Voice Agents - UnleashX

1. Introduction

1.1 What is an UnleashX Voice Agent?

Voice Agent talks to your customers on the phone, just like a human agent would. It listens, understands intent, responds naturally, and can take real actions during the call: looking up account details in your CRM, scheduling appointments, capturing structured information, sending emails or SMS, and transferring to a human when the situation calls for it. Voice agents on UnleashX are designed to handle the high-volume, repetitive calls that take up most of your team’s time, freeing your human agents to focus on the conversations where empathy and judgment really matter.

1.2 Who Is This For?

This documentation is written for product managers, contact center leaders, operations teams, and anyone responsible for designing, deploying, or maintaining voice agents on UnleashX. Some sections (Tools, API triggers) reference technical concepts, but no prior coding experience is required to build a working agent.

1.3 What You Can Do With UnleashX Voice Agents

Outbound Calling at Scale: Run lead outreach, renewal reminders, payment follow-ups, and customer surveys across thousands of contacts without growing your team. Use UnleashX Campaigns to handle batch dialing, pacing, and retries.
Inbound Support and Self-Service: Let customers call in for FAQs, order status, account help, and appointment booking, 24 hours a day, in any language you support. Inbound agents are always on, so customer queries get resolved 24*7.
Workflow-Triggered Calls: Automatically reach customers when something important happens: a new lead, a missed payment, a delivery delay, an abandoned cart. UnleashX Workflows let you drop the agent into any event-driven flow.
Conversation Intelligence: Every call is automatically transcribed, summarized, scored for sentiment, and tagged with dispositions and structured outputs that flow straight into your CRM, data warehouse, or analytics tool.

Key IdeaA voice agent is not a script reader. It’s a conversational AI that understands context, handles unexpected questions, and adapts in real time. The configuration steps in this guide are how you give it the right knowledge, voice, tools, and guardrails to do its job well on UnleashX.

2. Core Concepts

2.1 The Anatomy of a Voice Agent

Every voice agent on the platform is made up of six core components. Understanding what each one does will help you make better configuration decisions later.

The Brain (Prompt + LLM): The instructions you write are combined with a large language model that decides what the agent says next.
The Voice (TTS): The text-to-speech engine that turns the agent’s responses into spoken audio.
The Ears (Transcriber / STT): The speech-to-text engine that converts what the caller says into text the agent can understand.
The Knowledge Base: The documents, FAQs, and policies the agent draws on to answer factual questions accurately.
The Tools: External actions the agent can take during a call - CRM lookup, booking a meeting, sending an email, triggering a webhook.
The Guardrails: The trust and safety controls that keep the agent compliant, accurate, and respectful with DNC detection, restricted topics, PII redaction, and more.

2.2 Inbound vs Outbound Agents

Voice agents come in two flavors based on who initiates the call:

Type	Description
Inbound	The customer dials in. The agent waits for the call, picks it up, and handles the conversation. Best for support, FAQs, account help, order status, appointment booking, and self-service workflows.
Outbound	The agent makes the call. You feed it a contact list (via Campaign) or trigger it from a workflow event. Best for lead outreach, reminders, renewals, surveys, collections, and re-engagement.

2.3 Single-Prompt vs Multi-Prompt Conversation Flow

One of the most important decisions you’ll make when building an agent on UnleashX is how to structure its conversation logic. The platform supports two approaches, and the right choice depends on the complexity of your use case.

Single-Prompt Agent

You write one comprehensive prompt that defines everything: the agent’s role, goals, tone, conversation style, fallback handling, and instructions for every situation. The agent uses that prompt for the entire call and decides what to say next based on context. Best for: Short, focused calls where the agent has one clear job and the conversation doesn’t really branch. Quick to ship, easy to iterate. Real examples on UnleashX:

Order status and delivery tracking calls
Appointment reminders and confirmations
Quick post-purchase feedback surveys
FAQ bot answering questions from a product knowledge base
Outbound notifications (policy renewals, EMI reminders)

Why pick this: Fastest path from idea to live agent. Just one prompt to write, debug, and tune. Lower latency and lower cost per call. Single-Prompt Structure A well-written single prompt on UnleashX typically follows a six-section structure. Use this as a template when writing your own:

## Identity & Persona
You are Riya, a friendly customer service representative
calling on behalf of Acme Insurance. You are warm,
professional, and respectful of the customer's time.

## Goal
Remind the customer that their motor insurance policy
is due for renewal in 7 days, confirm whether they
want to renew, and capture their preferred renewal date.

## Conversation Flow
1. Greet the customer by name and introduce yourself
2. Confirm you are speaking to {customer_name}
3. Mention the policy due date and amount
4. Ask if they want to renew
5. If yes, capture their preferred renewal date
6. Confirm next steps and end the call politely

## Tone & Style
- Speak in short, natural sentences
- Use the customer's first name occasionally
- Never sound pushy or robotic
- Match the customer's energy

## Handling Edge Cases
- If the customer is busy, offer to call back
- If they decline, ask the reason briefly and thank them
- If they have questions about coverage, direct them
  to the knowledge base
- If they ask for a human, transfer to support

## Closing
Always thank the customer for their time and confirm
the next action before ending the call.

Each section serves a purpose: Identity tells the agent who it is, Goal tells it why the call is happening, Conversation Flow gives the high-level structure, Tone shapes how it talks, Edge Cases handle the unexpected, and Closing ensures consistent endings.

