An AI Instagram DM agent that qualifies leads in the founder's voice for an online coaching business
Most of an Australian online coaching business's leads start in Instagram DMs, and followers expect the founder to sound like himself. We built an AI agent that replies in his voice, holds a real conversation, qualifies the lead and gets their phone number for the head coach, and hands anything sensitive straight to a person.
Snapshot
- Client type
- Online fitness coaching business, built through a marketing agency
- Country
- Australia
- Industry
- Education & coaching
- System
- AI Chat Agent (client-facing)
- Use cases
- Social DM lead capture · lead qualification
- Stack
- GoHighLevel Conversation AI
- Timeline
- Delivered in June–July 2026
- Headline numbers
- One job: get the phone number · readiness rated high, medium or low · sensitive conversations go to a person
What did the client need?
Reply to every DM in the founder's voice, without sounding scripted, pushy or careless.What did we build?
An agent with one job, a voice built from real past DMs, and handover rules that override everything else.How does it run day to day?
The agent qualifies in the DM; the number and a summary go to the head coach, who calls and books the sales call.
What did the client need?
Leads arrive in Instagram DMs from comments, story replies and reels. The founder can't answer every message himself, and setters can, but the conversations have to sound like him, because followers message him personally.
These DMs aren't simple sales conversations. People share their training history, their setbacks and sometimes very personal things. Others ask about performance-enhancing drugs or blood work. A reply that sounds scripted, pushy or careless loses the lead and damages the account's reputation.
What did we build?
The agent's only job is to get the lead's phone number so the head coach can call. It doesn't sell coaching or book calls in the DM. It works towards seven things as a map, not a script, skipping anything the lead has already said:
- An opening that references why the lead is there
- Their goal, in their own words
- Where they are now, and whether it's working
- What's stopping them
- Whether they want to change now
- Their work and lifestyle
- A soft offer, then their number for the coach
The voice. Short messages, one question per turn, replies that react to what the lead just said before moving on, and a list of phrases it never uses. The patterns came from the business's own past DM conversations.
Handover to a person. The agent stops and passes the conversation to the team when a lead asks if they're talking to AI or asks for a person, sends a voice note or an image, is upset or shows signs of crisis, when a team member has already replied, or when it can't answer without making something up.
Guardrails. It never invents prices, results or what the coaching includes, never prescribes doses or protocols or gives medical advice, respects opt-outs immediately and never uses pressure or fake urgency. When someone opens up about something heavy, it responds to the person first and doesn't pivot to the call.
How does it run day to day?
Instagram DM ─► Conversation AI agent ─► number captured ─► CRM summary ─► head coach calls
└► handover trigger ─► agent stops ─► team repliesWhen the number is captured, the agent writes a short CRM summary so the coach knows who he's calling: goal, current training, main obstacle, why now, work and lifestyle, a financial signal, readiness (high, medium or low) and any sensitive context. It uses the lead's own words and leaves fields blank rather than guessing. From there, the lead runs through the business's CRM and sales system.
What were the results?
We don't have conversion figures to publish. What the agent does is defined by design: it has one job, getting the lead's number for the head coach; it rates every qualified lead's readiness as high, medium or low; and it hands every sensitive conversation, voice note, image or AI question to a person.
- Job: get the phone number for the coachBuild spec
- Readiness ratings: high, medium, lowBuild spec
- Conversation stages, used as a mapBuild spec
What did we learn, and what are the limits?
- Train on real conversations, not invented ones. Voice patterns from the business's own DMs made the agent sound like the founder, not a bot imitating him.
- Give the agent one job. Selling and booking stay with people, which keeps the DM a real conversation.
- Put always-needed rules where they always apply. GoHighLevel's bot goal fields have a length limit, so detail moved into knowledge bases, but handover and never-invent rules stayed in the bot goals so they apply on every message.
- Let it be human when it matters. The short-message rule is dropped when someone shares something heavy.