Build your own AI agent for GoHighLevel

Full video walkthrough

A 56 minute recording of a complete build, start to finish. The player has chapters, so you can jump straight to the part you are working on instead of watching the whole thing.

The written sections below cover every field in the same order the video moves through them, so you can follow along or skip ahead to the one you need.

1. The AI dashboard

Create a new AI and pick how you want to build it: co-pilot, an imported JSON config, or from scratch. Co-pilot is the fastest way to get a working agent. JSON is best when you already have a build you like. From scratch gives you full control over every field. Once you are inside, the builder is split into the sections below.

1.A. Strategy name, agent role, tag, and tone

The top of the strategy builder is small but it decides whether the agent ever talks to the right people. The LeadConnector tag is the piece people get wrong most often.

1.B. Timing and delays

Two different delays exist and they fire at different moments. The initial message delay applies only when the AI starts the conversation with proactive outreach, and a random wait within your range is taken before the first outbound message is sent. The response delay is the wait before the agent replies after a contact responds, and a fresh random wait is taken each time.

1.C. Qualification, objection handling, and creativity sliders

Objection handling, qualification priority, and creativity are not decorative sliders. They do an extremely important job. Increasing them determines how much the AI will objection handle versus letting things go, how much it will qualify leads versus letting them continue on in the conversation, and how much it is allowed to be creative versus sticking to your exact wording outside of scenarios.

1.D. The prompt editor

The prompt editor is the basic, freeform area where you can put broad instructions together, such as your agent's role and extra context. Do not repeat facts already in your knowledge base. Inside the prompt editor you can select character limits, and we suggest 160 character limits for iMessage lines, the language, typos per conversation to sound more human, and personality traits. These personality traits are in addition to the tone of the AI you select in the main instructions area.

Structured prompting takes priority over this editor. The editor is not the place to rely on for a detailed conversation plan. Put specific objectives, rules, scenarios, and phases into their structured sections so the AI knows exactly what to focus on.

1.E. Structured prompting

Structured prompting is a detailed layout for business owners that have specific lead context, phases of conversation, keywords used in conversations, rules the AI must follow, or scenarios the AI needs to pay attention to. The layers build in order: primary objectives, lead context, trouble points, rules, scenarios, keywords and affirmations, then phases.

The AI prioritizes these structured sections over the regular prompt editor. Each section gives the AI detailed instructions for that part of the conversation. If instructions in the two areas conflict, structured prompting takes priority. Use the regular editor for broad context and these sections for the specific behavior you want. Remember that high qualification priority can still cause the AI to largely ignore the rest of your prompting.

Primary objectives

Primary objectives need to be filled out in the order that the AI should try to hit them. Qualify the lead then book an appointment. Or build rapport, qualify the lead, book an appointment, then send a YouTube video.

Objectives are black and white: the exact step the AI has to hit. Each one also gets a matching phase with a little more context, so the AI has the overview of the conversation as well as the step itself.

Lead context

Lead context is important because it helps the AI understand whether the lead is hot, cold, stubborn, already booked, and so on. The same message lands very differently depending on the answer, so tell the AI up front.

Trouble points

Trouble points help the AI know where leads are going to get tripped up, or where they will try to trip up the AI. Pricing is the most common example. Give the AI some context on how to navigate these trouble points rather than only naming them.

Rules

Rules are must haves the AI needs to follow. They are not conditional. They are guardrails and guidelines the AI must follow every single time before it creates a response.

Best practices for a human sounding AI

Your AI should sound like a person texting, not a formal assistant. Add these default rules in Step 1 Instructions, under Rules. They apply throughout the conversation, not just to your first message.

1. 1. Speak at a fifth grade level

   Use short sentences and everyday words. Say “What do you need help with?” instead of “Could you elaborate on your requirements?” Simple does not mean childish. It means easy to read.

   Simple language rule

   Write at a fifth grade reading level. Use simple words and short, clear sentences. Avoid jargon and formal language.

2. 2. Keep emojis optional and natural

   If you want emojis, tell your AI to use them. Each emoji should fit the sentence and the feeling behind it. Use no more than one per sentence and only a few across the whole conversation. Do not force an emoji into every reply.

   Optional emoji rule

   Use emojis sparingly. Match each emoji to the meaning and feeling of the sentence. Use at most one emoji per sentence and only a few per conversation.

3. 3. Keep messages short and punctuation simple

   Keep every message under 160 characters, including spaces and punctuation. Avoid dashes, hyphens, and excessive punctuation. A simple question mark is enough. You do not need “!!!” or “???” to make a message feel friendly.

   Message length and punctuation rule

   Keep each message under 160 characters. Never use dashes or hyphens. Use normal punctuation without repeated question marks or exclamation marks.

4. 4. Skip routine acknowledgments

   You do not need to acknowledge everything your lead says. Starting every reply with “Got it,” “Absolutely,” or “Thanks for sharing” makes your AI sound scripted. Answer the question or move to the next useful question without repeating their answer.

