Voice Agents

Voice agents are AI personas that make and receive phone calls, generate personalized outreach copy, and respond to emails — all using your org’s context and knowledge base. Each agent has a configurable personality, voice, phone number, and set of skills.

How It Works

1. Go to Agents → Agents
2. Click New and choose Agent or Twin
3. Configure the agent’s persona, identity, and knowledge
4. Attach skills and activate

Agents vs. Twins

Type	Purpose	Extra Features
Agent	General-purpose AI persona for calls, emails, and chat	Preset voice selection
Twin	Digital replica of a real person	Custom voice cloning, LinkedIn profile extraction, writing style upload

Twins inherit their source agent’s persona, role, and voice — then layer on personal context from LinkedIn data and writing samples to match the real person’s tone.

Persona

The Persona tab controls how your agent communicates.

Setting	What It Controls
Role	SDR, AE, Receptionist, CSM, GTM Engineer, or Custom
Tone / Formality	Casual (0.1) → Moderate (0.5) → Formal (0.8)
Conciseness	Detailed (0.1) → Balanced (0.5) → Direct (0.8)
Assertiveness	Gentle (0.1) → Balanced (0.5) → Assertive (0.8)
Persona Prompt	Full free-text instructions for the agent’s behavior
Outbound Call Instructions	Override behavior specifically for outbound calls
Inbound Call Instructions	Override behavior specifically for inbound calls
Voicemail Message	Per-agent voicemail script (falls back to org-level if empty)

Voicemail messages support template variables: {first_name}, {bot_name}, {company_name}.

Identity

The Identity tab assigns communication channels to your agent.

• Phone Number — select from available Twilio numbers. Required to make or receive calls
• Voice — choose a preset Cartesia voice (Michael, Sarah, David, Emma, and others) or, for twins, clone a custom voice
• Calendar — link an appointment type for booking meetings during calls
• Company Knowledge — toggle access to the org-wide knowledge base

Voice Cloning (Twins Only)

1. In the agent form, switch to Custom voice mode
2. Click Clone New Voice
3. Select a sample script (Professional, Conversational, or Friendly)
4. Record up to 10 seconds of audio or upload a WAV/MP3 file
5. Click Clone Voice

Cloned voices are saved to your org and can be reused across multiple twins.

Knowledge

Agents can access two types of knowledge:

• Company Knowledge — org-wide context including company description, services, products, pain points, value proposition, pitch, FAQs, target audience, and competitive advantage
• Collections — up to 4 document collections per agent. Documents are chunked and vectorized for retrieval during conversations

Skills

Attach reusable AI skills to give your agent capabilities beyond conversation. Go to the Skills tab on the agent detail page and click Attach Skills.

Skills have three runtime types:

Type	What It Does
LLM	Runs a prompt template against an AI model with configurable temperature and token limits
API	Calls an external HTTP endpoint with method, headers, body template, and auth
Script	Executes Python, JavaScript, or TypeScript code with configurable timeout (up to 300s)

Skills can require human approval before execution. Execution history — including input, output, duration, token count, and cost — is tracked on the agent’s Activity tab.

On-Demand Content Generation

Voice agents power personalized content generation at send time across channels:

Channel	What Happens
Email	Per-contact personalized email generated using CDP data, campaign context, and conversation history
LinkedIn	Per-contact LinkedIn message generated at send time — no static templates needed
Call Script	Per-contact call script generated with live CDP data before dialing
Voicemail	Dynamic voicemail message personalized per contact

On-demand generation pulls from your Global Context, campaign documents, and the contact’s enrichment data. Twin tone profiles (formality, conciseness, assertiveness) are applied to match the sender’s style.

Phone Calls

Outbound Calls

Agents can place outbound calls via Twilio or Telnyx. Each call links to a campaign and sequence step for attribution. After the call, a callback delivers the transcript, summary, disposition, sentiment, recording, and duration.

Inbound Calls

Assign a phone number to an agent to handle inbound calls automatically. The agent uses your inbound instructions, company knowledge, and attached collections to respond. If the agent can’t resolve the call, it falls back to voicemail or queues for an SDR to claim.

Outbound Playbooks

A playbook is a structured outbound script attached to an agent. Playbooks make calls consistent across prospects and let you A/B test scripts.

Playbook Structure

Section	Purpose
Opening	First 15 seconds — pattern interrupt, name drop, permission to continue
Discovery	2–4 qualifying questions tied to ICP fit
Value pitch	Positioning aligned with the prospect’s role and pain
Objection handling	Responses to the 5 most common objections
Close	CTA (book meeting, email resource, callback) and next-step confirmation

Attaching a Playbook

1. Open the agent’s Persona tab
2. Paste the playbook into Outbound Call Instructions
3. Save — the agent reads the playbook top-to-bottom, branching on prospect responses

Tip

Create one playbook per persona. A VP Sales playbook should sound different from a Marketing Director playbook — pain points, proof points, and CTAs all shift.