Multi-Prompt Flow

You break the call into states — for example, ‘greet’, ‘identify caller’, ‘collect details’, ‘resolve issue’, ‘confirm and close’ — and write a focused prompt for each one. The agent moves between states as the conversation unfolds, carrying information forward through dynamic variables. Best for: Multi-step calls with branches, conditional logic, or distinct phases where the agent needs to behave differently depending on what the customer says. Real examples on UnleashX:

AI receptionist that books, reschedules, or cancels appointments
IT helpdesk that identifies the employee and issue, then walks through fixes
Insurance claim intake with verification, documentation, and assessment steps
Loan eligibility check with KYC, income verification, and document collection
Inbound support that triages, troubleshoots, and escalates to a human
Hospital appointment booking with department routing and slot confirmation

Why pick this: The agent stays on track. No drifting off-topic, no skipped steps, no repeating questions it already asked. Easier to debug and maintain at scale. Multi-Prompt Structure A multi-prompt agent on UnleashX is built as a tree of states (also called nodes). Each state has its own focused prompt, a clear goal, and transition rules that decide which state comes next based on what the customer says. Here’s how an AI receptionist for a dental clinic might be structured:

STATE 1: Greeting & Intent
Prompt: "Greet the caller, introduce yourself as the
Bright Smile Dental assistant, and ask how you can help.
Identify whether they want to book, reschedule, cancel,
or ask a question."
Transitions:
  → Booking flow (if intent = book)
  → Reschedule flow (if intent = reschedule)
  → Cancel flow (if intent = cancel)
  → FAQ flow (if intent = question)

STATE 2A: Booking Flow
Prompt: "Ask for the patient's full name, preferred
date, and reason for visit. Check available slots in
the calendar tool. Offer the closest 3 slots."
Transitions:
  → Confirmation (when slot selected)
  → Reschedule flow (if slots don't work)

STATE 2B: Reschedule Flow
Prompt: "Ask for the patient's name and current
appointment date. Look up the booking. Confirm the
new preferred slot."
Transitions:
  → Confirmation

STATE 2C: Cancel Flow
Prompt: "Confirm the patient's name and appointment
date. Ask for the cancellation reason. Update the
calendar."
Transitions:
  → Closing

STATE 2D: FAQ Flow
Prompt: "Answer questions using the clinic knowledge
base. If the question is outside scope, offer to
transfer to a human."
Transitions:
  → Closing
  → Transfer to human

STATE 3: Confirmation
Prompt: "Repeat back the appointment details, confirm
the patient's contact number, and let them know they'll
receive an SMS confirmation."
Transitions:
  → Closing

STATE 4: Closing
Prompt: "Thank the patient warmly, remind them of any
next steps, and end the call."

Notice how each state has a single, focused job. The Booking Flow only worries about booking - it doesn’t need to know how to cancel. This separation is what makes Multi-Prompt agents reliable: each state is small enough to debug and iterate independently. Information captured in earlier states (like patient name and contact number) is automatically passed forward as dynamic variables, so the agent never has to ask the same question twice. Common Multi-Prompt Patterns Most Multi-Prompt agents on UnleashX follow one of these proven patterns:

Linear Flow — States happen in a fixed order: Greet → Verify → Collect → Confirm → Close. Used for forms, surveys, and intake calls where the sequence doesn’t change.
Branching Tree — An initial state routes to one of several specialized flows based on intent. Used for receptionist agents, support triage, and any call where the customer’s first answer determines what happens next.
Loop with Exit — The agent loops through a state (e.g., ‘answer questions’) until a condition is met (customer says they’re done). Used for open-ended Q&A and consultation calls.
Verification Gate — Mandatory verification states must pass before the agent moves into sensitive flows. Used for KYC, account access, and high-risk actions.

When to switch from Single to Multi-Prompt Start with Single-Prompt — it’s the fastest way to get something live on UnleashX. Move to Multi-Prompt the moment any of these signals appear: your prompt grows past ~1000 words, the agent uses more than 5 tools, the agent keeps skipping steps or repeating itself, or the conversation naturally splits into distinct phases that need different behavior.

3. Building Your First Agent

Building a voice agent on the platform is a six-step process. You can move forward and backward freely between steps — nothing is final until you click Publish.

