GoHighLevel AI Agent Studio: Setup & Pricing (2026)
GoHighLevel AI Agent Studio: Setup & Pricing (2026)
By Dr Priya Jaganathan, GoHighLevel Certified Admin · HL Growth Partner, Australia · Updated 27 August 2026 · 8 min read
GoHighLevel AI Agent Studio is the build layer where you configure a single AI agent once — goals, knowledge base, guardrails, tone, escalation rules — and then point it at SMS, web chat, Facebook and Instagram DMs, or voice. It sits above the older bot settings rather than replacing them. Cost is usage-based on top of your plan: most agencies I work with land somewhere between two and six cents per text conversation and roughly a dollar to a dollar-fifty per connected voice minute, billed out of agency wallet credits. If you run one sub-account and answer every enquiry yourself within ten minutes, you do not need it. If you run fifteen sub-accounts and leads sit unanswered overnight, it pays for itself in a fortnight.
What follows is how I actually set these up in client sub-accounts, not the marketing version. I will cover where the Studio lives, how it differs from Conversation AI and Voice AI, how to build an agent that does not embarrass your client, how to hand off cleanly to Workflows and appointment booking, what the usage charges really look like on an invoice, and how to roll the whole thing across a portfolio using snapshots. I will also flag the mistakes that cost me time so you can organise your build around them instead of discovering them at 11pm the night before a client demo.
What GoHighLevel AI Agent Studio actually is
Think of AI Agent Studio as an agent factory rather than a channel setting. Before it existed, you configured a bot inside each channel — Conversation AI had its own training tab, Voice AI had its own prompt box, and the two never shared anything. Every time a client changed their pricing you edited the same information in three places and forgot one. The Studio inverts that. You define the agent as an object: its role, its objective, the knowledge base it reads from, the actions it is permitted to take, and the conditions under which it stops talking and hands the conversation to a human. Channels then subscribe to that agent.
How it differs from Conversation AI
Conversation AI is still the underlying text engine for SMS, web chat and social DMs, and it still has its own training modes. The practical difference is scope and control. Conversation AI answers questions and books appointments within one channel using intents and training URLs. An agent built in the Studio carries a persistent objective across channels, can call multiple actions in sequence, and gives you far more granular guardrails. If your build is a simple FAQ responder on a website widget, the older setup is genuinely fine — my GoHighLevel Conversation AI setup and training guide walks through that path end to end and it will get most local service businesses answered inside sixty seconds.
How it differs from Voice AI and AI Employee
Voice AI handles inbound and outbound calls through LC Phone or your own Twilio account, with its own latency constraints and its own per-minute billing. AI Agent Studio can drive a voice agent, but the voice pipeline — speech to text, response, text to speech — is still Voice AI underneath. If you are weighing which channel to build first, the trade-offs are laid out in GoHighLevel Voice AI vs Conversation AI. AI Employee, meanwhile, is the commercial bundle: it is the SKU you enable at agency level that switches on the AI features, not a separate product you build in.
Where to find it and how to enable it
Agency level first
Nothing works at sub-account level until the agency enables it. In your agency view, open Settings, then the AI section, and switch on the AI features for the accounts you want. You will also need billing configured — a card on file and, in most cases, a topped-up wallet with auto-recharge turned on. I set auto-recharge deliberately low on new builds, around fifty dollars with a hundred-dollar top-up, so a runaway agent hits a wall rather than a four-figure invoice. You can rebill usage to clients at agency level with a markup; decide your margin before you switch anything on, not after the first invoice arrives.
Sub-account level
Inside the sub-account, AI Agent Studio appears in the left navigation under the AI section. Before you build anything, get the plumbing right: LC Phone or Twilio connected and a number purchased, Mailgun connected if the agent will trigger email, and A2P 10DLC registration completed for US numbers. Australian clients do not go through 10DLC, but ACMA rules on consent and identification still apply — an AI agent sending unsolicited SMS is the same breach as a human doing it. Confirm the business hours and timezone on the sub-account too, because an agent that books 3am appointments is a fast way to lose a client.
Building your first agent in HighLevel AI Agent Studio
Start with one narrow goal
The agents that work have one job. "Qualify inbound enquiries from the Google Ads landing page and book a fifteen-minute consult" is a goal. "Handle customer service" is not. Write the objective in plain language, name the single success condition, and resist adding a second job until the first one runs clean for a fortnight. I usually set the agent to collect three or four data points — service required, suburb, urgency, budget range — and write each to a custom field so the sales team sees structured data rather than a transcript.
