How to Build an AI Voice Agent in GoHighLevel

A GoHighLevel Voice AI agent can pick up a call, book the slot on your calendar, and tag the contact before your phone finishes its second ring.

Getting there means working through three layers inside the platform: what the agent is told to say, what it's allowed to look up, and which phone number actually puts a caller in front of it. Skip a layer and the agent either repeats itself, guesses at an answer, or never rings at all.

This walks through a working GoHighLevel AI voice agent setup from the first screen to the test call, using HighLevel's own documentation for every step.

The three things GoHighLevel AI voice agent setup depends on

Before you open the builder, three pieces need to already exist, or you'll be filling in placeholders you have to redo later.

You need a Voice AI agent with a defined goal, a knowledge base it can search for real answers, and a phone number or chat widget to deploy it on.

HighLevel treats these as separate configuration phases rather than one form, and jumping to deployment before the agent's prompt is solid is the most common way a launch goes wrong.

None of it needs a developer. It does need someone willing to write out what the agent should and shouldn't say, and that takes longer than the technical setup itself.

Where to build it: AI Agents, then Voice AI

Setup lives inside your sub-account under AI Agents, then Voice AI, then the Agent List tab. From there, Create Agent gives you three starting points.

• Browse Marketplace installs a prebuilt agent built for a specific use case, like a front desk receptionist.
• Create Custom Agent opens a full-screen builder where you configure everything from a blank slate.
• Flow Based Builder gives you a visual, node-based way to design the conversation, starting from one of five ready-made Voice AI templates covering use cases like lead qualification and post-call follow-up.

Whichever path you pick, you'll end up setting the same core fields: the agent's name, its voice, which model and language it runs on, a welcome message for inbound or outbound calls, and how long it pauses before responding.

HighLevel's own guide to creating a Voice AI agent is specific about the prompt itself: it should define the agent's role and primary objective, the information it should collect, the topics it may answer, and the topics or claims it must avoid.

Basic Mode or Advanced Mode

Early in the builder you'll choose between two levels of control.

Basic Mode keeps it simple. You pick a knowledge base and tell the agent which details to collect from a caller, such as their name or the reason for the call. There's no prompt writing involved at all.

Advanced Mode hands you the full prompt, plus call transfer logic, workflow triggers, and the rest of the actions below. A simple intake line runs fine on Basic Mode. Anything that needs to route calls, update records, or book appointments needs Advanced.

Connecting a knowledge base so it stops guessing

An agent without a knowledge base either stays silent outside its script or, worse, answers anyway. Attaching one gives it something real to pull from.

Under the agent's Knowledge Base section, you select which base it can access, then add or update the sources behind it: business hours, pricing, and policies.

You also tell the agent when it should bother searching the knowledge base rather than answering straight from the prompt. Then test it against edge cases and the odd questions real callers ask, not just the ones you expected.

This is where agents fail quietly. A prompt that reads well in a demo still invents an answer the first time someone asks something the knowledge base doesn't cover.

Giving the agent something to do

A Voice AI agent that only talks is a glorified voicemail greeting. Actions let it change something in your CRM while the caller is still on the line.

Call Transfer sends the caller to a person or an external number. Agent Transfer hands the conversation to a different Voice AI agent instead, useful when one agent screens calls and another handles a specific department.

Trigger a Workflow launches whichever automation you've already built with GoHighLevel's workflow builder, whether that's a follow-up sequence or an internal alert.

Send SMS and Update Contact Field let the agent text the caller or save what it just learned straight onto the record you set up during your CRM setup.

Appointment Booking puts the call directly onto your calendar's open slots. A Custom Action can call out to a webhook when you need to reach something outside GoHighLevel entirely.

Each action needs a sharp trigger condition, not just a destination. An agent set to transfer "when the caller sounds frustrated" without a tighter definition will transfer far more calls than you intended.

