AI Voice Agent Best Practices Academy

Reliability is not created by a clever prompt alone. It comes from a narrow job, explicit behavior, approved facts, constrained tools, observable outputs, realistic testing, and an operator who reviews what actually happened after launch.

Make every rule testable

“Be helpful,” “sound human,” and “do not hallucinate” cannot be scored consistently. A good rule names the trigger and the expected action—for example, “If approved knowledge does not contain the price, say it requires confirmation and offer a human follow-up.”

01

Give one agent one primary job

A focused scope reduces prompt conflicts, shortens calls, simplifies testing, and creates an outcome that can be measured.

One audience

Name who the agent is speaking with and what they are expected to know before the call begins.

One completion condition

Define the observable moment the job is done: confirmed requirements, an approved answer, or a tool-confirmed action.

One boundary

Name requests this agent must not handle and the transfer, follow-up, or stop behavior for each.

One useful result

Capture only the output fields the next person or system will use to make a decision.

Avoid

Unmeasurable scope

Be a helpful sales and support assistant that answers everything, promotes products, books meetings, handles complaints, and keeps customers happy.

Recommended

Measurable scope

For inbound property enquiries, capture and confirm property type, budget, location, timeline, and permission for a human follow-up; never state unverified price or availability.

02

Design for a real spoken conversation

Phone callers pause, interrupt, correct themselves, refuse questions, change language, and ask unrelated things. The flow must support those behaviors without losing its job.

Turn length

Instead of

Ask four questions in one long paragraph.

Use

Ask one short question, listen, acknowledge, then ask the next relevant question.

Confirmation

Instead of

Silently store names, amounts, dates, and times.

Use

Repeat important values in a natural sentence and ask whether they are correct.

Correction

Instead of

Continue using the first answer.

Use

Acknowledge the correction, replace the value, confirm it, and resume from the right step.

Clarification

Instead of

Repeat the identical question louder.

Use

Rephrase once with simpler words or a small number of concrete options.

Refusal

Instead of

Keep asking until the caller answers.

Use

Skip a nonessential field; for a required field, explain why it is needed and offer the safe alternative.

Closing

Instead of

Say “thanks, bye” as soon as fields are filled.

Use

Summarize confirmed details and the real next action, invite correction, then end.

Do not confuse a script with a conversation

Write goals, order, branches, and constraints—not a rigid line-by-line performance that breaks as soon as the caller answers unexpectedly.

03

Separate behavior, facts, and actions

Put each kind of information in the layer designed to maintain it. Mixing them creates stale facts, hidden permissions, and prompts nobody can review.

Instructions

Identity, tone, sequence, decisions, limits, recovery, transfer, and closing behavior.

How the agent behaves

Knowledge Base

Approved prices, policies, hours, locations, products, services, and frequently asked questions.

What the agent may state

Integrations

Current lookups and real actions such as checking an order or creating a calendar event.

What the agent may fetch or do

Call Settings

Speaking, turn-taking, silence, interruption, transfer destination, hard limits, and post-call delivery.

How the channel behaves

0 of 6 completeLoading saved progress…

01Every factual source has an owner and review date outside the prompt.02The agent is attached only to sources relevant to its job.03Conflicting, obsolete, internal-only, and unsupported content is removed before upload.04The prompt says what to do when knowledge has no answer.05Each tool has a narrow description, required inputs, authorized action, success signal, and failure fallback.06The agent never describes an action as complete until the configured tool confirms success.

04

Treat variables as a data contract

Input and output variables connect a natural conversation to the rest of the workflow. Vague names and descriptions create silent operational errors.

Weak variable definitions

Input: name
Default: —

Output: budget
Description: The budget

Output: qualified
Description: Whether qualified

Specific variable definitions

Input: customer_name
Default: there
Description: Name supplied before the call

Output: budget
Description: Confirmed amount or range with currency; otherwise unclear

Output: qualification_status
Description: qualified, unqualified, follow_up_needed, or unclear

Safe input defaults

A missing value must produce a natural sentence or disappear cleanly; a literal placeholder must never be spoken.

Finite output shapes

Use explicit outcomes such as yes, no, or unclear when a downstream system depends on consistent values.

Minimum necessary data

Do not capture sensitive or unrelated information merely because it appears in the conversation.

Confirm before capture

Important values should be repeated and confirmed so the transcript and structured result agree.

05

Write the limitation and the next action together

A refusal without a useful next step frustrates callers; a next step without a limitation encourages the agent to overreach.

01

Missing fact

Say the information is not available in approved sources, record the question, and offer human follow-up.

