Lacy.ai Help Center

AI Phone Agent Engine Guide

Learn how Dualplex and Pipeline engines work, where each performs best, and how to choose the right foundation for every AI phone agent.

Book a Demo Create an Account
Quick ruleConversation-heavy → DualplexData-heavy → Pipeline

Start With the Call's Primary Job

What is an AI phone agent engine?

The engine determines how a caller's speech is interpreted, how the AI reasons and follows instructions, and how the response is spoken back. The right choice depends less on which engine is “better” and more on what the call must accomplish.

01 · Realtime engine

Dualplex

Built for natural, free-flowing conversation.

Caller audio moves directly into the realtime model, which listens, reasons, and responds in one continuous loop. This creates fluid pacing and a highly conversational experience.

Caller audio→Realtime AI→Voice

Best for

  • ✓AI receptionists and call routing
  • ✓Lead qualification and discovery
  • ✓Objection handling and open conversation
  • ✓Customer support and knowledge responses
02Structured engine

Pipeline

Built for accurate data and dependable execution.

Pipeline separates the call into three specialized stages: speech becomes text, the language model processes the transcript, and a voice service speaks the response.

Speech-to-text→LLM→Voice

Best for

  • ✓Intake, ordering, and verification
  • ✓Phones, emails, addresses, and IDs
  • ✓Appointment booking and CRM updates
  • ✓Structured workflows and tool calling

Engine Comparison

Choose based on the outcome that matters most

Most businesses need both conversational quality and accurate data. The deciding factor should be the part of the call that cannot fail.

Decision factorDualplexPipeline
Conversation styleMost natural and fluidClear and structured
Exact data captureRequires careful confirmationStrongest fit
Tool executionGood for lighter workflowsReliable for complex workflows
Best call typesReception, qualification, supportIntake, booking, verification
Primary advantageHuman-like responsivenessTranscript-driven accuracy

Recommended Starting Points

Configure the engine around the work

These baselines provide a practical starting point. Production performance should always be validated against real calls and edge cases.

Language model

Pipeline default: GPT-5.1

A reliable starting point for instruction following, structured workflows, and tool calling.

Prompt rule

“Read important values back for confirmation. Never assume, complete, modify, or round a value that was not clearly provided.”

Dualplex turn-taking

Responsive without clipping callers

Allow interruptions
On
Interrupt sensitivity
≈ 0.5
Turn detection
≈ 0.4
AEC warmup
≈ 3 sec
Noise cancellation
Off

A sensitivity value that is too high can miss short answers such as “yes” and “no.” Noise cancellation may clip the beginning of speech.

Production voice

Start with a stable voice baseline

CartesiaElevenLabs
Stability
0.70
Similarity
0.50
Avoid stability below 0.30

Very low stability can make the voice sound inconsistent or erratic across a call.

Before Production

Test the call, not just the configuration

Recommended settings are a baseline. The final configuration should reflect actual callers, phone lines, integrations, and business rules.

01

Short answers

Confirm the agent hears quick responses such as yes, no, and dates.

02

Structured data

Test phone numbers, spelled emails, addresses, IDs, and corrections.

03

Real interruptions

Verify that callers can pause, interrupt, clarify, and change direction.

04

Workflow execution

Validate bookings, CRM updates, transfers, notifications, and failures.

Frequently Asked Questions

AI phone agent engine FAQs

Which AI phone agent engine sounds the most natural?+

Dualplex typically provides the most fluid, human-like pacing because the caller's audio moves directly into the realtime model. It is a strong fit for reception, qualification, objection handling, and other conversation-heavy calls.

Which engine is better for capturing phone numbers and emails?+

Pipeline is generally the better choice for structured data capture. Its dedicated speech-to-text layer creates a transcript before the language model acts, improving reliability for phone numbers, spelled emails, addresses, IDs, and order details.

Can one engine be used for every AI phone agent?+

Yes, but the best result comes from choosing the engine around the call's primary job. Use Dualplex when natural conversation is the priority and Pipeline when accurate data capture and dependable tool execution matter most.

Every Conversation, Operationalized

Build the right AI phone experience for your business.

See how Lacy.ai can answer calls, capture customer information, book appointments, update your CRM, and trigger the next workflow.

Book a Demo ↗Create an Account