GoHighLevel Conversation AI: Setup & Training (2026)
GoHighLevel Conversation AI: Setup & Training (2026)
By Dr Priya Jaganathan, GoHighLevel Certified Admin · HL Growth Partner, Australia · Updated 21 August 2026 · 8 min read
GoHighLevel Conversation AI is the built-in bot that reads inbound messages in your Conversations inbox and replies for you, either answering questions or pushing a lead onto a calendar. It runs on SMS, Live Chat, Facebook, Instagram and WhatsApp, is configured per sub-account, and is billed as usage rather than a seat fee — you pay per message the bot generates, drawn from your wallet or your plan's AI allowance. A few hundred bot replies a month is a small dollar figure, not a hundred-dollar line item.
The honest boundary: it works beautifully on high-volume, low-variance inbound — "are you open Saturday", "how much is a service", "can I book Thursday" — where the answer is knowable and the outcome you want is a booking. It embarrasses you when it goes live on Auto-Pilot with a thin knowledge base, no sleep hours and no handoff rule, because it will invent a price, promise a tradesperson who does not service that postcode, or text someone at 2am. Below is the setup order I use on client builds: prerequisites, training, guardrails, testing, measurement.
What Conversation AI actually does inside HighLevel
HighLevel Conversation AI sits between the inbound message and your team. When a contact replies, the bot checks it is enabled for that channel, waits out the message delay, matches the message against its training material and Bot Goal, then either sends the reply (Auto-Pilot) or drops a draft into the inbox for a human to approve (Suggestive). It is not a workflow engine and not a phone agent — for outbound calling see the GoHighLevel Voice AI guide.
The most important setting is the Bot Goal. Appointment Booking steers relentlessly toward a calendar slot; Query & Answer answers and stops. Choosing wrong is the commonest misconfiguration I see: a clinic sets Query & Answer, the bot answers perfectly, the lead says "great, thanks" and never books. SMS replies go out through LC Phone or Twilio, so every A2P and sender-ID rule that applies to normal SMS applies to the bot too.
Prerequisites before you switch it on
Build the plumbing first. A booking bot with a broken calendar is worse than no bot at all.
- A working calendar — correct time zone (Australia/Sydney, Brisbane and Perth are not interchangeable), realistic availability, buffers, minimum notice of a few hours, and confirmation plus reminder workflows already firing.
- Business hours defined at account level, matching the hours a human could plausibly follow up.
- A tagged pipeline with stages the bot's outcomes map to — New Enquiry, Bot Engaged, Booked, Needs Human. Without it you cannot measure anything.
- Custom fields for what you want captured (suburb, service type, preferred day). The bot can collect them only if they exist first.
- Registered SMS sending — A2P completed, sender details correct, opt-out handling live. Sort your SMS A2P compliance before the bot sends a single text.
- A chat widget that captures a phone number if you are running on Live Chat — the GoHighLevel chat widget setup covers the fields that keep conversations attached to a contact record.
Bot Goal and mode configuration compared
Pick one row per axis — goal, mode, channel set.
| Configuration | Best for | Risk | What to set |
|---|---|---|---|
| Bot Goal: Appointment Booking | Lead-gen — trades, clinics, brokers, agencies | Pushes for a booking when the person only wanted a price; double-books if buffers are wrong | One calendar, minimum notice 4+ hours, capped booking attempts, 2–3 questions answered before it pivots |
| Bot Goal: Query & Answer | Support inboxes, e-commerce, memberships | High reply volume, near-zero bookings — a very expensive FAQ page | Tight training set, instruction to offer the booking link on commercial intent, handoff on pricing or complaints |
| Mode: Auto-Pilot | Mature bots with four weeks of reviewed transcripts and a narrow subject area | Wrong answers ship instantly with no human in the loop | Only after Suggestive review; keep sleep hours, handoff triggers and a 1–3 minute delay on |
| Mode: Suggestive | Every new build, weeks one to four, and any regulated advice | Slow responses if nobody watches the inbox; staff rubber-stamp drafts unread | A named inbox owner, every draft reviewed, edits logged as training data |
| Channels: SMS + Live Chat | Highest-intent inbound; where speed genuinely converts | SMS carries real legal exposure; chat sessions expire and threads orphan | SMS only where consent is recorded; capture name and mobile before the bot engages |
| Channels: Facebook, Instagram, WhatsApp | Real social DM volume with a community manager in place | Messaging windows close; DMs are often existing customers, not leads | Enable one social channel at a time, shorten replies, route complaints to a human |
Training the bot properly
Training is where results are made. The bot draws on three sources: URLs you crawl, FAQs written as question-and-answer pairs, and the personality and instruction fields.