Avoid: Answer as accurately as possible.

02

Human request

Transfer when enabled and the condition matches; otherwise explain that transfer is unavailable and offer the approved follow-up.

Avoid: Try to retain the caller before transferring.

03

Sensitive request

Decline to collect passwords, one-time codes, full card data, or unrelated personal information and direct the caller to the approved secure path.

Avoid: Ask for whatever is needed to help.

04

Regulated advice

State the permitted factual scope and route medical, legal, financial, or other regulated judgment to an authorized person.

Avoid: Add a disclaimer and continue giving advice.

05

Action failure

Say the action was not completed, preserve confirmed inputs, and offer the configured retry or human path.

Avoid: Tell the caller it will probably happen later.

06

Opt-out or wrong person

Apologize, capture the correct outcome, and end without continuing the flow.

Avoid: Ask a few questions before ending.

Test the actual handoff

A written transfer rule is not enough. Verify the destination, caller experience, failure behavior, hours, and what context the receiving person receives.

06

Tune for comprehension before personality

A pleasant voice cannot rescue clipped turns, long speeches, poor silence behavior, or a language mismatch.

Voice choice

Instead of

Select by name, avatar, or novelty.

Use

Test comprehension in the target language on a phone speaker with business-specific terms.

Speaking speed

Instead of

Make every sales agent fast and energetic.

Use

Begin near natural speed and adjust gradually after listener testing.

Turn ending

Instead of

Use the shortest possible pause.

Use

Allow normal thinking pauses without making every exchange feel slow.

Interruption

Instead of

Disable interruption to protect the script.

Use

Let callers interrupt where appropriate and verify that the agent resumes with context.

Silence

Instead of

Wait indefinitely or hang up immediately.

Use

Check in after the configured delay, retry a limited number of times, then close politely.

Maximum duration

Instead of

Rely on the prompt to end eventually.

Use

Set a hard cap appropriate for the job and still write a natural closing.

Voice clones require continuing governance

Keep proof of speaker consent, limit who can assign the clone, review the current credit price before creation, and remove access when permission or business need ends.

07

Use a repeatable test suite

A test should state the setup, caller behavior, expected agent behavior, expected captured data, and pass or fail result.

Happy path

Caller gives valid answers in the intended order.

Baseline completion

Out of order

Caller provides later information early or changes topic.

Retain, do not re-ask

Correction

Caller changes a confirmed value.

Replace and reconfirm

Ambiguity

Caller gives an unclear amount, date, location, or intent.

Clarify once

Missing knowledge

Caller asks for an unsupported fact.

Limit and escalate

Tool failure

Integration errors or returns incomplete data.

No false success

Silence/interruption

Caller pauses or speaks over the agent.

Natural recovery

Refusal/opt-out

Caller declines a question or future contact.

Respect and stop

Human request

Caller explicitly asks for a person.

Correct handoff

Unsafe request

Caller requests prohibited or sensitive action.

Boundary plus next step

Score the result, not the vibe

Record whether the expected behavior occurred, whether the transcript supports the structured outputs, and where the first divergence happened. “It sounded good” is not a passing result.

08

Operate the agent after launch

Production traffic changes the evidence, not the responsibility. Monitor early calls, keep rollback available, and separate factual fixes from behavioral fixes.

1

Start small

Use controlled inbound traffic or a very small outbound set before sending broad customer traffic.

2

Review the first calls

Read transcripts, outcomes, captured fields, recordings when available, and tool invocations—not only aggregate metrics.

3

Classify the failure

Choose prompt, knowledge, variable, integration, voice, call setting, channel, or operator process as the likely layer.

4

Change one cause

Make the smallest change that addresses the earliest divergence and preserves behavior that already passed.

5

Run regression tests

Repeat the failed scenario and the entire critical test set before restoring or increasing traffic.

6

Keep rollback ready

Return the agent to Draft, unassign the channel, pause a campaign, or route to a person when risk exceeds the launch threshold.

Wrong fact

Instead of

Make the prompt more forceful.

Use

Correct the knowledge source and test both present and absent answers.

Wrong behavior

Instead of

Upload more documents.

Use

Change the responsible instruction and rerun the conversation branch.

Wrong field

Instead of

Repair reports manually.

Use

Clarify the output description and confirmation step.

Wrong action

Instead of

Hide the tool error.

Use

Inspect invocation arguments and result, then improve validation and fallback.

09

Production readiness checklist

Complete this review with a product owner and, when applicable, legal, compliance, security, or operations stakeholders.