AI training: knowledge base, FAQs, brand voice
Teach the AI the business: knowledge base, FAQs and brand voice.
Read the lesson
Goal: Teach the voice agent the business's facts, approved answers and voice. Then prove it with a scored test.
You need: The workshop sheet (lesson 2). Services built (lesson 4). The owner's written prices and emergency wording. Any plan; testing in the browser is free.
Why this matters
The agent can only say what it knows. If the knowledge is thin, it either says "I'm not sure" too often or fills the gap with a guess. A guess about price, hours or service area costs the owner money and trust.
Good training does three things. It gives the agent the facts. It gives it the exact words for sensitive answers. And it tells it how to sound like this business, not like a call center.
Most of the quality of an AI receptionist is decided here, not in the voice settings. Plan to spend more time on this lesson's work than on any other setup step.
The four layers of what the agent knows
Think of the agent's knowledge in four layers: identity and voice, rules, exact answers and background facts. Each layer has its own place in Voice Agents.
Four layers of agent knowledge
| Layer | What goes in | Where it lives |
|---|---|---|
| Identity and voice | Name, greeting, tone, words to use and avoid | Basics: Initial greeting, voice instructions, Voice |
| Rules | What to book, transfer and never say | Voice instructions, Advanced instructions |
| Exact answers | Approved prices, policies, emergency wording | Knowledge: Quick answers, Add FAQ |
| Background facts | Service details, areas, warranties, about us | Knowledge: Knowledge sources, Manage sources |
Use Quick answers when the words must be exact. Use Knowledge sources when the agent needs facts but may phrase them freely.
Identity and brand voice
Start with the Initial greeting. It is the first thing every caller hears. Keep it to one or two short sentences. There is no separate AI-disclosure setting, so put the disclosure here if the client wants one. For example: "Thanks for calling Northside Plumbing. This is Sam, the virtual assistant. How can I help?"
Then write the voice instructions. The box takes up to 32,000 characters, but long is not better. Clear is better. Write in plain lines:
- Who the business is and who it serves.
- How the team talks. Friendly and short? Formal? Do they say "folks" or "customers"?
- What the agent should do first on most calls.
- What the agent must never do.
Listen to two or three real calls by the office, if the client has them. Copy the phrases the best person on the phone uses. That is the brand voice. Do not invent a personality the owner does not have. Callers can tell.
Pick the Language and the Voice. Eight languages are available: English (US, UK, Australia, India), Spanish (Spain, Mexico), French and Hindi. If many callers speak Spanish, ask the owner whether to offer it.
Rules: what to book, transfer and refuse
Write each rule as one plain line. Number them. Rules the agent can follow look like this:
- Book these services: the active list in
Services. - Do not book outside these towns: list them.
- If the caller describes a case on the owner's emergency list, read the owner's safety line, then transfer.
- Quote only prices in
Quick answers. For anything else, say a tech will give the price on site. - Never give repair steps or safety advice beyond the owner's approved line.
- Never agree to a discount.
There is no price-guard switch in the app. Rules 4 and 6 are the guard. Test them hard.
Emergency wording comes from the client. Each client writes its own, with its own advisors. You paste it in exactly. You do not edit it, and you do not supply your own.
Exact answers and background facts
In Quick answers, click Add FAQ for each question that needs exact words. Common ones:
- What is your service call or diagnostic fee?
- What areas do you serve?
- What are your hours?
- Do you offer financing?
- Do you work on weekends?
In Knowledge sources, use Manage sources to add documents and the website. Uploads can be PDF, Word, Excel or text files, up to 10MB each. Good sources are the service descriptions, warranty terms and the "about us" page. Remove old price sheets first. An old PDF with last year's prices will cause wrong answers.
Use Website knowledge & answer testing to check what the agent learned from the site. Use Pronunciation help for town names and brand names the agent says wrong.
The Knowledge tab
What the agent knows
Build it in ServiceAgent
- Open
Voice Agents. Open the client's agent, or create one.Use templategives a starting point. - In
Basics, setAgent name,Language,Initial greetingandVoice. - Paste the voice instructions. Add the numbered rules.
- In
Knowledge, add eachQuick answersentry withAdd FAQ. - In
Knowledge sources, add the approved documents and the website. Remove anything out of date. - Add
Pronunciation helpentries for local names. - Open
Test & publish. UsePractice conversationfor typed tests. It makes no call and changes no records. - Use
Talk in browserfor spoken tests. It is free and does not change customer records. - Click
Publishonly when the scored test passes. Publishing updates live calls. - Check it works. Run the 20-question test below and record the score.
The 20-question test
Write 20 test questions before you test. Use real call reasons from the workshop. Include:
- 8 normal booking requests across the top services.
- 4 price questions, 2 with approved prices and 2 without.
- 3 questions about hours, area and policies.
- 2 emergency cases from the owner's list.
- 2 out-of-area or out-of-scope requests.
- 1 discount request.