Step 1: Agent Setup & Brain

The first step is where your agent gets its identity and its instructions. Everything you configure here shapes how the agent thinks and behaves on every call. About Give your agent a name, a short description, and a category. This is how you’ll find it later in lists, and how teammates understand what it does at a glance. A good name is specific to the use case, like ‘Q2 Outreach Agent’ or ‘Insurance Renewal Bot’ — not just ‘Agent 1’. Language Pick the primary language your customers speak. This decision is important because it determines which transcribers and voices you can choose in Step 2 — not all engines support every language with the same quality. If you serve multiple languages, you’ll typically build a separate agent for each one and route calls based on the customer’s preferred language. Purpose Tell the platform what the agent is for: Sales, Support, Reminders, Surveys, Collections, Onboarding, and so on. We use this to suggest the right prompt templates, recommended dispositions, and best practices specific to that use case. You can always change it later, but starting with the right purpose saves time in later steps. Inbound or Outbound Choose whether this agent receives calls (inbound) or makes them (outbound). This isn’t just a label, it changes the configuration options that show up in later steps and the triggers available in Step 6. Prompt & Templates This is the most important configuration in the entire setup. The prompt is the agent’s brain, it tells the agent what to say, how to say it, and what to do in tricky situations. Don’t start from a blank page. The platform ships with industry-specific templates for common use cases like outbound sales, insurance renewals, e-commerce support, hospitality bookings, and automotive service reminders. Pick the closest template, then customize the parts that matter for your business. This is also where you choose between Single-Prompt and Multi-Prompt Flow. Refer back to Section 2.3 if you’re not sure which to use.

Best Practice: Writing Effective PromptsStart your prompt with the agent’s identity and tone (e.g., ‘You are Riya, a friendly customer service representative for Acme Corp’). Then specify the goal of the call, the conversation flow, and what to do when something unexpected happens. Avoid vague instructions like ‘be helpful’ instead, give specific examples of what the agent should and shouldn’t say.

Input Variables If you want the agent to call each customer by name, reference their order ID, or personalize the conversation in any way, you’ll need input variables. Define placeholders like {customer_name}, {order_id}, or {appointment_date} once in your prompt, then pass real values at call time. The agent fills them in automatically i.e. same agent, personalized for every call. Knowledge Base Drop in your product docs, FAQs, policies, pricing sheets, or any reference material the agent might need. The platform supports two ways to add knowledge:

Document Upload — Upload files directly: PDF, DOCX, TXT, or Markdown. Best for static content like policy documents, product catalogs, and internal SOPs.
URL Extraction — Paste a website URL and we’ll automatically extract and index the content. Best for documentation that’s already published online and changes regularly.

With a knowledge base attached, the agent answers factual questions using grounded information instead of guessing or making things up. This is one of the most effective ways to keep the agent accurate and trustworthy. Tools Tools let the agent do things during the call, not just talk. The platform supports several tool types out of the box:

CRM Lookup — Fetch customer details, account history, or order information from your CRM in real time.
Send Email — Trigger a transactional email mid-call (confirmation, follow-up, document).
Calendar Booking — Check availability and book meetings directly from the call into your team’s calendar.
Send SMS — Send a text message to the caller during or after the conversation.
Custom Webhook — Trigger any HTTP endpoint with call data to integrate with your own systems.
API Action — Call any REST API as an action — useful for custom integrations and lookups.

Start with the minimum tools you actually need. Every additional tool adds complexity to the prompt and increases the chance of the agent calling the wrong tool at the wrong time. Add tools incrementally as you validate behavior.

Step 2: Voice & Transcription

Step 2 is about how the agent sounds and how it hears. These three settings have a direct impact on the customer experience, pick them carefully. Transcriber (Speech-to-Text) This is how the agent hears the caller. The transcriber converts spoken audio into text in real time, which is then passed to the LLM. Accuracy here is non-negotiable: if the agent mishears the caller, the entire response will be wrong. Pick a transcriber optimized for your language and the typical accents of your customers. The platform supports multiple engines, and the available options will be filtered based on the language you chose in Step 1. Voice (Text-to-Speech) How your agent sounds. Pick a voice that matches your brand and use case:

Warm and friendly for customer support and care
Confident and crisp for outbound sales and lead qualification
Calm and clear for healthcare, insurance, and financial services
Professional and neutral for B2B and enterprise calls

Always preview voices before deciding. The platform lets you play a sample greeting in each voice so you can hear how it actually sounds with your script. LLM Model The brain behind the agent. Different models offer different trade-offs:

Faster Models — Lower latency means more natural, conversational pacing. Better for high-volume use cases where call duration matters and conversations are relatively simple.
Smarter Models — Better at handling complex questions, long conversations, multi-step reasoning, and edge cases. Better for sales, support, and any use case where conversation quality matters more than raw speed.

Most teams start with a balanced model and switch only if they see specific issues like too much latency or too many comprehension failures.

Step 3: Phone Number

Your agent needs a phone number either to receive calls (inbound) or to make calls (outbound). The platform gives you three options. Buy a Number Need a number fast? Buy one directly from the platform. Pick a country, area code, and number type (local, toll-free, or mobile) and you’re ready to go in minutes. This is the simplest option for new deployments. Bring Your Own Number Already have phone numbers with Twilio, Plivo, Exotel, or any SIP provider? Connect them via SIP trunk and keep the numbers your customers already know. This is the right choice for established businesses with existing telephony relationships. Number Assignment Once you have a number, assign it to the agent:

For inbound agents, this is the number customers will dial to reach the agent.
For outbound agents, this is the caller ID customers will see when the agent calls them.

You can assign multiple numbers to the same agent (useful for regional caller IDs in outbound campaigns), or one number to handle multiple agents based on routing rules.