Knowledge base
Feed the knowledge base actual source material: the services page, the pricing page, the FAQ, a short document of things staff say on the phone. Do not dump the entire website. Every irrelevant page increases the chance the agent quotes a 2019 promotion. Ask the client one direct question — what are the five things people always ask before they book? — and make sure those five answers exist in the knowledge base verbatim. Re-crawl whenever pricing changes, and diarise a quarterly review, because stale knowledge bases are the single most common cause of an agent confidently saying something wrong.
Guardrails, tone and escalation
Guardrails are where practitioner builds separate from demo builds. Set explicit prohibitions: never quote a fixed price, never discuss competitors, never promise a timeframe, never give clinical or legal advice. Cap the conversation length — if the agent has not reached its goal in six or seven exchanges, it should stop and hand over. For tone, give two or three example replies written the way the client actually speaks rather than adjectives like "friendly and professional", which produce identical bland output every time.
Escalation needs a real destination. Configure the agent to add an ai-escalated tag, assign the conversation to a user, and send an internal notification. Trigger escalation on frustration signals, on any mention of a complaint or refund, on repeated questions, and on an explicit request for a human. Then check that someone is genuinely watching that inbox. An escalation path that leads to an unmonitored conversation is worse than no agent at all.
Connecting channels, Workflows and booking
Choosing channels
Attach the agent to one channel, prove it, then add the next. SMS is the usual starting point because response rates are high and the failure modes are visible. Web chat is second. Facebook and Instagram DMs need the pages connected and permissions re-authorised after any Meta password change — check this monthly, because expired tokens fail silently. Voice comes last, always, because latency and interruption handling need genuine tuning. If outbound calling is on the roadmap, read GoHighLevel Voice AI outbound calls setup before you point an agent at a cold list.
Handing off to Workflows
The agent should decide and the Workflow should execute. When the agent completes its goal it writes to custom fields, applies a tag, and that tag fires a Workflow trigger that creates the opportunity in the right pipeline, notifies the owner, and starts the nurture. Keeping the logic in Workflows means it is testable, auditable and portable into a snapshot. If you are structuring this properly, the patterns in my write-up on GoHighLevel AI workflows will save you rebuilding the same branch three times.
Appointment booking
Give the agent access to one calendar with sensible availability, buffers and a minimum scheduling notice of at least two hours. Use custom values for the business name, address and booking link so the same agent behaves correctly in every sub-account after a snapshot import. Always confirm the booking with a separate SMS from the Workflow rather than relying on the agent's last message — clients want a record, and so do you when someone claims they were never booked.
Pricing and usage charges in plain terms
Everything here stacks on top of your agency plan. The plan gives you access; usage is metered separately and drawn from your agency wallet. In practice you are paying for three things: the AI conversation or per-minute charge, the underlying LC Phone or Twilio carrier cost for the SMS segments and call minutes, and any email sending through Mailgun. People budget for the first and forget the other two, then wonder why the wallet drains faster than expected.
As a working estimate, budget a few cents per text-based conversation and roughly a dollar to a dollar-fifty per connected voice minute, plus carrier costs. A sub-account handling three hundred inbound text conversations a month typically sits under twenty dollars of AI usage; a busy voice agent taking two hundred calls averaging three minutes is a different order of magnitude entirely. Confirm current rates on the official GoHighLevel pricing page and in the HighLevel help documentation before you quote a client, because these numbers move. For how the bundled SKU is structured, see GoHighLevel AI Employee pricing and what's included.
| Feature | What it does | Best use case | Channel | How it's billed | Setup effort |
|---|---|---|---|---|---|
| AI Agent Studio | Builds reusable agents with goals, knowledge base, actions and guardrails | Multi-channel qualification and booking across a portfolio | SMS, web chat, FB/IG DM, voice | Usage-based per conversation or per minute from agency wallet | High — half a day per agent done properly |
| Conversation AI | Answers text enquiries and books using intents and trained content | Single-channel FAQ and booking bot | SMS, web chat, social DM | Per message or conversation | Low to medium |
| Voice AI | Handles live inbound and outbound phone conversations | After-hours call answering and missed-call recovery | Voice via LC Phone or Twilio | Per connected minute plus carrier cost | Medium to high — tuning required |
| AI Employee | Commercial bundle enabling the AI feature set at agency level | Agencies rebilling AI usage to clients | All of the above | Monthly add-on plus usage | Low — billing configuration only |
Testing and rolling out across sub-accounts
Test like a difficult customer
Build in one sandbox sub-account and test from your own phone. Run the happy path, then deliberately break it: misspell things, ask about a service the client does not offer, ask for a discount, go silent mid-conversation, say "just call me". Read the full transcripts. Nine times out of ten the fix is a guardrail sentence, not a new tool. Only after twenty or thirty clean conversations do I turn an agent loose on real leads, and even then I read every transcript for the first week.