   Acknowledgment rule

   Do not acknowledge or repeat every lead response. Skip routine acknowledgment phrases and move the conversation forward naturally. Still answer their questions and respond to what matters.

   Sounds scripted

   Lead: “I need help with my roof.”
   AI: “Got it, thanks for sharing that you need help with your roof! How old is it?”

   Sounds natural

   Lead: “I need help with my roof.”
   AI: “How old is your roof?”

Test these rules in a full conversation before you launch. Check that your AI stays short, uses simple words, and moves naturally between questions instead of falling back into the same phrases.

Scenarios

Scenarios are conditional and only fire when their condition is met. If you do not explain a scenario correctly it may not fire, so write your scenarios with intent. Scenarios let you deliver the wording word for word with verbatim, or make it act like a rule by selecting always check, which means the agent should always check whether the lead asked this. A widely used example is: if a lead wants a phone call instead of booking a consultation, then let them know that someone will be reaching out to them and do not ask any more questions to the user. Keep verbatim toggled off and always check toggled on.

Keywords and affirmations

These are important for making the AI talk exactly how you want, with the language, tonality, and slang you desire. The keyword might be Cool, and you can write an affirmation that says That's cool! which gives the AI extra context on how to use these words inside sentences.

Phases

Phases match your primary objectives, one phase per objective, in the same order. Your objectives are black and white: the exact step the AI has to hit. Phases say the same thing with a little more context, so the AI has an overview of the conversation to follow.

If objective 1 is build quick rapport, phase 1 is build quick rapport with the lead, plus the extra detail on how that step should look and feel. Do not invent a phase that has no objective behind it and do not skip one. When you change an objective, update its phase so the two still line up.

Phases are also different from qualifying questions, which ultimately lead the conversation. The qualifying questions lead, but phases are how the AI knows which step of the process it is currently in and where it needs to get to next.

2. How the conversation actually runs

Once the prompting is in place, the agent runs the same order for every tagged contact in GoHighLevel: initial message, reply and rapport, qualification questions, then the call to action. Follow ups sit behind that as a backup. They are for the lead that never responds, or the conversation that never makes it to the call to action. You control each stage.

2.A. Initial messaging

The initial message is the first thing the AI sends. Provide 5 variations for testing and the AI cycles through them for every new lead. Enable Spintext to rotate the wording so repeated sends do not get caught by spam filters, and so you can test different hooks against each other. Contact fields and custom fields from GoHighLevel work inside initial messages and throughout your prompting.

2.B. Qualification questions

Qualification questions are the main driver of the conversation. They are the prompts the AI uses to steer the thread, asked in the order you list them, and you can add as many questions or routes as you need. If a guide is not needed, disable it with the toggle on the right. The qualification priority slider decides how seriously the AI treats them. At high or maximum priority, getting these questions answered correctly becomes almost the AI's only goal, and the rest of your prompting is largely ignored. Keep the priority lower if you want your phases and other instructions to guide the conversation too.

Each question has its own toggles. Required means the AI must get an answer before moving on, and turning it off lets the AI skip the question if the lead does not answer. Accept refusal means a lead declining to answer still counts as satisfying the question, so the AI moves on instead of pressing. Conditions let you branch a question into different routes.

2.C. Smart follow ups

Follow ups are the backup path. They only come into play when the lead never responds, or when the conversation never reaches your call to action. They are automated messages sent when a lead has not responded. They send in the order sequence you set, and only after the specified delay has passed since the lead's last response, so follow ups are timer based. They can be spin texted too, which means the AI slightly changes what you wrote so the same line is not sent every time, and you can review those variations yourself.

Use one to three follow ups, not a long stack. Everything goes over iMessage, and a line pushed with too many messages works against you instead of winning the lead. Every follow up you write should be trying to trigger a response, so it earns a reply rather than just checking in. Follow ups do not count toward your daily lead limit, so the limit is never what stops you. The reason to stop is that a lead who has not answered after a few short nudges needs a different approach, not more messages.

3. Configuring the bot

With the prompting and conversation flow set, the configuration layer decides how messages are delivered, what outcome the agent drives toward, and when it should stop.

3.A. Multi message

Multi message is a toggle that lets the AI break longer replies into short sequential messages instead of one block, so the thread feels more natural.

3.B. End goal, calendar, and appointment title

The end goal is the primary outcome the agent should drive toward, such as booking an appointment or a custom call to action, and downstream logic adjusts accordingly. Choose the LeadConnector calendar the agent should use for availability and booking.

The custom appointment title sets the title the appointment gets on your calendar, and in confirmation emails if those are enabled. Leave it blank to use the calendar's own title. It supports merge tags like {{contact.firstName}} and {{contact.lastName}}.

3.C. Disqualification trigger

The disqualification trigger is important whether you want leads formally disqualified or you simply want the outreach to stop. Add the keywords or phrases that signal a lead is lost or has opted out. When one is detected, outreach stops and the lead is marked as disqualified.

