How to Build an AI Voice Agent Without Code

What you need before starting

You do not need to write any code, but a small amount of preparation makes the build far faster. Gather these before you open the platform:

• A clear picture of the calls you want automated first - one call type, not everything at once.
• Your website URL and any FAQ, pricing or policy documents you already have written down.
• Access to the calendar, CRM or booking tool you want the agent to check or update.
• A phone number to use - either a new one or your existing number for forwarding.

Deciding what "done" looks like

Before building anything, write down two or three sample calls exactly as you would want them to go, including what the agent says at each turn. This becomes your test script later, and it forces the scope decisions - what counts as in-scope, what should trigger a transfer - before you are staring at a blank prompt field trying to invent them on the spot.

Define the use case

The single biggest driver of a good result is scoping the job precisely. "Answer the phone" is too broad; "answer after-hours calls, book appointments, and take a message for anything else" is buildable. Write down the exact outcomes you want: which questions it answers, which actions it completes, and exactly when it should hand off to a person.

Create the agent

1. 1

   Create the agent

   Start a new agent in the platform and give it a name and purpose. This sets up the shell you will configure in the following steps.

2. 2

   Train it on your knowledge base

   Add your website, FAQ documents and policies so the agent answers from your real business information instead of generic assumptions.

3. 3

   Write the prompt

   Set the agent's role, tone, scope and escalation rules in plain language - what it should do, what it should never claim, and when to transfer to a human.

4. 4

   Choose the voice

   Pick a voice and language that suit your brand and callers, and preview it in a live test call before deciding.

5. 5

   Connect integrations

   Link the calendar, CRM, SMS or webhook tools the agent needs to check availability, log calls and trigger follow-up actions.

6. 6

   Attach a phone number

   Connect a new number or forward your existing line so real calls reach the agent.

7. 7

   Test the agent

   Place test calls yourself covering the happy path, edge cases and deliberately confusing questions, then read the transcripts.

8. 8

   Launch

   Route real traffic to the agent, starting with a portion of calls if you want a controlled rollout.

Train it on your website and knowledge base

Most platforms let you point the agent at your website and upload documents directly, so it can answer from your actual services, pricing rules and policies rather than improvising. Keep this material current - an agent trained on an outdated price list will confidently repeat it.

• Import your website pages for general business information.
• Upload FAQs, service lists and policy documents for anything more specific.
• Review what the agent retrieves during test calls and correct gaps in the source material, not just the prompt.

Structuring documents so the agent retrieves well

Knowledge retrieval tends to work best on documents broken into clear, self-contained sections with a heading per topic, rather than one long undifferentiated page. If a policy document mixes pricing, opening hours and cancellation rules in one dense paragraph, split it into labelled sections - the agent is more likely to pull the exact relevant fact instead of a nearby but wrong one.

Write the prompt

The prompt is the instruction layer that sits above the knowledge base. A strong prompt states the agent's role, the tone to use, the specific tasks it can complete, and firm boundaries: what it must never say, and exactly which situations trigger a transfer to a person.

• Define the role and tone in one or two sentences.
• List the tasks it can complete, in the order a caller would naturally ask for them.
• State explicit escalation triggers - complaints, medical or legal questions, anything it cannot confirm.
• Avoid vague instructions like "be helpful"; specify what helpful means for this business.

A short worked example

A weak instruction reads: "Be friendly and answer questions about our clinic." A workable instruction reads: "You are the after-hours receptionist for a dental clinic. Book, reschedule or cancel appointments using the calendar tool. Answer questions about opening hours, location and accepted insurance from the knowledge base. If a caller describes a dental emergency, direct them to the emergency line immediately. If you are unsure whether something is an emergency, treat it as one." The second version gives the agent a decision rule it can actually apply mid-call, rather than a general disposition.

Choose the voice and connect integrations

Voice choice affects how trustworthy and on-brand the agent feels, so preview several options on a real call before settling. Integrations determine what the agent can actually do beyond talking - connect a calendar for bookings, a CRM for lead records, and SMS for confirmations, using the tools already built into the platform rather than custom code.

Attach a phone number

You can provision a new local number directly in the platform or connect your existing business number by forwarding it or linking it via SIP trunking. Many businesses start by routing only after-hours or overflow calls to the agent, then expand once they trust the results.

Test the agent thoroughly

Before any real caller reaches the agent, call it yourself repeatedly. Try the obvious happy path, then try to break it: interrupt it mid-sentence, ask something outside its scope, give an unusual name or address, and confirm it escalates correctly rather than guessing.

Launch and optimise over time

Launching is not the end of the work. Read transcripts regularly for the first few weeks, particularly calls that ended in a transfer or an unclear outcome. Adjust the prompt and knowledge base based on what real callers actually ask, and widen the agent's scope one call type at a time rather than all at once.