Why it matters
A strong AI receptionist setup depends on several parts working together: clear call flows, well-written agent instructions, correctly configured tools, a reliable knowledge base, structured data parsing, and properly configured voice and interaction settings.
If one of these areas is weak, the receptionist may sound inconsistent, route calls incorrectly, miss important details, fail to trigger actions, or create poor reporting. When these areas are configured together, your receptionist can handle common calls more reliably and give your team better visibility into what callers need most.
Key Concepts
Call Flow: The sequence the receptionist follows during a call, including greeting, intake, action, confirmation, and close.
Agent Instructions: The main instructions that define how the receptionist behaves, what it says, what it asks, and when it should escalate or use tools.
Begin Message: A separate field outside the main instructions that defines the exact opening line used at the start of a call.
Capability Scope: The set of call types the receptionist should handle directly, route, escalate, or capture for follow-up.
Caller Intake: The process of identifying the reason for the call and collecting the information needed to move the request forward.
Tools: Agent actions configured in Whippy, such as sending SMS, call transfer tools, triggering automations, logging outcomes, or routing to a queue.
Knowledge Base: The approved source of company information the receptionist should use when answering factual or company-specific questions.
Data Parsing: Structured fields extracted from calls, such as intent, outcome, issue type, caller details, or escalation reason.
Custom Reports: Reports built in Whippy from parsed call data so teams can measure patterns such as FAQ volume, payroll inquiries, escalations, and follow-up needs.
Voice Settings: Configuration that controls how the receptionist sounds, including voice choice, pacing, consistency, and expressiveness.
Interaction Settings: Configuration that controls responsiveness, interruption behavior, backchanneling, reminders, voicemail handling, and duration limits.
Language Mode: A setting that controls whether the receptionist uses a single language or supports multiple languages.
Speech-to-Text Mode: A setting that balances transcription speed and accuracy when interpreting caller audio.
Guardrails: Explicit constraints that prevent incorrect or unsafe behavior, such as never inventing information or always confirming critical details.
Fallback Logic: Rules for handling unclear requests, failed actions, missing information, silence, or confusion.
Escalation Rules: Conditions that define when the receptionist should transfer to a human, route to a team, or stop attempting resolution.
Step-by-Step: Build a Best-Practice AI Receptionist
Define scope and call handling goals.
Decide what the receptionist should handle, what it should route, what it should escalate, and what it should capture for follow-up.
Choose the main call types it will support, such as company FAQs, payroll questions, applicant support, current worker support, HR requests, client calls, or urgent issues.
Decide which call types should be fully resolved by the receptionist and which should always be routed.
Define the outcome you want for each workflow, such as resolved, transferred, escalated, logged, or follow-up required.
Set the call flow and core agent instructions.
Build the foundation for how the receptionist should behave on every call.
Set the Begin Message in the dedicated field. Do not place it inside the main instructions.
Write structured instructions with sections for identity, objectives, primary workflow, intent handling, guardrails, fallback logic, escalation rules, and call ending rules.
Use clear call stages: greet, identify intent, collect information, take action, confirm next step, and close.
Define caller intake questions for each workflow so the receptionist consistently collects the right details.
Here is a very high level overview of a basic call flow
Configure tools and knowledge, and reference them in agent instructions.
Make sure the receptionist can take actions and knows when to use them.
Enable the tools needed for your workflows, such as call transfer tools, SMS tools, routing tools, and automation triggers.
For each tool, define usage directly in the agent instructions: when to use it, what information is required, and whether confirmation is required before execution.
Do not rely on tool configuration alone. The agent must be explicitly instructed to use tools in the correct scenarios.
Connect a knowledge base for approved company information.
Add a Knowledge Base Usage section in the agent instructions that tells the receptionist when to use it, what it contains, and that it must not guess if information is missing.
Ensure instructions clearly tie workflows to tools and knowledge usage. (For example: use knowledge base for FAQs, use call transfer tool for escalation.)
Set up data parsing and reporting.
Define what information should be extracted from calls so you can measure outcomes in Whippy.
Add parsed fields for the categories you want to report on, such as
intent,outcome,issue_type,escalation_reason,lead_name, orlead_email.Use enums where possible for cleaner reporting and more consistent values.
Keep parsed fields aligned with your actual workflows and reporting needs.
Build custom reports in Whippy from parsed data so you can track questions such as how many calls were company FAQs, payroll questions, or escalation cases.
Configure voice and interaction settings.
Adjust how the receptionist sounds and behaves during live calls.
Choose the right language mode and primary language. Use multi-language only when needed.
Set speech-to-text mode based on speed or accuracy needs, and enable denoising if callers are often in noisy environments.
Select a voice that fits the use case, then tune speed, stability, and expressiveness so the receptionist sounds natural and consistent.