Snapshots and portfolio rollout
Once an agent is stable, capture it in a snapshot along with the Workflows, tags, custom fields, custom values and calendar it depends on. Snapshot behaviour for AI configuration has improved but is not perfect, so keep a written build checklist of anything that does not carry across — phone number assignment, knowledge base re-crawl, calendar reconnection, Meta page permissions. Import into one live sub-account, verify against the checklist, then batch the rest. If you are also scoring the leads these agents produce, GoHighLevel AI lead scoring pairs neatly with this and stops the sales team chasing everyone equally.
Measuring whether it is actually working
Track four numbers per sub-account: time to first response, conversation-to-booking rate, escalation rate, and cost per booked appointment. Time to first response should collapse from hours to seconds — that alone usually lifts booking rates. A conversation-to-booking rate under about ten per cent on warm inbound traffic means the agent is talking too much or the calendar is too restrictive. An escalation rate above roughly a quarter means the knowledge base has holes. Cost per booked appointment is the number to show the client: if the agent costs eighteen dollars a month and produces nine bookings, nobody argues about the invoice. Review monthly, adjust the knowledge base, and keep the agent's job narrow.
Common mistakes to avoid
- Switching an agent live before A2P 10DLC registration is approved, or before checking ACMA consent obligations for Australian sub-accounts — messages fail silently or expose the client to a complaint.
- Leaving auto-recharge uncapped on the agency wallet, so a misconfigured voice agent or a spam wave burns hundreds of dollars overnight with no alert.
- Loading the entire client website into the knowledge base instead of the five pages that answer real pre-booking questions, then wondering why the agent quotes expired promotions.
- Building escalation that adds a tag but never assigns the conversation or notifies a human, so escalated leads sit in an unwatched inbox for days.
- Letting the agent make decisions that belong in Workflows — pipeline moves, opportunity creation and nurture sequences should be triggered by tags, not improvised inside the conversation.
- Rolling a snapshot across fifteen sub-accounts without a post-import checklist, then discovering the calendar link, phone number and Facebook permissions never carried over.
If you want your GoHighLevel AI agents built, tested and rolled out across your sub-accounts without burning a month of trial and error, book a strategy call with the HL Growth Partner team.
Frequently asked questions
Is GoHighLevel AI Agent Studio included in my plan or does it cost extra?
Access is tied to your agency plan and the AI feature set being enabled, but the actual usage is charged separately from your agency wallet. You pay per text conversation or per connected voice minute, plus the underlying LC Phone or Twilio carrier cost. Budget the usage as a variable expense, decide your rebilling markup before launch, and cap auto-recharge so a misconfiguration cannot run away.
Do I still need Conversation AI if I use AI Agent Studio?
Conversation AI remains the text engine underneath, so you are not choosing one over the other in a technical sense. What you are choosing is where the configuration lives. For a single-channel FAQ and booking bot in one sub-account, the older Conversation AI setup is quicker and perfectly adequate. For multi-channel agents rolled across a portfolio, build in the Studio so you maintain one definition instead of three.
How long does it take to build an agent properly?
Allow half a day for the first agent in a new sub-account: an hour on the objective and knowledge base, an hour on guardrails and escalation, an hour wiring Workflows, custom fields and the calendar, and the rest on testing. Subsequent sub-accounts running from a snapshot take under an hour, mostly spent reconnecting phone numbers, calendars and Meta page permissions.
Can the agent book appointments without a human checking?
Yes, and it works well provided the calendar has buffers, a minimum scheduling notice of at least two hours, and correct business hours and timezone on the sub-account. Always send a confirmation SMS from a Workflow rather than relying on the agent's closing message, and notify the assigned user immediately. That gives both sides a record and catches double bookings early.
What happens when the agent does not know the answer?
With guardrails set correctly it should say it will get a human to help, apply an escalation tag, assign the conversation and notify staff. Without those guardrails it will guess, which is how clients end up honouring prices nobody offered. Configure escalation on frustration signals, complaints, refund requests and explicit requests for a person, then audit transcripts weekly.