Assigning a phone number and routing the calls

Photo by cottonbro studio on Pexels

None of the above does anything until the agent is deployed. That happens on the Deploy tab, where you assign the agent to an eligible phone number or a number pool.

Call routing gives you two modes. Answer Calls Directly means the agent is the first thing a caller hears. Use as Backup means it only picks up when your primary call path, usually a person, doesn't answer.

You also set working hours, either all hours or a custom schedule, and HighLevel is upfront that assigning a number doesn't guarantee the agent answers every call. Forwarding rules, timeouts and voicemail settings can all still get in the way first.

If you'd rather have the agent reachable from your website than a phone line, the Voice AI chat widget runs the same agent in a browser instead, which is worth weighing against a text-based chatbot built for your site.

Turn it on under Settings, then Labs, then the sub-account tab, pick the agent, and paste the generated script into your site's tracking code.

Visitors tap a microphone icon to start talking, and the widget shows a recording notice automatically. A single sub-account can run up to 100 of these agents, so testing a few variations side by side costs nothing structurally.

Testing it before a real caller finds the gaps

Save the configuration and you get two ways to test it. A Web Call runs in the browser but doesn't support call transfers. Phone Call testing uses the full telephony path instead, transfers included.

Run both against the same scenarios: the questions you expect, the ones the knowledge base should answer, the ones the agent should refuse, and a few unclear or incomplete requests.

Then test what happens when you go quiet mid-call, when you interrupt the agent while it's talking, and when the call just ends.

Before flipping it live, work through the checklist HighLevel builds the testing flow around:

1. The greeting sounds natural, not read off a script.
2. Knowledge base answers are accurate against your actual sources.
3. The agent declines rather than invents when it doesn't know something.
4. Every action fires only under the condition you set for it.
5. A real external call actually reaches the deployed agent.

Skipping that last line is how an agent passes every test in the browser and then never rings when someone dials the real number.

What it costs on top of your plan

Voice AI isn't a separate product you buy. It rides on whichever GoHighLevel plan you're already on, with usage billed in addition to the base price.

GoHighLevel's pricing page lists three core plans, plus two optional add-ons that bundle AI usage instead of billing it as you go.

Plan or add-on	Monthly price	What it buys
Starter	$97/mo	3 sub-accounts, core CRM, AI usage billed separately
Unlimited	$297/mo	Unlimited sub-accounts, phone and email rebilling without markup
Agency Pro	$497/mo	SaaS Mode, rebilling with markup, deeper API access
AI Employee Growth (add-on)	$50/mo per sub-account	Bundled usage of GoHighLevel's AI Employee tools, including Voice AI
AI Employee Unlimited (add-on)	$97/mo per sub-account	Same AI Employee bundle, higher usage tier

Without one of those add-ons, Voice AI minutes get billed individually against your account's usage balance instead. A 14-day free trial covers every plan, so you can build and test an agent before any of this applies.

Writing the prompt and action logic well enough that the agent is worth deploying is the part that takes real time. If that's not where your team wants to spend it, it's the kind of build our AI & Automation team takes on directly, inside whichever CRM you're already running.

Frequently asked questions

Can I run Voice AI without using a phone number at all?

Yes. The Voice AI chat widget puts the same agent inside your website instead of on a phone line, and visitors start the conversation by clicking a microphone icon in the browser.

Can the agent book appointments directly onto my calendar?

Yes, through the Appointment Booking action in Advanced Mode. It checks your calendar's open slots and books the call in without anyone touching the schedule by hand.

Does it work with workflows I've already built in GoHighLevel?

Yes. The Trigger a Workflow action lets the agent launch any existing workflow mid-call, from a follow-up sequence to an internal notification.

What happens if the agent can't answer a question?

That depends on how it's configured. The pre-launch checklist specifically asks you to confirm the agent declines rather than guesses, and a call transfer action can hand the question to a person instead.

Cover photo by Yan Krukau on Pexels