GoHighLevel Conversation AI Setup: Train It, Test It and Go Live

GoHighLevel Conversation AI is the feature clients ask for most and switch on worst. Left on defaults, it answers with generic enthusiasm, invents opening hours, and offers appointments the business does not do. Set up properly — with real instructions, real training and a proper test — it holds a text or DM conversation well enough that most people do not ask if it is a bot, and it books. This is the setup I use, in the order I do it.

It builds on the general chatbot guide (what a bot should and should not do) and the qualification design (the questions it should ask). Here it is the GoHighLevel specifics.

Where it lives and what the settings mean

In the sub-account, under Settings → AI Agents → Conversation AI (the menu has moved before and may again; search “Conversation AI” in settings). The settings that matter:

Step 1: Write the instructions before anything else

The instructions (the bot’s personality or prompt field) are the setup. Everything else fills in facts. A structure that works:

Who you are: You are the assistant for Northside Plumbing, a family plumbing and heating business in Salford. You reply by text in a friendly, brief, professional tone, one or two sentences at a time, no emojis.

What you can do: answer questions about our services, areas and how we work; take details for a quote; book a callback or a visit in the calendar. Ask one question at a time. Qualify with: what the job is, the postcode, and how urgent.

What you must not do: never quote prices except the fixed prices listed in training; never promise a time slot that is not in the calendar; never diagnose a gas issue — for gas smells say “leave the property and call the emergency line on 0800 111 999”, then hand off; never say you are a human if asked directly.

When to hand off: complaints, anything about invoices or refunds, commercial or landlord jobs, anyone who asks for a person, and any question you cannot answer from training. Say “I’ll get Dave to message you directly” and stop replying.

Always end with a next step: a question, a booking offer, or “Dave will call you within the hour.”

Notice that it is mostly boundaries. Bots fail by being helpful outside their knowledge; the instructions are the fence.

Step 2: Train it

1. Website. Add the URLs of the service pages, the about page, the FAQ and the areas page — not the whole site, and not the blog. The bot reads them for facts.
2. Q&A pairs. The twenty questions customers actually ask, answered the way the owner would: do you charge for quotes, do you do weekends, how soon can you come, which areas, do you do gas, what are your fixed prices, how do I pay. This source is more reliable than crawled pages; write it.
3. Fixed prices and policies as explicit pairs, so the bot has something specific to say and a reason not to invent.
4. Calendar. One calendar with the right availability, buffer and duration. If the business needs two (callbacks and site visits), decide which the bot books and put the other behind a handoff.

Step 3: Connect it to workflows

Conversation AI is a good talker and a bad record-keeper unless you tell it what to write down. Set it to update contact fields with the qualification answers (job type, postcode, urgency) and to add a tag when a booking is made. Then workflows do the rest: the “appointment booked” trigger sends the confirmation and reminders; the “handoff” tag creates a task and notifies the right person; urgent jobs ring a phone. The missed-call text-back can hand its replies to the bot, which is the most common way it earns its cost.

Step 4: Test with the ten-scenario script

In Suggestive mode, or using the bot trial panel, run these from a test contact and read every reply:

1. “Do you cover M6?” — correct area answer, then a qualifying question.
2. “How much for a boiler service?” — the fixed price if listed; otherwise a quote path, not a number.
3. “Can someone come today?” — real availability from the calendar, not a promise.
4. “I can smell gas” — the safety line and a handoff, immediately.
5. “Are you a bot?” — honest answer, keeps helping.
6. “I want to complain about last week” — handoff, no defensiveness.
7. Book an appointment end to end — confirm it appears in the calendar and the confirmation text sends.
8. Gibberish or a wrong-number message — a polite short reply, no loops.
9. A question not in training (“do you fit solar panels?”) — “I’m not sure, I’ll ask Dave to message you,” not an invention.
10. A human replies mid-thread — the bot goes quiet and stays quiet.

Fix the instructions or training for every wrong answer, re-run the failed ones, and only then flip to Auto-pilot.

Step 5: Go live, one channel at a time

• SMS first (make sure the number is A2P-approved, or texts will be filtered and the bot will look broken).
• Read every conversation for the first week. Add Q&A pairs for every gap.
• Add web chat, then Facebook and Instagram, then WhatsApp, a week apart, testing each.
• Check the AI usage and cost in the billing area after the first fortnight and adjust wait times or channels if needed.

“It’s not responding” — the checklist

Other common problems

• Invents facts. The instructions lack a “never say what is not in training” rule, or the training is thin. Add pairs; tighten the fence.
• Too long, too chirpy. Instruct: one or two sentences, no exclamation marks, no emojis, plain English.
• Keeps asking the same question. Usually a contact field it is trying to fill that the customer answered differently; simplify the qualification list.
• Books the wrong calendar or offers slots the business cannot do. One calendar per bot goal; check availability settings.
• Does not hand off. Handoff phrases need to be explicit in the instructions, and a workflow must act on the handoff tag; otherwise it hands off into silence.

What good looks like after a month

Most first-contact texts and DMs answered within a minute at any hour; a majority of qualified conversations ending in a booking or a callback task; handoffs arriving with the details already collected; and a Q&A list that has grown by twenty pairs from real conversations. If the bot is also answering the phone, that is Voice AI, which follows the same discipline with a different setup.

Further reading: GoHighLevel Help Center.

FAQ

Want it written, trained, tested against the ten scenarios and switched on for you? See the GoHighLevel AI services — a half-day build that answers messages for years.