Compliance NoteDifferent regions have different rules about caller ID, number type, and disclosure requirements. If you’re calling customers across countries, work with your compliance team to ensure the right number type and consent handling for each region.

Step 4: Outputs & Success Metrics

This step is where you define what the agent should capture from each call and how to measure success. It’s also one of the most overlooked steps — but it’s where the long-term value comes from. Structured Outputs Tell the agent exactly what information to extract from every call. Each output is defined as a key-value pair with a type:

Key	Type	What it captures
`customer_name`	string	The customer’s name as stated during the call
`customer_email`	email	The email address the customer wants to be contacted at
`appointment_date`	string	The date and time of any booked or rescheduled appointment
`demo_scheduled`	boolean	Whether the customer agreed to schedule a product demo
`satisfaction_score`	number	The customer satisfaction rating captured at end of call

Once defined, the agent extracts these values automatically from every call and pushes them to your CRM, analytics tool, or data warehouse. This is how voice agents become a structured data source instead of just a black box of audio recordings. Dispositions Dispositions are the labels the agent assigns to each call’s outcome. Think of them like the tags a human agent would put on a CRM record after wrapping up a call: Interested, Not Interested, Callback Requested, Demo Booked, Issue Resolved, Voicemail Left, Wrong Number. Define the dispositions that matter to your business. The agent will automatically tag every call with one based on what happened during the conversation. Dispositions then power your reporting, your campaign retry logic, and your overall agent performance metrics. Disposition Templates Don’t reinvent the wheel. UnleashX ships with disposition templates for common industries. Pick the closest one and customize the labels. Each template comes with recommended success markers.

BFSI / Insurance Renewal Agent

Renewal Completed — Customer agreed and the renewal was processed end-to-end on the call. (Success)
Renewal Promised — Customer agreed to renew but wants to pay later or via a different channel. (Success)
Callback Requested — Customer is busy but asked to be called back at a specific time.
Not Interested — Customer explicitly declined the renewal.
Already Renewed — Customer has already completed the renewal through another channel.
Switched Provider — Customer has moved to a competitor.
Documents Pending — Customer wants to renew but needs to submit additional documents first.
Wrong Number — The number does not belong to the intended customer.
Do Not Call — Customer requested no further calls (DNC).
Unreachable — Customer did not answer after maximum retry attempts.

E-Commerce Inbound Support Agent

Issue Resolved — Customer’s question was answered or problem fixed on the call. (Success)
Order Tracked — Customer received the order status they were calling about. (Success)
Refund Initiated — A refund was processed during the call. (Success)
Ticket Created — Issue is too complex for self-service; ticket created and assigned.
Escalated to Human — Call was transferred to a human agent.
Information Requested — Customer asked for product or policy information.
Complaint Logged — Customer raised a complaint that needs offline follow-up.
Repeat Caller — Customer called about an existing open ticket.

Healthcare Appointment Booking Agent

Appointment Booked — New appointment confirmed in the calendar. (Success)
Appointment Rescheduled — Existing appointment moved to a new date or time. (Success)
Appointment Cancelled — Existing appointment was cancelled at customer’s request.
Slot Unavailable — No suitable slots were available for the customer’s preference.
Wrong Department — Customer needs a specialty not handled by this agent.
Emergency Routed — Urgent medical query was transferred immediately to a human.
Follow-up Reminder — Customer was reminded of an upcoming visit.

Lending / Loan Outreach Agent

Qualified Lead — Customer met all eligibility criteria and showed interest. (Success)
Application Started — Customer began the loan application during the call. (Success)
Documents Pending — Customer is interested but needs to submit KYC documents.
Not Eligible — Customer did not meet basic eligibility criteria.
Not Interested — Customer is not currently looking for a loan.
Income Verification Failed — Customer’s stated income did not match required thresholds.
Callback Requested — Customer asked to be reached at another time.

These are starting points. Customize the labels, add your own, and remove what doesn’t apply to your business. The goal is to capture every meaningful outcome of a call so you can report on agent performance accurately. Agent Success Benchmark Out of all the dispositions you define, which ones actually count as a win? Mark the successful ones for example, Renewal Completed, Issue Resolved, Appointment Booked, Qualified Lead and UnleashX will calculate your Agent Success Benchmark as the percentage of calls that hit one of these outcomes. This is the single most important number on your dashboard. It’s not ‘how many calls did the agent make’ or ‘how long did they last’, it’s ‘how many calls actually accomplished what we wanted.’ Benchmark Against Your Human Team On UnleashX, you can also benchmark the agent’s success rate against your human team’s success rate on the same use case. This is how leaders make confident decisions about where to deploy AI agents — you can see where the agent is outperforming humans, where the two are evenly matched, and where the agent still needs work before scaling up.

