How to write the knowledge base for an AI voice agent

When people try an AI voice agent and come away unimpressed, the cause is usually not the voice or the model. It's the instructions. The agent was given a vague description of the business and asked to "be helpful," and it filled the gaps with guesses. Writing a good knowledge base for an AI voice agent is mostly about being specific.

The document that fixes this is the knowledge base: a plain-language description of how your business wants calls handled. Here's how I'd write one.

Building the knowledge base for an AI voice agent: start with the calls you actually get

Before writing anything, list the ten most common reasons people call. For most businesses it's some mix of: booking or rescheduling, hours and location, "do you do X?", "how much does Y cost?", status of an existing job, and a handful of urgent situations.

Write the knowledge base around those. A knowledge base organized around real calls will outperform a polished company bio every time.

For each call type, write four things

1. What the caller usually wants. One sentence.
2. What the agent can say. The approved answer, in the words you'd want used.
3. What the agent should ask. The details you need to move forward: name, callback number, address or service area, the problem in their words, preferred times.
4. What happens next. Book it, transfer it, take a message and promise a callback within a stated time, or send a text with a link.

Be explicit about what the agent must not do

This is the most important section, and the one most often skipped. Write down plainly:

• Topics that always go to a person (complaints, billing disputes, anything legal or medical, anything safety-related).
• Things the agent should never promise (exact prices, arrival times you can't guarantee, outcomes).
• What to say in an emergency. Usually that's directing the caller to call 911 or the appropriate emergency line. A voice agent is not an emergency service.
• How to handle someone who asks whether they're talking to a person. Answer honestly.

Handle pricing questions deliberately

Callers ask about price constantly. Decide your policy and write it down. Some businesses give ranges, some explain what affects the price and offer an estimate visit, some only quote in writing. Any of those can work. What doesn't work is leaving it undefined, because then the agent improvises.

Write the way you talk

The agent will speak your knowledge base out loud, more or less. Short sentences. Common words. No internal jargon unless your customers use it too. Read it aloud yourself. If it sounds stiff, it'll sound stiff on the phone.

Keep it current

Hours change, services change, staff change. Put the knowledge base somewhere you can edit it quickly and make updating it part of your routine: new service, new holiday hours, new policy, update the document the same day.

Review real calls

Once the agent is live, read the transcripts and summaries for the first couple of weeks. You'll find questions you didn't anticipate and answers that need rewording. Every fix goes back into the knowledge base. This review loop is what turns a decent agent into a good one, and it's the same loop a good manager would use to train a new receptionist.

Writing all this down is useful even if you never deploy a voice agent. Most businesses discover their own call handling was never written down at all.

Related reading

• Why RingIris is built around an AI voice agent
• Speed-to-lead automation
• AI receptionist vs answering service (RingIris)

Frequently asked questions