $50?").
• Names if they'll be used in writing.

Don't confirm:

• That the caller heard a brief acknowledgment.
• Trivia ("you said you have a question — yes?").
• Anything obvious from context.

The confirm-back rule: "Read it back when getting it wrong would matter."

Patterns for specific contexts

Capturing a date. Ask "what day?" first. Get a day. Then ask "what time?" Get a time. Confirm both together. Don't ask for both in one question.

Capturing a name. "And your first name?" Get it. "Last name?" Get it. Confirm pronunciation if uncommon. Don't ask the caller to spell unless the audio was unclear.

Capturing a phone number. Read it back digit by digit. "That's six-five-zero, five-five-five, zero-one-nine-nine, correct?"

Diagnosing an issue. Start broad ("what's happening?"), narrow with each turn. Don't ask 10 diagnostic questions in a row — get the lay of the land first.

How to test question quality

Listen to 20 calls. For each, look for:

• Questions the caller had to clarify ("you mean what time today?")
• Questions the agent could have skipped (info already known)
• Compound questions where the caller answered partially
• Confirms that felt unnecessary

Each is a signal to tighten the prompt.

What the prompt should say

Some specific rules to include:

Ask one question at a time. Never compound multiple questions
into one.

If you can look up info via a function, do that before asking
the caller. For example, if you have the caller's phone number
from caller ID, don't ask for it again — call lookup_caller
first.

Confirm dates and times by reading them back in natural
language ("so that's Tuesday the fifteenth at three PM,
correct?"). Don't confirm trivial details.

Use short questions. Under 10 words. Front-load the key word.
"Date?" beats "What date would you like to book for?"

For more, see designing system prompts for multi-turn voice conversations.

Related reading

• How Large Language Models Power Voice Agents
• Open-Source vs Closed-Source LLMs for Voice Agents
• How LLMs Decide What to Say Next in a Voice Conversation
• Why Context Windows Matter Less Than You Think for Voice
• Multi-Agent Architectures for Customer Service

FAQ

Should the agent always ask one question at a time? For voice, yes. For chat, multi-question is more tolerable.

What about open-ended questions like "how can I help?" Use them as openers; don't repeat them. Once the caller stated their issue, narrow with closed questions.

How do I know if my agent is asking too many questions? Average call length. If your agent's calls are 8+ turns and a human handles the same case in 3, you're over-asking.

Should I script every question? Define the patterns; let the model phrase naturally. Pure scripts feel robotic.

What about cultural differences in question style? Real consideration for multilingual deployments. Direct vs indirect questioning conventions differ. Adjust prompt per language.