URLs: crawl narrowly
Feed it services, pricing, about, service-area and genuine FAQ pages. Do not crawl the whole domain — blog archives and expired campaign pages are the number-one source of a bot quoting a price that ended in 2023. Re-crawl after every pricing change; the bot does not know your site updated.
FAQs: write them like a receptionist would answer
Thirty to fifty tight Q&A pairs beats a two-hundred-page crawl. Pull the real questions from your Conversations inbox. One or two sentences each, with specifics — suburbs serviced, hours, "from" prices, what is not offered. Include explicit negative answers: "Do you do commercial work? No, residential only, metro Melbourne." Negative answers stop hallucination better than any prompt instruction.
What NOT to feed it
Anything you would not want read aloud to a customer: internal SOPs, staff rosters, supplier costs, margins, discount authority, terms you cannot honour, and any advice needing a licensed human. Leave out competitor comparisons — they invite arguments the bot cannot win. Generative range for marketing copy belongs in Content AI, not your customer-facing bot.
Tone
Be specific in the personality field: "Friendly Australian receptionist. Short sentences. No exclamation marks. Never says 'I'm just an AI'. Never guesses — if unsure, offers a call back." Generic instructions produce generic American-sounding replies, and Australian customers notice.
Intents and fallback behaviour
Intents are how the bot categorises what a person wants: booking, pricing, hours, location, complaint, existing customer, spam. Map each to a defined behaviour rather than leaving it to the model. Booking gets the calendar, pricing gets a range plus a booking offer, complaints get an apology and immediate handoff. Anything unrecognised triggers the fallback.
Your fallback must acknowledge, promise a human and set a time expectation — then fire a workflow that tags the contact, creates a task and notifies the team. A fallback that says "I don't understand" and stops is a lost lead. Set a two-strike limit: after two consecutive fallbacks the bot disables itself for that contact and hands over.
Message delay, sleep hours and human handoff
Message delay. Instant replies read as robotic, and the bot can respond mid-thought when someone sends three texts in a row. I use 1–3 minutes on SMS and 10–30 seconds on Live Chat. The delay also batches inbound messages into one context.
Sleep hours. Set them. A bot texting at 2:10am reads as a scam and generates complaints. My default is 8am to 8pm local, overnight enquiries queued and answered at 8am. Confirm the sub-account time zone matches the business's location — a Perth client on a Sydney-zoned sub-account texts three hours early every night.
Human handoff. Define triggers explicitly: the contact asks for a human, uses "complaint", "refund", "cancel", "lawyer" or "urgent", hits the fallback strike limit, or the conversation runs past a set number of turns without progress. On handoff the bot stops replying, applies a Needs Human tag, creates a task and notifies the right person. Plenty of Australian leads ask for a human early — make it easy rather than making the bot fight them.
Australian compliance: consent and SMS
An AI bot is not a legal exemption. Under the Spam Act 2003 and ACMA rules every commercial electronic message needs consent (express or reasonably inferred), clear sender identification, and a functional free unsubscribe actioned within five working days. That applies identically to a message Conversation AI composed. Practically: only enable the bot on SMS where consent is recorded, keep the business name in the first outbound message, make sure STOP handling still works mid-conversation, and never let the bot re-engage someone who opted out. Add a workflow condition checking DND status and opt-out tags before the bot may reply, and record the consent source in a custom field. "The AI sent it" is not a defence — the business is the sender.
