Skip to content

outbound-agents.md

Latest commit

 

History

History
298 lines (211 loc) · 8.69 KB

outbound-agents.md

File metadata and controls

298 lines (211 loc) · 8.69 KB
Raw

Outbound Agent Design

How to build outbound AI voice agents that handle the unique challenges of calling people who didn't ask to be called.

Why Outbound Is Different

Inbound agents talk to people who chose to call. Outbound agents face a fundamentally different set of problems:

Challenge Why it matters
Unknown destination May reach human, voicemail, IVR, fax, or nothing
Recipient skepticism They didn't initiate contact — may assume spam
First 3–5 seconds are critical Determines whether they hang up or engage
Varied phone systems Different carriers, business PBXs, VoIP providers, Google Voice

The Three Scenarios

Every outbound call lands in one of these:

Scenario Detection method Required action
Human Natural greeting, conversational speech Proceed with task
Voicemail Carrier/personal greeting with "leave a message" signals End call immediately with voicemail tool
IVR "Press 1 for..." menu options Navigate to reach a human

For detailed voicemail/human detection logic, see voicemail-detection.md.

IVR Navigation

Phase 1: Aggressive — Try Shortcuts First

Most IVR systems have hidden shortcuts. Try these before listening to the full menu:

  1. Press 0 three times (DTMF tool) — many systems route 0 to an operator
  2. Say "representative" 2–3 times
  3. Say "operator" 2–3 times

During this phase: don't introduce yourself, don't say anything else, don't press other buttons.

If transferred to hold music, wait silently for a human.

Phase 2: Structured — Navigate the Menu

If shortcuts fail, listen to the full menu and select strategically:

Option type Action
"Speak to someone" / "Representative" / "Operator" Select this
"Front desk" / "Store" / "Host" / "Manager" Good second choice
"Orders" / "Billing" Often has humans
"Store hours" / "Directions" Avoid — usually recordings
"For English, press 1" Press 1 first, then navigate

When to Give Up

End the call when:

  • You've tried both phases and the IVR keeps looping
  • You're redirected to "visit our website" or "call another number"
  • The IVR ends in voicemail

Don't give up after just one attempt. Make a good-faith effort with multiple paths.

DTMF Tool Configuration

type: dtmf

The built-in DTMF tool lets the agent press phone keypad buttons. No additional configuration needed.

Opening Statement Design

The opening line determines whether the recipient hangs up. Optimize ruthlessly.

What Works

"Hi, are you open right now?"
"Hi, I'm calling about your recent order."
"Hi, this is a quick call to confirm your appointment tomorrow."

Why: Under 5 seconds, asks a simple question, gets to the point, doesn't trigger "sales call" defenses.

What Doesn't Work

"Hello, my name is Alex and I'm an automated assistant calling
on behalf of Acme Corp regarding your account. How are you today?"

Why: Too long, robotic framing, "How are you today" is an obvious sales opener that triggers immediate hang-up.

Rules

  • Get to the point in the first sentence — no preamble
  • Keep it under 5 seconds of speech
  • Ask a simple question — gives them something to respond to
  • Don't volunteer your identity — but answer honestly if asked

Identity Handling

Don't reveal that you're an AI in the opening. But never lie when asked directly.

They say You respond
"Who is this?" "This is an automated system from [Company] calling to [purpose]."
"Are you a robot?" "Yes, I'm an automated assistant calling on behalf of [Company]."
"Are you a real person?" "I'm an automated assistant. I'm calling to [purpose]."
Unrelated questions "I can only help with [specific task]. For other questions, please contact [Company] support."

