Updated 20 September 2026: first published on Elite Agent in April 2025. The article has been refreshed as an AI basics guide to preparing and testing a small voice-assistant prototype.
In January 2025, I used ElevenLabs to create a voice assistant around one of my real estate leadership articles. I called it Agent Samurai. Get it?
The initial setup took about ten minutes. That was enough to explore an idea, but it was the beginning of testing, not the time required to build a dependable service for clients.
I gave the assistant my article as source material. I was configuring its behaviour and supplying knowledge, rather than training a new underlying AI model. ElevenLabs’ knowledge-base documentation describes how documents can provide information for an assistant’s answers.
Screenshot from the original Agent Samurai experiment. The linked article is background reading; this is not a current widget setup guide.
What could my first voice assistant answer?
Pick a narrow, low-stakes topic with a source you can check. Explaining your published open-home process is a better first experiment than attempting to answer every legal, finance and property question a visitor might raise.
Write its job in one sentence: “Help a visitor understand our published inspection process and find the right person for questions the document does not answer.”
That sentence gives you a boundary to test. It also helps you choose which information belongs in the assistant’s source material.
Prepare the source before choosing a voice
Start with one short, approved document. Give it a title, review date and owner. Resolve contradictory instructions and remove expired details before uploading it.
List questions the document can answer and questions it cannot. If the source does not establish an inspection time, a school catchment or a property dimension, the assistant should not fill the gap with a plausible answer.
Decide what happens at that boundary. It might direct the visitor to a published contact route or collect a request for follow-up, if your setup actually supports that. Supplying a document does not automatically connect an assistant to your calendar or CRM.
Try this prototype-planning prompt
Help me plan a small internal voice-assistant prototype for a real estate agency.
One job: [TASK]
Approved source material, owner and review date: [DOCUMENT]
Actual route to a person: [DETAILS]
Actions the prototype is allowed to take: [ACTIONS OR NONE]
First identify gaps or conflicting statements in the source. Ask me to resolve them before drafting instructions.
Then propose a brief for the assistant: introduce itself as AI, stay within the supplied information, acknowledge missing answers and use the stated human-contact route. Do not invent facts, appointments or integrations.
Create eight test questions: three answerable from the source, two outside its scope, one with an unclear request, one containing a wrong assumption and one requesting a person. Give the expected answer or behaviour and the source passage supporting it.
This is a test plan for internal review, not a claim that the assistant is ready for clients.Run the tests yourself
Check the proposed answers before testing the voice assistant. Otherwise, you may end up judging one AI answer against another incorrect AI answer.
Ask the questions aloud in different ways. Interrupt, correct a detail and ask a follow-up. Record where the assistant loses context, guesses or fails to offer the agreed next step.
Choose a voice you have permission to use and make the AI introduction clear. If another language is involved, have someone competent in that language check the experience. A convincing voice does not establish that the information is correct.
Before any client-facing trial, settle who maintains the source, reviews conversations and handles requests. Confirm the arrangements for recording and data use, and make it possible to stop or reach a person.
Build on what the small experiment teaches you
The first success might simply be an assistant that explains one process clearly and admits when it cannot help. Keep the test results and repeat them when the source or configuration changes.
Agent Samurai was one step in my own experimentation. Read how Scoop became the beginning of Ailsa for the next part of that story.
Start with a source you trust, a job you can describe and questions whose answers you can check. That gives your prototype a useful purpose before it ever speaks to a client.
The original version appeared in Samantha McLean’s AI Powered Agents newsletter. Read the refreshed source article on AI Powered Real Estate by Samantha McLean on Substack.