Step 5: Configurations

Step 5 is where you fine-tune the operational settings that affect how the agent behaves on every call. Most settings have sensible defaults and you only need to change them if the default isn’t working for your use case. Scheduling & Timing Decide when your agent is allowed to handle calls. Set the timezone, working hours window, and start time. The configuration depends on whether your agent is inbound or outbound:

Outbound Agents — Stick to permitted calling hours typically 9 AM to 9 PM in the customer’s local time. Calling outside these hours is illegal in many regions and damages your brand. UnleashX will only place calls during the window you define.
Inbound Agents — Set this to 24 hours so customers can get help any time of day, including nights, weekends, and holidays. This is one of the biggest wins of voice AI so your customers don’t have to wait for business hours to get their queries resolved.

Recommended DefaultOutbound: 9:00 AM start, 12-hour window in customer’s local timezone. Inbound: 24 hours, always-on coverage. Adjust based on your specific use case and local regulations.

Background Denoising Strips out background noise like traffic, chatter, music, machinery so the agent hears the caller clearly. This is especially valuable when calling people who are out and about, in cars, or in busy environments. The intensity is adjustable from light (0.1) to aggressive (1.0). Lower values preserve more natural sound; higher values prioritize clarity at the cost of some audio quality.

Recommended Default0.5 works well for most call types. Increase if you’re seeing comprehension issues from noisy backgrounds; decrease if calls sound unnaturally compressed.

Voicemail Detection Stops the agent from talking to a voicemail or answering machine. The moment the platform detects an answering machine, the call is either ended (default) or a pre-recorded voicemail message is left, depending on your setting.

Recommended DefaultON for outbound campaigns. OFF for inbound (since voicemail isn’t relevant). Voicemail detection saves significant cost by preventing wasted minutes on machines.

Voice Detection Confidence How confident the system must be that audio is actually human speech before the agent reacts. Higher values reduce false interruptions from background noise but make the agent slower to respond. Lower values are more responsive but may cause the agent to react to noises that aren’t actually speech.

Recommended Default0.9 keeps the agent stable in noisy environments while still being responsive. Lower to 0.7 if the agent is missing real speech; raise to 0.95 if it’s reacting to background noise.

Noise Ambience Optionally adds subtle background ambience (like a soft call center hum) so the agent sounds more like a real person in a real place. This is a small detail, but it makes the conversation feel more natural and reduces the ‘uncanny valley’ effect that pure-silence AI calls can have. Agent Interruption Handling Should the agent stop talking the moment a customer says anything? Probably not as that would include background ‘mm-hmm’s, ‘okay’s, and other filler sounds that aren’t actually interruptions. The Interruption Handling setting lets you require a minimum number of words before the agent yields the floor.

Recommended Default3 words. This filters out filler sounds and brief acknowledgments while still letting the agent stop quickly when the customer genuinely interrupts. Lower for more sensitive interruption handling; higher to make the agent more assertive.

Inactivity Handling If the customer goes silent for too long, the agent gently checks in (‘Are you still there?’) instead of immediately hanging up. Useful for letting people think, search for information, or talk to someone else briefly. Call Follow-Up When a customer asks to be called back later, the agent automatically schedules and triggers the follow-up call at the requested time. No manual queue management or callback list maintenance required. Call Transfer (Warm Handoff) Sometimes the agent needs to hand off to a human for complex issues, escalations, or when the customer explicitly asks for a person. Configure the destination phone number (your call center, a specific team, or a fallback queue) and the agent will transfer the call seamlessly. Currently supported with Twilio and select SIP providers. Do Not Call (DNC) Detection If a customer says ‘don’t call me again,’ ‘remove me from your list,’ ‘stop calling,’ or anything similar, the agent picks that up immediately and adds them to your DNC list. Critical for compliance with telemarketing regulations (TCPA in the US, TRAI in India, GDPR in Europe) and for protecting your sender reputation. Memory & Context Recall Let the agent remember details from previous conversations with the same caller. If a customer called last week about a refund, today’s call doesn’t start from zero. The agent picks up where the previous conversation left off. This is a game-changer for customer experience in support and account management use cases.

Step 6: Guardrails

Guardrails are the trust and safety controls that keep your agent compliant, accurate, and respectful. Every guardrail in this section is built-in and ready to enable you don’t need to write any code. We strongly recommend turning on every guardrail relevant to your use case before going live. Restricted Topics Define topics the agent should never discuss — things like competitor pricing, legal advice, medical diagnoses, internal discount policies, or unreleased product information. When these topics come up, the agent gracefully steers the conversation away or escalates to a human. Profanity & Abuse Filters Blocks offensive language from the agent’s responses (even if the model would generate it) and detects when callers are being abusive. You decide what happens next: de-escalate with calming language, end the call politely, or hand off to a human. This protects both your brand and your team’s wellbeing. Hallucination Prevention This is one of the most important guardrails for any agent. With hallucination prevention enabled, the agent only answers factual questions using information from your knowledge base. If it doesn’t know something, it says so honestly — instead of inventing an answer that could mislead your customer.

Why This MattersAn agent that confidently makes up answers will erode customer trust faster than any other failure mode. Hallucination prevention is the single most effective control for keeping your agent reliable in production. Always enable it.

