Add AI Voice Agents to Your Existing Phone System

Why add AI to the phone system you have, instead of replacing it

Many AI voice products assume you will move your numbers to them, or forward every call to a new platform. For a small business with one line that can be fine. For anyone with a working PBX, it is usually the wrong first move.

Your phone system already does a lot: it owns your numbers and carrier contracts, routes calls to queues and extensions, records calls, feeds reports, and connects to desk phones, softphones and perhaps a CRM. Staff know how it works. Replacing it to add AI means re-creating all of that, retraining people and accepting risk on every call while you do.

Adding the agent to the existing system turns AI into one more destination, like a queue or an extension. You decide which calls reach it, you can switch it off by changing one route, and the rest of your telephony keeps working exactly as before. That makes it far easier to start small, measure, and grow only when the results are clear.

What changes and what stays the same

Four routing patterns, from lowest risk to highest reach

The first decision is which calls the agent gets. These four patterns cover almost every deployment, and many businesses move from the first to the last over a few months.

A fifth, quieter option is to let the AI listen rather than talk: real-time transcription, summaries and agent assist on calls your staff handle. That is AI VoIP integration rather than a voice agent, and it is often a sensible first step. See our AI VoIP integration page.

How each kind of phone system connects

Almost every modern business phone system can send a call to another SIP destination. The details differ, and a few systems have limits worth knowing before you plan.

Our SIP trunk to AI voice agent guide covers trunk configuration in detail, and LiveKit with Asterisk and FreeSWITCH walks through a self-hosted set-up.

Example: after-hours and overflow routing on Asterisk

This dialplan sends calls to the AI agent outside business hours, and during business hours tries the support queue first, overflowing to the agent if nobody answers in time. The agent is a PJSIP endpoint named ai-agent.

; extensions.conf
[from-carrier]
exten => _X.,1,NoOp(Inbound call from ${CALLERID(num)})
 same => n,Set(__ORIGINAL_DID=${EXTEN})
 same => n,GotoIfTime(09:00-18:00,mon-fri,*,*?open)
 ; closed: AI agent takes the call
 same => n,Set(__ROUTE=after-hours)
 same => n,Dial(PJSIP/${EXTEN}@ai-agent,,b(add-context^s^1))
 same => n,Goto(voicemail,s,1)              ; fallback if the agent is down
 same => n(open),Queue(support,t,,,25)      ; staff first, 25 seconds
 ; overflow: nobody answered, AI agent takes it
 same => n,Set(__ROUTE=overflow)
 same => n,Dial(PJSIP/${EXTEN}@ai-agent,,b(add-context^s^1))
 same => n,Goto(voicemail,s,1)

[add-context]
; runs on the outbound channel: pass details to the agent as SIP headers
exten => s,1,Set(PJSIP_HEADER(add,X-Original-DID)=${ORIGINAL_DID})
 same => n,Set(PJSIP_HEADER(add,X-Route-Reason)=${ROUTE})
 same => n,Return()

Two details matter in production. First, always keep a fallback (voicemail or a queue) for when the agent is unreachable. Second, tell the agent why it got the call: an overflow caller who has already waited needs a different greeting from an after-hours caller.

Handing calls back to your staff

Handover is where AI deployments are judged. Callers forgive an AI that says "let me get a colleague"; they do not forgive repeating everything to that colleague.

Our AI-to-human handover guide compares the methods in more depth.

Numbers, caller ID, recording and calling rules

• Numbers: keep them on your carrier and PBX; route to the agent rather than porting numbers to an AI vendor.
• Caller ID: make sure the caller's number reaches the agent, so it can look up the customer; forwarding sometimes replaces it.
• Recording: record on the PBX as you do today, and store the agent's transcripts under the same retention rules.
• Disclosure: tell callers they are speaking with an AI assistant where the law or your policy requires it, and how to reach a person.
• Outbound calls: AI-voiced outbound calls are regulated in many countries, including consent rules in the US under the TCPA; check each country you call.
• Data: decide where audio and transcripts are processed and stored, and whether a self-hosted option is needed.

See our guide to TCPA rules for AI outbound calling, and self-hosted vs managed voice AI for where call data goes.

Audio quality and latency

Routing through your PBX adds very little delay. The things that make an agent feel slow are elsewhere, and most can be fixed in design.

A six-step rollout plan

Measuring whether it works

Our voice agent testing and monitoring guide covers how to collect these numbers.

A real example

CelloIP built an AI voice agent for a contact centre in North America that already ran FreeSWITCH. Nothing in the existing telephony was replaced:

• FreeSWITCH kept the carrier trunks, inbound routing, outbound dialing, queues and human agents.
• Chosen calls go to a LiveKit Agents voice agent over SIP, and return to FreeSWITCH queues for handover, with a summary.
• The agent handles inbound support, outbound calls and appointment booking in English and Spanish.
• It works with the contact centre's CRM, booking system and helpdesk through defined tools.

Read the full case study.

What it costs

Ranges match our AI voice agent development page. Running costs (telephony, speech and model usage, hosting) are separate; see our AI voice agent cost guide.

Common mistakes

Frequently asked questions