Testing before you go live
Run Suggestive mode with staff reading every draft for at least two weeks or fifty conversations, whichever comes later. Run adversarial tests yourself first:
- Ask for a price you deliberately have not trained — it should defer, not guess.
- Ask about a suburb outside the service area.
- Send an angry complaint; confirm handoff fires and the bot goes quiet.
- Message at 11pm; confirm nothing sends until sleep hours end.
- Try to book inside minimum notice; confirm the calendar refuses cleanly.
- Reply STOP mid-conversation; confirm DND is set and the bot stops.
- Send three rapid messages; confirm the delay batches them into one sensible reply.
Read every transcript from that fortnight; the edits staff make to suggested replies are exactly what belongs in the FAQ set. Configuration references sit in the GoHighLevel help documentation, and current plan inclusions and AI usage rates sit on the GoHighLevel pricing page.
Measuring GoHighLevel Conversation AI performance
"Replies sent" is a vanity metric — it says the bot is running, not working. The number I report is booked appointments per 100 bot conversations, monthly. A well-trained Appointment Booking bot on warm inbound usually lands in the teens to low twenties; under ten, the training set is the problem, not the model.
Track four supporting numbers: handoff rate (rising means a training gap), fallback rate (each one a missing FAQ), show rate on bot-booked versus human-booked appointments, and median first-response time. These come from opportunity stages and tags, which is why the tagged pipeline was a prerequisite. Running several AI products in one account? The AI Employee pricing breakdown covers how the usage adds up.
Common mistakes to avoid
- Going straight to Auto-Pilot on day one. Two weeks in Suggestive costs nothing and catches the answers that would have cost you a customer.
- Crawling the whole website. Old posts and expired offers are where phantom prices come from. Crawl five to ten current pages.
- Sleep hours off, or the wrong sub-account time zone. The fastest route to a 2am text and a complaint.
- No handoff path. With no way out, frustrated leads just stop replying and you never learn why.
- Booking mode with no buffers or minimum notice. The bot books a slot fifteen minutes away and the technician will not make it.
- Reporting "messages sent" to the client. Report bookings per 100 conversations, show rate and handoff rate instead.
If you want your Conversation AI bot trained, tested and wired into a booking workflow that actually fills the calendar, book a strategy call with the HL Growth Partner team.
Frequently asked questions
How much does GoHighLevel Conversation AI cost?
It is billed as usage, not a per-seat licence. You pay a small amount per message the bot generates, drawn from your account wallet or the AI allowance included with your plan. Most single-location service businesses spend a modest monthly figure; check the GoHighLevel pricing page for current per-message rates, because they change.
Which channels does the Conversation AI bot support?
SMS, Live Chat, Facebook Messenger, Instagram DMs and WhatsApp, enabled individually per bot. Start with SMS and Live Chat only, then add social channels once transcripts look clean, because tone expectations differ noticeably between a text message and an Instagram DM.
Should I use Appointment Booking or Query & Answer as the Bot Goal?
If a booked appointment is the outcome you are paid for, use Appointment Booking — it still answers questions, it just steers to the calendar. Use Query & Answer for support inboxes, order questions or memberships where most messages come from existing customers rather than leads.
Can the bot text people at night?
Only if you let it. Set sleep hours and confirm the sub-account time zone matches the business's actual location. Overnight enquiries queue and are answered when sleep hours end. This matters legally as well as commercially, given ACMA's expectations around contact hours.
What happens when the bot cannot answer a question?
It sends your fallback message. Write it to acknowledge the question, promise a human and give a time expectation, then attach a workflow that tags the contact, creates a task and notifies your team. Set a two-strike limit so the bot disables itself for that conversation rather than looping.