PII Handling & Redaction Automatically detects and redacts personally identifiable information — credit card numbers, government IDs (Aadhaar, SSN, passport numbers), passwords, security questions — from transcripts and recordings. The actual conversation still happens normally; the redaction only affects the stored transcript, so your data stays clean and compliant. Recording Consent Plays a quick consent disclosure at the start of every call (‘This call may be recorded for quality and training purposes’) and respects customer opt-outs. Required by law in many jurisdictions (most of Europe, parts of the US, and elsewhere). The disclosure is configurable so you can match your specific legal requirements. Escalation Rules Define what triggers an automatic handoff to a human:

Frustrated or angry customer (detected by sentiment analysis)
Repeated misunderstanding (the agent failed to answer the same question twice)
Sensitive topic (complaint, legal issue, urgent medical situation)
Explicit customer request (‘I want to speak to a person’)
High-value or VIP customer (based on input variables)

The agent recognizes when it’s out of its depth and gets a human involved before the customer experience degrades.

4. Testing & Publishing

4.1 Why Test Before You Publish

A voice agent is a complex system. Even a small change to the prompt, the voice, or a tool configuration can have unexpected effects on the conversation. Always test before you publish, and test again any time you make a meaningful change.

4.2 Browser Test Calls

The fastest way to test your agent is a browser call. Click ‘Start Browser Call’ on the Agent Test tab and talk to the agent through your laptop microphone. You’ll hear it speak, see the live transcript, and watch which tools it triggers in real time. Browser calls are great for fast iteration cycles, change the prompt, click test, talk to the agent, refine, repeat. Most teams do dozens of browser calls during the design phase before ever going live.

4.3 Live Phone Test Calls

Once the browser call sounds right, do a live phone test. Click ‘Dial via Phone’ and the agent will call your number from the assigned caller ID. This gives you a real-world feel for how the call sounds over an actual phone connection including network latency, audio compression, and the speaker on the receiving device. Live phone tests catch issues that browser calls can’t: voice quality on a real handset, behavior when there’s background noise, how the agent sounds on speakerphone, and how it handles real telephony edge cases like one-sided audio glitches.

4.4 Pre-Publish Checklist

Before you click Publish and send live traffic to your agent, walk through this checklist:

Run at least 5 browser test calls covering happy paths and edge cases
Run at least 2 live phone tests from different devices and network conditions
Verify that all structured outputs are being captured correctly
Verify that the agent assigns the right disposition for each scenario
Confirm that all guardrails are enabled (DNC, restricted topics, hallucination prevention, recording consent)
Test the call transfer to a human agent if you have escalation enabled
Validate the recording consent message matches your legal requirements

5. Putting Your Agent to Work

Once your agent is published, you have four ways to put it to work. Pick the one that fits your workflow and you can use multiple at the same time for the same agent.

5.1 Campaigns (Batch Calling)

Got a list of customers to call? Upload it as a campaign and the agent will work through the list automatically. Campaigns are the right choice for any outbound use case where you have a known set of contacts to reach: lead outreach, renewal reminders, payment follow-ups, surveys, win-back outreach, and re-engagement. Campaigns include:

Smart Pacing — Controls how many calls happen per minute, so you don’t overwhelm your phone provider or violate calling rate limits.
Retry Logic — Automatically retries failed calls based on rules you define (number of attempts, gap between attempts, time-of-day restrictions).
Time-of-Day Rules — Only places calls during permitted windows in each contact’s local timezone.
Disposition-Based Branching — Use the disposition from one call to decide whether to retry, schedule a follow-up, or move on.

5.2 Workflows

Trigger a call when something happens in your business: a new lead signs up, a payment fails, a delivery is delayed, an order is placed, a support ticket is created. Drop the agent into a workflow and it will call automatically, in real time, when the moment is right. This is the right choice for event-driven outreach where the value of the call depends on it happening close to the triggering event. A workflow-triggered renewal reminder one hour after the renewal failed will perform much better than a batch reminder a day later. Workflows can be triggered from:

Webhook events from your own systems
CRM triggers (new lead, status change, opportunity stage)
Form submissions
Scheduled events (renewal date, anniversary)
Custom triggers via API

5.3 Inbound

Give the agent a phone number and let customers call in. Best for:

Customer support and FAQ handling
Order status and delivery tracking
Account help and self-service
Appointment booking and rescheduling
After-hours coverage when human agents are off duty

Inbound agents handle the high-volume, repetitive incoming calls that would otherwise tie up your support team leaving humans to focus on the calls that actually need a person.

5.4 API

For engineering teams: trigger calls programmatically from your own backend using the platform’s REST API. This gives you complete control over when and how calls happen, and lets you build custom integrations that fit exactly how your business works. API triggers are ideal for:

Custom workflows that don’t fit the standard workflow builder
Integrations with proprietary internal systems
High-frequency, low-latency event-driven calls
Any use case where you need complete programmatic control

6. Monitoring & Analytics

Once your agent is live, the work isn’t over, the work is just starting. Voice agents need ongoing monitoring, tuning, and improvement based on what’s actually happening on real calls. This section covers everything available on the Agent Detail page.

