A useful assistant starts with a small task and a testable definition of success. For an agency, a good first use case is receiving enquiries about a service, collecting the necessary information and suggesting the next step. Before writing a prompt, decide which conversation your team needs to improve and what information is usually missing when an enquiry reaches sales.
HighLevel documents that Conversation AI can answer questions, collect information and support bookings, depending on its configuration and supported channels. Choosing a guided, prompt-based or flow-based assistant depends on the complexity of the conversation. Official documentation.
Define the assistant’s response boundaries#
Write a one-page specification with four sections: objective, approved information, information it may request and situations it must hand over. For an SEO service, the assistant can explain the published scope, ask for the prospect’s website and offer a consultation. It should not invent custom prices, team availability or predictions about search rankings.
Prepare a concise knowledge base with current services, confirmed commercial terms, frequently asked questions and responsible team members. Add a review date to each document. If two pages describe the same service differently, resolve that contradiction before training the assistant.
Build a short conversation#
- Introduce the assistant as automated and explain how it can help.
- Ask one question at a time: main need, website and preferred contact method.
- Repeat critical details to confirm they were understood correctly.
- Offer a concrete action that fits the person’s need.
- Hand over cases requiring human judgment and leave context for the person who continues.
Human handover supports explicit requests for a person, missing information and unresolved problems. Its availability may require activation in the account. Also check who receives the conversation and what happens while the assistant is paused. Official documentation.
Test uncomfortable situations#
Create twenty test conversations: common questions, spelling mistakes, an angry user, someone who refuses to share a phone number and an enquiry about a service that does not exist. A friendly answer is not enough; record whether it was correct, requested unnecessary information and ended with an appropriate action.
Hypothetical example: sixteen of twenty tests are handled well, three need better handover and one invents a price. That last failure is enough reason to correct the reference material before expanding use. Repeat exactly the same tests after the change.
During the pilot, review conversations every week. Measure verifiable answers, handovers that receive attention and abandoned enquiries. Record new questions too, so you can improve the reference material. The pilot should produce a process you can explain and supervise, with clear boundaries for your team and the person writing to it.