Adjust responsiveness, interruption sensitivity, and backchanneling to match the desired call style.
Configure reminders, voicemail behavior, duration limits, boosted keywords, and pronunciation settings where needed.
Test, review, and improve.
Validate the full setup before launch and continue refining it over time.
Test in agent chat to review logic, workflows, knowledge use, and instruction quality.
Test in preview call to hear pacing, tone, interruption handling, and closure behavior.
Test common call types, edge cases, and failure cases such as silence, confusion, tool failure, or urgent escalation.
Review parsed data and reporting output after testing to make sure the right information is being captured.
Update instructions, tools, parsing fields, and voice settings based on real test results.
Tips and Best Practices
Start with structure before configuration. Define call flows and write clear agent instructions before enabling tools or advanced features.
Keep instructions and configuration aligned. If a tool or knowledge base is enabled, it must also be clearly referenced in the agent instructions.
Use explicit tool instructions. Always define when to use a tool, what inputs are required, and when to confirm before execution. Do not assume the agent will infer this.
Name tools clearly in instructions. Reference tools exactly as configured, such as “call transfer tool” instead of generic phrasing.
Tie tools to workflows. Each major call type should clearly map to a tool action, resolution, or escalation path.
Add a dedicated knowledge usage section. Tell the agent to use the knowledge base for all factual company questions and to never guess.
Keep call flows simple and operational. Start with a narrow scope and expand after testing.
Break actions into atomic steps. Clear, single-step instructions reduce skipped steps and inconsistent behavior.
Standardize caller intake. Define what information must be collected before any action is taken.
Separate resolve, route, and escalate logic. Be explicit about what happens in each scenario.
Design for reporting. Configure data parsing fields based on what you want to measure in Whippy, such as payroll calls, FAQs, or escalations.
Use enums for consistency. Standardized values improve reporting and filtering.
Tune voice settings to match the use case. Use calm, steady voices for support and adjust expressiveness carefully.
Keep voice pacing natural. Avoid speeds that sound rushed or robotic.
Adjust responsiveness and interruption intentionally. Match these settings to the type of conversations you expect.
Use backchanneling sparingly. Low frequency works best for most professional calls.
Configure voicemail and reminders intentionally. Avoid overuse and ensure behavior matches your workflows.
Use boosted keywords and pronunciation where needed. This improves accuracy for company-specific terms.
Test in both preview call and agent chat. Each mode reveals different issues.
Iterate based on real data. Use call outcomes and custom reports to improve over time.
Troubleshooting
Issue | Possible Cause | Fix |
The receptionist handles too many call types | Scope is too broad in agent instructions | Narrow responsibilities, simplify workflows, and route more calls |
Calls are routed incorrectly | Routing logic is unclear or incomplete | Update call flows, intake questions, and escalation rules |
The receptionist gives inconsistent answers | Instructions are unstructured or conflicting | Rewrite instructions into clear sections with atomic steps and explicit rules |
The receptionist hallucinates or guesses | Knowledge base rules are weak or missing | Add strict knowledge usage rules and state that the receptionist must never invent information |
The receptionist cannot take an action | Required tool is not enabled or not defined clearly | Enable the tool and define exactly when it should be used in both configuration and instructions |
Tools trigger at the wrong time | Trigger conditions or required inputs are unclear | Add explicit tool rules and confirmation requirements in the agent instructions |
The receptionist collects incomplete information | Caller intake questions are not standardized | Define required intake fields for each workflow |
SMS, follow-up, or automations do not run | Tool configuration is missing or not referenced in instructions | Review tool setup and update instructions to reference the tool properly |
Parsed data is missing from reports | Data parsing fields are not configured clearly | Add or revise parsed fields, descriptions, required fields, and enum values |
Reports are hard to use | Parsed data structure is too broad or inconsistent | Simplify fields and standardize values for easier filtering and grouping |
The receptionist misunderstands caller speech | Speech-to-text is optimized for speed or denoising is off | Switch to a more accurate transcription mode and enable denoising if needed |
The voice sounds unnatural or robotic | Voice speed or consistency settings are not tuned | Adjust speed and stability for more natural delivery |
The receptionist interrupts callers too often | Interruption sensitivity is too high | Lower interruption sensitivity |
The receptionist reacts too slowly | Responsiveness is too low | Increase responsiveness |
Voicemail behavior is incorrect | Voicemail action is not configured properly | Set the correct voicemail action and review the message setup |
Calls end awkwardly or without resolution | Call ending rules are missing | Add explicit confirmation and close steps |
The receptionist gets stuck on unclear requests | Fallback logic is missing | Add clarifying questions, escalation paths, and failure handling |
Testing looks good in chat but not on calls | Voice and interaction settings were not validated in voice mode | Test in preview call and refine pacing, interruption, and responsiveness settings |