6.1 The Agent Dashboard

The Dashboard is your agent’s report card. It shows the key performance indicators that matter most:

Total Calls — How many calls the agent has handled in the selected time window.
Successful Calls — How many calls hit a success disposition (as defined in Step 4). This is your main KPI.
Unanswered Calls — How many outbound calls didn’t connect- voicemails, no-answers, busy signals.
Failed Calls — How many calls failed due to technical issues, hangups, or errors.
Average Duration — How long the average call lasts. Useful for spotting changes in conversation length over time.
Call Status Distribution — A bar chart showing how calls ended, connected, failed, unanswered, busy, user-ended, voicemail.
Disposition Breakdown — A horizontal bar chart showing the percentage of calls that ended in each disposition category.
Agent Success Benchmark — The percentage of calls that hit a success disposition, plus secondary metrics like CSAT, Resolution Rate, Average Handle Time, and First Call Resolution. This is where you compare against your human team’s benchmarks.
Recent Calls — A live feed of the most recent calls the agent has handled, with one-click View Details to dive into any conversation.

6.2 Activity Logs

The Activities tab shows every single call your agent has ever handled. Each row shows:

From and to phone numbers (or ‘Live Browser Call’ for tests)
Call ID for searching and referencing
Date and time
Duration (if connected)
Who ended the call (Agent or User)
Final status (Completed, Failed, No Answer, Busy)
Campaign or workflow that triggered the call
Disposition assigned by the agent

You can filter by status, duration, campaign, or date range, and search by phone number or call ID. Click View Details on any row to open the full conversation analytics.

6.3 Conversation Analytics (Per-Call Deep Dive)

This is where you really understand what happened on a specific call. Open any call from the Activities tab and you’ll see a complete breakdown. Call Information At the top: phone numbers, status badge, recording player, and the disposition the agent assigned with its description. Summary An AI-generated summary of the call in 3-5 sentences. This is the fastest way to understand what happened without listening to the recording or reading the full transcript. Alongside the summary, you’ll see Key Outcomes, the most important results extracted from the call. Analysis The Analysis section breaks down the call along three dimensions:

Sentiment — Overall positive/neutral/negative ratio with a percentage breakdown. Helps you spot calls where the customer was unhappy.
Call Metrics — Talk-to-listen ratio, average response time, interruption count, total silence, words per minute. These are the operational KPIs that tell you whether the agent is conversational or robotic.
Actionable Insights — AI-generated recommendations on what went well and what could be improved. Each insight is tagged as a Strength or an Improvement opportunity, with an Impact rating (High/Medium/Low) so you know what to focus on first.

Structured Outputs All the variables the agent extracted from the call, displayed as a clean key-value table with the data type for each field. This is the data that flows into your CRM and analytics systems. Transcript The complete conversation, with timestamps for every message and a sentiment indicator for each turn. The visual layout makes it easy to scan the conversation flow and spot any moments where the conversation went off-track.

Use Analytics to Improve Your AgentThe most successful teams review their bottom 10% of calls weekly. Look for patterns: Is the agent failing on specific question types? Is it making the same mistake repeatedly? Use the Actionable Insights to guide prompt updates and keep iterating.

7. Best Practices

7.1 Designing Effective Prompts

Start with identity — Open with who the agent is, who it represents, and what its goal is. Be specific — ‘You are a customer service representative for Acme Corp, helping customers with order status’ beats ‘You are a helpful AI’.
Use examples, not just instructions — Show the agent what to say with example exchanges. LLMs follow examples more reliably than abstract rules.
Handle the unhappy path — Spend at least as much time on what to do when things go wrong (customer is angry, doesn’t know the answer, asks something off-topic) as on the happy path.
Keep it focused — Resist the urge to make one agent do everything. A focused agent that does one thing well outperforms a generalist agent every time.
Iterate based on real calls — Your first prompt won’t be perfect. Review real call transcripts and refine the prompt based on what you actually see.

7.2 Choosing the Right Voice

Match the brand — A premium financial brand needs a different voice than a casual e-commerce brand. Think about what your customers expect to hear.
Match the use case — Calm and reassuring for healthcare. Confident and energetic for outbound sales. Warm and patient for elderly customer support.
Test with real customers — Run a small pilot before scaling up. The voice that sounds best in your office may not be the one your customers respond to best.

7.3 Building Trust with Guardrails

Always enable hallucination prevention — There’s no use case where you want the agent to make things up. This should be on by default for every agent.
Always enable DNC detection — Even if you think it’s not relevant, enable it. It’s free protection against compliance issues.
Always disclose recording — If you’re recording calls (and you should be, for quality and analytics), always disclose it. It’s the law in many places and builds customer trust everywhere.
Be ready to escalate — Configure call transfer to a human agent for any sensitive situation. Customers value the option to speak to a person, even if they rarely use it.