Conversation Flow Design

The Five Golden Rules

  1. Be direct — get to the point in the first sentence
  2. Be brief — short sentences, simple words
  3. Don't rush — wait for responses after each statement
  4. Stay in scope — only do what you're designed to do
  5. End gracefully — always thank them before hanging up

Step-by-Step Flow Template

## Step 1: Opening
- Deliver opening line
- Wait for response

## Step 2: Handle Response
- Clear answer → acknowledge and continue
- Unclear → ask for clarification (max 2 times)
- No response → repeat once, then close

## Step 3: Provide Context (if needed)
- Explain why you're calling (1–2 sentences max)
- Wait for response

## Step 4: Listen & Acknowledge
- If they share info → acknowledge empathetically
- Do NOT prompt for additional information unprompted
- Do NOT rush to the next step

## Step 5: Close
- Summarize outcome (if applicable)
- Thank them
- End call gracefully

Pacing and Interruption

LLMs tend to rush through outbound scripts. Counter this explicitly in your prompt:

# Pacing Guidelines

- DO NOT rush through the conversation
- Wait for the user to respond after each statement
- Use natural pauses: <break time='0.3s'/>
- If they're speaking, do not interrupt
- Give them time to process — they weren't expecting this call

Model Parameters for Outbound

model:
  provider: openai
  model: gpt-4.1
  temperature: 0.1
  maxTokens: 250
  • Low temperature (0.1) — consistent, predictable responses
  • Low maxTokens (250) — forces brief responses, prevents monologues

Error Handling

Situation Response
No response after opening Repeat question once, then thank and end call
Unclear answer after 2 attempts Acknowledge and proceed to closing
Complaints Acknowledge empathetically, do not offer solutions, direct to support
Hostility Thank them for their time and end call gracefully
"Take me off your list" Acknowledge, apologize, end call. Flag for suppression list.

Required Tools for Outbound

Every outbound agent needs at minimum:

# 1. Voicemail detection (ends call with 'voicemail' endedReason)
type: voicemail
function:
  name: end_call_on_voicemail
  description: End the call immediately when voicemail is detected.
beepDetectionEnabled: false

# 2. DTMF (for IVR navigation)
type: dtmf

# 3. End call (for graceful termination)
type: endCall
function:
  name: end_call
  description: End the call after completing the task or when unable to proceed.

Prompt Structure for Outbound

Recommended section order for outbound agent system prompts:

# Identity
[Who the agent is, what company, what purpose]

# Identity Lock
[Fixed identity, scope limits, jailbreak prevention]

# Call Type Awareness
[How to detect voicemail/IVR/human — brief version]

# Voicemail Detection
[Trigger phrases and action — use voicemail tool]

# IVR Navigation
[Phase 1 aggressive + Phase 2 structured]

# Task & Goals
[Step-by-step conversation flow]

# Response Guidelines
[Tone, pacing, brevity rules]

# Error Handling
[What to do when things go wrong]

# Examples
[Concrete conversation examples with tool calls]

# Tools Available
[List of tools and when to use each]

Metrics and Testing

Key Metrics

Metric Target Notes
Connection rate >60% Calls that reach human or voicemail
Voicemail detection accuracy >95% Test against carrier/personal/business VM
Task completion rate >80% Of calls that reach humans
Average call duration <60s For simple tasks
Hang-up rate (first 5s) <20% Opening effectiveness

Test Matrix

Scenario What to verify
Human — cooperative Completes task successfully
Human — confused ("Who is this?") Handles identity questions gracefully
Human — hostile Exits gracefully without escalating
Human — doesn't answer question Clarifies appropriately
Voicemail — carrier greeting Detects and ends call
Voicemail — personal greeting Detects and ends call
Voicemail — Google Voice Detects and ends call
Voicemail — business voicemail Detects and ends call
IVR — simple menu Navigates to human
IVR — complex menu Makes good-faith attempt
IVR — no human option Ends call appropriately
IVR — Spanish option first Handles language selection
Silence Handles via timeout
Background noise Functions with kitchen/office noise
Interruption Handles being interrupted mid-sentence
Off-topic question Stays in scope
Jailbreak attempt Maintains identity lock