4. The knowledge base

The knowledge base is where everything about your company should live, along with the answers to factual questions. Your website can be pasted and scraped. You can upload text documents and PDFs of your past human conversations. You can add canonical facts so the AI knows exactly what to say and when, and attach dates so it can reference when those changes or facts occurred.

Knowledge is segmented on purpose. Each section is read differently by the agent, so putting content in the right place matters more than the volume you load in.

4.A. Company information

This gives the AI context on your brand, services, audience, and key value props so replies stay accurate and on brand.

4.B. FAQ

Common questions and answers the AI can reuse. Keep the responses concise and factual. For conditional objection handling, the kind where the lead says X and you want response Y, use scenarios in Step 1 then Instructions instead of the FAQ.

4.C. Knowledge sources

Upload documents, add website URLs, or jot down notes. The AI uses the extracted text as extra context when replying to leads. We recommend uploading 200 of your best conversations in a correctly titled list, plus your company help docs.

4.D. Canonical facts

Founder authored source of truth facts, stored as a key and a value, that the AI must treat as authoritative. Set an optional effective date and a status to control when a fact applies.

5. Automations and nodes

Automations tell the AI how to stop, start, book, cancel, reschedule, notify, and update contacts. You attach workflow tasks the agent can trigger, such as zap steps and CRM updates, and tasks are selected from the current sub account. Every automation is tracked with analytics and can be added to or swapped between agents.

Each automation you add shows up as an AI task on the automations step, and opening one drops you into the canvas where you drag nodes out of the palette and connect them under a trigger.

5.A. Triggers

The trigger works like a GoHighLevel trigger except there are no static fields. You tell the AI exactly when it should fire in your own words, which makes it fully dynamic, then attach the nodes you want to run after it. A trigger does not disable the AI by itself, so your team can be notified while the agent keeps handling the conversation.

Set your trigger to match your final primary objective. Your last objective is the thing you want the lead to agree to or do, so your trigger describes that same moment. If your final objective is getting the lead to agree to a booked consultation, your trigger reads run when a lead agrees to a booked consultation. When the two match, the automation knows exactly when to fire. When they do not, the booking node waits for a moment the conversation never describes.

5.B. AI custom actions

Multi step chain lets you tell the AI exactly what to do with the lead, such as collecting everything and splitting it into a conversation summary, the pain points, and what the closer could say on the sales call to win the lead over. Assign to agent hands one AI off to another with no human intervention, which is how a booking agent passes the thread to a call confirming agent.

5.C. CRM actions

Add tag labels the lead after the automation runs, which opens up everything inside GoHighLevel: pipelines, other workflows, rep notifications. Generate summary is built to deliver a clean summary for your sales reps into any custom or contact field. Update custom field and update contact field log answers in real time, so if a lead says to use their new email the AI updates it on the spot.

5.D. Calendar nodes

Calendar nodes are how you segment leads across your sales team. Book appointment has no controls: you attach it under a trigger like run when a lead wants to book an appointment and the backend reads the calendar, suggests times, and gets the lead booked. Update calendar provider can sit inside the book appointment node to switch to a specific rep's calendar, so one agent books multiple reps instead of you building an agent per rep. Reschedule booking and cancel booking work the same way, each in its own automation with its own trigger.

5.E. Control center

Agent status lets you activate the AI, disable it, or turn off follow ups while keeping the agent on for customer support. Trigger, book appointment, generate summary, disable AI shuts the agent off right after the booking. Swap the last step for disable follow ups and the agent stays available for questions without chasing the lead. To hand off, extend it with assign to agent and activate AI.

5.F. Notifications

Notifications alert sales teams by email or phone number. Most clients handle this inside GoHighLevel instead by adding a tag in the automation and building a GHL workflow that notifies the team whenever that tag lands on a lead.

6. Building specific agents

Once the strategy is built, the next step is wiring it to real outcomes inside GoHighLevel. This section walks through a complete booking agent and the GHL workflow that fires it.

6.A. Building a booking agent

Build the agent the same way as any other strategy, but keep the primary objective focused on a booked appointment. In the end goal, select Book appointment, choose GHL as the calendar type, then pick the calendar from the list that syncs from LeadConnector.

In the automations section, create a new automation and name it Booked Appointment. Drag a Trigger onto the canvas and write a prompt like Run when a user chooses a time to book an appointment. Then drag the Book Appointment node underneath the trigger and connect them. It does not matter if the meeting is a Zoom, Google Meet, or phone call, the node follows whatever is configured inside your GHL calendar.

6.B. Making sure the AI fires correctly from GHL

Inside GoHighLevel, create a new workflow or add this to an existing one. The workflow trigger decides if the lead should enter. The first action should add the same tag that is inside the AI agent. After that, add the Sendify node by typing Sendify into the GHL actions panel and selecting Send Initial Message.

In the Sendify action, map the contact phone number, the full name, and the same tag you added in the previous step. Some workflows also include a Toggle Agent step before Start Sendify AI to make sure the agent is active. Save the automation. Every lead that enters this GHL workflow will get the tag and be fired into the AI with that tag.