7.4 Iterating from Real Data

The best voice agents are not built once and forgotten. They’re tuned continuously based on what’s happening on real calls. Set aside time each week to:

Review the dashboard and look for trends. Is the success rate going up or down?
Listen to a sample of recent calls (especially the failed ones)
Read the Actionable Insights for individual calls
Update your prompt based on patterns you see
Test the changes with browser calls before publishing
Compare metrics before and after changes to validate improvements

8. Glossary

Term	Definition
Agent	An AI-powered conversational system configured to handle voice calls for a specific use case.
Agent Success Benchmark	The percentage of calls that result in a disposition you’ve marked as successful. The primary KPI for measuring agent effectiveness.
BYO Number	Bring Your Own Number: connecting an existing phone number from a third-party telephony provider via SIP.
Campaign	A batch outbound calling job that processes a list of contacts according to defined pacing, retry, and timing rules.
Conversation Flow	Another term for Multi-Prompt Flows: a stage-based conversation structure where the agent moves between focused prompts.
Disposition	A label assigned to each call to indicate its outcome (Interested, Not Interested, Resolved, etc).
DNC	Do Not Call — a customer’s request to be excluded from future outbound calls. Detection of DNC intent and respect of DNC lists is required for telemarketing compliance.
Guardrails	Built-in trust and safety controls (DNC detection, hallucination prevention, PII redaction, etc.) that keep the agent compliant and trustworthy.
Hallucination	When an LLM generates information that sounds plausible but is not true. Hallucination prevention forces the agent to only answer from the knowledge base.
Input Variable	A placeholder in the prompt (like `{customer_name}`) that gets filled with a real value at call time, enabling personalization.
Knowledge Base	Documents and reference material attached to the agent so it can answer factual questions with grounded, accurate information.
LLM	Large Language Model: the AI model that powers the agent’s reasoning and response generation.
Multi-Prompt Flow	A conversation structure where the agent uses different focused prompts for different stages of the call.
PII	Personally Identifiable Information: sensitive customer data that should be redacted from transcripts and recordings.
Prompt	The instructions you write that define how the agent behaves: its identity, goals, tone, and conversation flow.
Single-Prompt	A conversation structure where the entire agent behavior is defined in one comprehensive prompt.
SIP	Session Initiation Protocol: the standard for connecting third-party phone numbers and PBX systems to the platform.
Structured Output	A defined variable the agent extracts from each call (customer name, email, demo date, etc.).
Transcriber	The speech-to-text engine that converts caller audio into text for the agent to process. Also called STT.
TTS	Text-to-Speech: the engine that converts the agent’s text responses into spoken audio.
Voicemail Detection	Automatic detection of an answering machine on the receiving end of an outbound call, allowing the agent to either hang up or leave a message.
Webhook	An HTTP callback the agent can trigger during a call to integrate with your own systems.
Workflow	An event-driven trigger that calls the agent in response to something happening in your business systems (a new lead, a failed payment, etc.).

Documentation Index

​1. Introduction

​1.1 What is an UnleashX Voice Agent?

​1.2 Who Is This For?

​1.3 What You Can Do With UnleashX Voice Agents

​2. Core Concepts

​2.1 The Anatomy of a Voice Agent

​2.2 Inbound vs Outbound Agents

​2.3 Single-Prompt vs Multi-Prompt Conversation Flow

​Single-Prompt Agent

​Multi-Prompt Flow

​3. Building Your First Agent

​Step 1: Agent Setup & Brain

​Step 2: Voice & Transcription

​Step 3: Phone Number

​Step 4: Outputs & Success Metrics

​Step 5: Configurations

​Step 6: Guardrails

​4. Testing & Publishing

​4.1 Why Test Before You Publish

​4.2 Browser Test Calls

​4.3 Live Phone Test Calls

​4.4 Pre-Publish Checklist

​5. Putting Your Agent to Work

​5.1 Campaigns (Batch Calling)

​5.2 Workflows

​5.3 Inbound

​5.4 API

​6. Monitoring & Analytics

​6.1 The Agent Dashboard

​6.2 Activity Logs

​6.3 Conversation Analytics (Per-Call Deep Dive)

​7. Best Practices

​7.1 Designing Effective Prompts

​7.2 Choosing the Right Voice

​7.3 Building Trust with Guardrails

​7.4 Iterating from Real Data

​8. Glossary

1. Introduction

1.1 What is an UnleashX Voice Agent?

1.2 Who Is This For?

1.3 What You Can Do With UnleashX Voice Agents

2. Core Concepts

2.1 The Anatomy of a Voice Agent

2.2 Inbound vs Outbound Agents

2.3 Single-Prompt vs Multi-Prompt Conversation Flow

Single-Prompt Agent

Multi-Prompt Flow

3. Building Your First Agent

Step 1: Agent Setup & Brain

Step 2: Voice & Transcription

Step 3: Phone Number

Step 4: Outputs & Success Metrics

Step 5: Configurations

Step 6: Guardrails

4. Testing & Publishing

4.1 Why Test Before You Publish

4.2 Browser Test Calls

4.3 Live Phone Test Calls

4.4 Pre-Publish Checklist

5. Putting Your Agent to Work

5.1 Campaigns (Batch Calling)

5.2 Workflows

5.3 Inbound

5.4 API

6. Monitoring & Analytics

6.1 The Agent Dashboard

6.2 Activity Logs

6.3 Conversation Analytics (Per-Call Deep Dive)

7. Best Practices

7.1 Designing Effective Prompts

7.2 Choosing the Right Voice

7.3 Building Trust with Guardrails

7.4 Iterating from Real Data

8. Glossary