Receptionist Routing Rules

Receptionist agents qualify and route inbound calls. Rules run in order — the first match wins.

Rule Types

Rule Type	When to Use
Keyword match	Caller mentions a word or phrase (e.g., “billing”, “support”, “demo”)
Intent classification	Agent classifies caller intent (e.g., sales vs. support vs. existing customer)
Business hours	Route differently based on time of day and weekday
Caller identity	Route known contacts (from CRM) differently from unknown callers
Voicemail fallback	Default action when no other rule matches

Routing Actions

Action	Behavior
Transfer to human	Warm or cold transfer to a phone number or user
Transfer to another agent	Hand off to a specialized AI agent
Take a message	Record a voicemail and email it to a recipient
Book a meeting	Pull up calendar and schedule on the call
Hang up	For known-bad numbers (sales lists, spam)

Warm vs. Cold Transfer

Transfer Type	Behavior
Warm transfer	Agent briefs the human before connecting — explains who’s calling and why
Cold transfer	Agent connects the caller directly with no briefing

Warm transfers take ~10 extra seconds but reduce friction for the human picking up.

Post-Call Summary

After every call, graph8 generates a structured summary for CRM and coaching.

Generated Fields

Field	Description
Disposition	Call outcome (Booked, Callback, Not Interested, etc.)
Summary	2–3 sentence recap of what was discussed
Key points	Bulleted list of topics, objections, and commitments
Next steps	Action items for the rep (send case study, schedule follow-up)
Sentiment	Positive, Neutral, or Negative based on tone analysis
Qualification score	0–100 based on discovery answers vs. ICP criteria
Intent signals	Buying signals surfaced during the call

Summaries push to the connected CRM as activity notes and appear in the contact’s Inbox timeline.

Coaching Loop

Turn calls into training material without manual review.

Auto-Flagged Calls

graph8 auto-flags calls that match coaching criteria:

• Calls with low qualification scores (missed discovery)
• Calls with unresolved objections
• Calls where transfer was triggered by frustration
• Calls with long silences or interruptions

Flagged calls appear in the agent’s Activity tab with timestamps on the problem moments.

Review Workflow

1. Open the Activity tab
2. Filter by flag type (e.g., “Missed Objections”)
3. Click a call to open the transcript
4. Click any line to jump to that timestamp in the recording
5. Update the agent’s call instructions based on what you learned

Tip

Batch coaching weekly. Review 10 flagged calls, extract the 3 most common gaps, and update the playbook or objection-handling section. Redeploy and measure conversion delta.

Webhook Events

Subscribe to voice events to build custom integrations.

Event	When It Fires
call.started	Agent picks up or dials out
call.answered	Far side answers
call.ended	Call completes (any disposition)
call.transferred	Call handed off to another party
call.voicemail	Agent leaves voicemail
call.summary.generated	Post-call summary is available
agent.flagged	Call auto-flagged for coaching

Configure webhook URLs at Settings → Developer → Webhooks. Events are signed with HMAC-SHA256.

Privacy and Recording Retention

Setting	Default	Notes
Recording retention	90 days	Extend in Settings → Compliance
Transcription retention	Indefinite	Used for search, coaching, and analytics
Two-party consent states	Auto-notice	Agent plays a recording notice at call start in states requiring consent
Do Not Call list	Auto-scrubbed	Numbers on the DNC list are excluded from outbound campaigns

Caution

Check local call recording laws. graph8 enables the auto-notice by default, but you are responsible for compliance with state and country-level wiretap rules.

Testing

Click Test Digital Agent on the agent detail page to run a test before going live.

Phone Call Test

• Outbound — enter a phone number and click Start Test Call. A live status card shows call state (Dialing → Ringing → Active → Completed) with a real-time transcript
• Inbound — call the agent’s assigned number from your phone. The modal shows the expected greeting and instructions

Email Test

• Respond — select a pre-built scenario (Billing Inquiry, Technical Support, Sales Inquiry, Feature Request, Service Complaint), review the incoming email, and generate a response
• Compose — choose a scenario or write custom context, then generate a fresh email with recipient info and context

Activity & Analytics

The Activity tab on the agent detail page shows a filterable feed of all interactions:

Filter	What It Shows
Calls	Inbound/outbound calls with transcript, audio player, summary, disposition, sentiment, and duration
Skills	Skill executions with status, token count, cost, input/output JSON, and errors
Workflows	Workflow runs with status, duration, input/output data

The Analytics tab shows aggregate metrics: total calls, success rate, average duration, and total duration.

Tip

Test both inbound and outbound calls before activating an agent. The real-time transcript in the test modal lets you verify persona behavior, knowledge retrieval, and call instructions without affecting live contacts.

Next Steps

• Skills → — Create reusable AI actions for agents
• Workflows → — Chain skills and actions into automated multi-step processes
• Dialer → — Use agents with the parallel dialer for outbound calling
• Sequencer → — Enroll contacts in sequences that use agent-generated content