Map the journey, define SPINE, prepare current knowledge, and test both the conversation and the human handoff.
One customer, one offer, one conversation
In the September Conversation AI workshops, I asked participants to build around one costly customer conversation. The goal was an agent they could test and demonstrate, with a clear reason for the agent to exist.
On the second day, we worked backward from the person. What happened that made them contact the business? How do they describe the problem? What result do they want? Where should the conversation take them?
“Answer everything about our business” is difficult to design or test. “Help a first-time customer understand the service and request a callback” gives you a specific journey to work on.
Write the journey before the replies
- Trigger: What just happened that brought the customer here?
- Starting point: What do they already know, and what is still uncertain?
- Desired result: What are they trying to accomplish in their own words?
- Offer: What can this business actually provide, with what limits?
- Next step: What should happen after the conversation?
This prevents the agent from sounding helpful while taking the customer somewhere irrelevant. The right response depends on the person and the moment.
Use SPINE to define the role
SPINE is the planning structure I introduced during the first workshop. Use it to describe the job before configuring the agent.
| Part | Decision | Example |
|---|---|---|
| S: Scope | What is the agent responsible for, and what is outside its role? | Explain the service and request a callback. Do not negotiate a custom contract. |
| P: Persona | How should it communicate with this audience? | Clear, warm replies with one question at a time. No forced jokes. |
| I: Intake | What information is needed for the next step? | The question, preferred callback time and an approved contact method. |
| N: Next step | Where does the conversation lead? | Pass the request to the coordinator with the relevant context. |
| E: Escalation | When does a person take over? | A complaint, conflicting information or a request outside the approved offer. |
The examples above are hypothetical. Replace them with the actual decisions and permissions of the business you are building for.
Build the supporting knowledge
Prepare the current offer and FAQs, the agent's job description, a short operations manual and examples of the brand voice. In the workshops, we worked on these before expecting the agent to handle a real customer journey.
Write down who maintains each changing fact. Prices, schedules, availability and locations need a current source. If the agent cannot establish the correct answer, give it a useful way to ask for help.
A booking request and a confirmed appointment are different. I used that distinction in the workshop because some businesses want a person to confirm the slot. Decide which step your agent owns and make its language match that step.
Test the conversation and the handoff
Start with a small test set: a normal inquiry, an unclear request, missing information, an out-of-scope question and a customer asking for a person.
Check more than the reply. Did the right information reach the next person? Did an action happen once? Did the customer receive an accurate explanation of what happens next? If a connection fails, does someone know the request still needs attention?
Record the scenario, expected result, actual result and correction. Repeat failed scenarios after the fix. Having another person test the conversation helps reveal assumptions you no longer notice.
Your first demonstration
Show the customer situation, the agent's responsibility, one normal conversation and one human handoff. Explain what the demonstration establishes and what still requires testing in the business's actual setup.
Once you have built and tested the agent, bring the conversation plan, operating knowledge, working version and test notes into your demonstration. Those make the discussion specific: what the agent does, how you checked it and what still needs work.
Continue into the platform setup
This planning guide draws on my September 26–27, 2026 Conversation AI workshops. The recorded course page explains the build sequence. For current platform instructions, consult HighLevel's official Conversation AI setup guide and human handover documentation. Match the available settings to your account and agreed business permissions.
Adapted from Cris Vinson's recent teaching in Cebu in October 2026 and the September 2026 Conversation AI workshops. Edited into a practical guide; worksheets and hypothetical examples are added for application.