The 20-question test mix
- Booking requests8
- Price questions4
- Hours, area, policy3
- Emergency cases2
- Out of area or scope2
- Discount request1
Questions20
Ask a second person to play some callers. You know the answers too well to test your own work fairly. Score each answer right or wrong. Right means the agent said what the owner would say, and did what the rule says. Fix every wrong answer at the cause, then retest that question and its neighbors.
When an answer is wrong, find which layer failed. Wrong fact: fix the knowledge source. Wrong words: add a Quick answers entry. Wrong action: fix the rule. Wrong service name: fix the service or the knowledge article, then test again.
Find the layer that failed
- An answer is wrong
- Was a fact wrong?YesFix the knowledge sourceNoNext question
- Were the words wrong?YesAdd a Quick answers entryNoNext question
- Was the action wrong?YesFix the ruleNoWrong service name: fix the service or article
- Retest in Practice conversation
The target before go-live is 20 out of 20. A receptionist who gets 1 call in 20 wrong is not ready either.
Worked example: Northside Plumbing
You train Northside Plumbing's agent and run the 20-question test in Practice conversation.
First run: 16 right, 4 wrong. Score: 16 / 20 = 80%.
| Question | What went wrong | Layer | Fix |
|---|---|---|---|
| Drain clearing price | Quoted an old price | Background facts | Removed the 2024 price PDF |
| Water heater flush | Called it a "tank repair" | Exact answers | Added the service name to Quick answers |
| Caller in a town across the river | Booked it | Rules | Listed served towns in the rules |
| Asked for 10% off | Said "I can check" | Rules | Added rule 6 and a polite reply |
Second run: 19 right, or 95%. The out-of-area caller was refused, but the agent did not offer a referral. You add the owner's referral line. Third run: 20 / 20 = 100%.
Northside's test score by run
Cost of all this testing: 0 credits. Every test ran in Practice conversation and Talk in browser. The real-number calls wait for go-live day, where five planned calls prove the phone path as well as the answers.
One brain, many channels
The same knowledge helps more than the phone. Website chat can answer from the knowledge base too. On Growth and Franchise, the chat replies with AI. On Launch and Core, chat replies stay manual, so the office answers them.
So write knowledge once, in a form that works for both voice and text. Short sentences help both. A caller cannot scroll back, and a chat visitor will not read a wall of text.
Website chat by plan
- Chat replies stay manual
- The office answers
- Chat replies with AI
- Answers from the same knowledge
- Growth: $95 billed yearly, $119 monthly
A Core client may want AI chat at night. That is a plan question, not a training question. AI website chat starts on Growth, at $95 a month billed yearly or $119 month to month.
Keep it current
Knowledge goes stale. Prices change in spring. A new tech joins. The business stops serving a town. Each change must reach the agent the same week.
Set a monthly knowledge check in your service rhythm. Put it on the same day as the monthly report, so it never gets skipped. It takes 15 minutes:
- Ask the owner: "Did any price, area, hour or service change?"
- Read five recent call transcripts. Look for "I'm not sure" and for answers you did not expect.
- Update
Quick answersandKnowledge sources. - Rerun the questions that touch what you changed.
Size changes the load. An owner-operator changes little, so the check is short. A fleet adds services and staff more often. A multi-location business may have a different price or area per location. Then each location's rules need their own lines, and the test needs questions for each location.
Common mistakes
- Uploading everything. Old price sheets leak wrong answers. Fix: add only current, approved documents.
- Writing the emergency script yourself. Fix: use the client's own words exactly.
- Testing only easy calls. Fix: build the 20-question test with hard cases on purpose.
- Fixing one answer, not the cause. Fix: find the layer that failed and fix it there.
- Publishing before the test passes. Fix: publish only at 20 out of 20.
Do this now
Write the 20 test questions for one real or practice client. Train a practice agent on the free Launch plan. Run the test in Practice conversation. Record the first score, the fixes and the final score.
Practice the conversation
- Operator I need three things from you before the AI talks to anyone.
- Owner Go ahead.
- Operator Your current prices, your emergency wording, and the towns you serve.
- Owner Can't the AI work out emergencies on its own?
- Operator No. It matches what callers say against cases you write down.
- Owner Okay. What else?
- Operator I'll run 20 test calls. You'll see the score before it goes live.
- Owner What if it says something dumb?
- Operator Then I fix the cause and test again. It goes live at 20 out of 20.
Quick check
- In a test, the agent calls a drain clearing a "pipe repair". What do you do?
- Go live, because callers will understand
- Run the same test again
- Fix the service name and the knowledge, then test again
- Ask techs to correct callers on site
- The owner wants the agent to say the exact $89 diagnostic fee every time. Where does it go?
- `Quick answers`, with `Add FAQ`
- A long PDF in `Knowledge sources`
- `Pronunciation help`
- `Voice`
- Your first test run scores 17 of 20. What percent is that, and is the agent ready?
- 85%, ready to publish
- 85%, not ready until it scores 20 of 20
- 17%, not ready
- 90%, ready to publish
ServiceAgent UniversityLearn to run a service business on autopilot. Free courses, Blueprints and certification for Service Operators.