AI Receptionist
Training Your AI Receptionist: Building a Custom FAQ Knowledge Base
Every business has the same twenty questions. Your hours, your prices, whether you serve their area, how booking works, what happens if it rains. Your staff answers them without thinking. An AI receptionist can too, but only if someone teaches it. Custom FAQs, the knowledge base you train the AI on, are what separate a receptionist that sounds like your business from one that sounds like a generic call center. Here is how to build one that actually works.
What the knowledge base is
The knowledge base is a structured collection of your business facts: services, prices, policies, service area, hours, and the answers to questions callers ask repeatedly. When a caller asks something, the AI checks the knowledge base first and answers from it. When the answer is not there, the AI follows your escalation rule, which is usually to take a message or transfer, rather than guessing.
That last part is the important design decision. A knowledge base is a closed book, not an invitation to improvise. The AI answers what it knows from your material and admits what it does not know. That constraint is what makes the answers trustworthy. An AI that guesses at your pricing is worse than one that says “let me have someone get back to you on that.”
Starting with your real questions
Do not start by writing FAQs from memory. Start with evidence. Pull your last month of calls, emails, and contact form submissions and list the questions that repeat. You will find the list is shorter than you expect and more specific than you would have guessed. “Do you work on Sundays” outranks “what is your mission statement” by a wide margin.
If you do not have call records yet, ask your team. Whoever answers your phones can list the top twenty questions from memory in ten minutes. Write them down in the callers’ words, not your marketing words. Callers ask “how much for a drain cleaning,” not “what are your service rates for plumbing interventions.” The knowledge base should match the language callers actually use, because that is what the AI will be matching against.
Writing answers the AI can use
Knowledge base answers are not website copy. They are scripts for spoken conversation, and they follow different rules.
Short and direct. A spoken answer should land in two or three sentences. “We’re open Monday through Friday, 8 to 5, and Saturdays 9 to 1.” If the full answer needs more detail, give the short version first and let the AI offer the rest. “Want the details on after-hours rates?” Callers tune out of long spoken paragraphs.
Specific, not approximate. “Starting at $89” is a usable answer. “Competitive rates” is not. If a price varies, give the range and the factors that move it. “Drain cleaning runs $89 to $240 depending on the clog. I can have someone give you an exact quote.” Vague answers train callers to distrust everything else the AI says.
One question, one answer. Do not bundle. A knowledge base entry for pricing should not also explain your warranty. Separate entries let the AI answer precisely what was asked instead of reciting a paragraph that half-applies.
Include the follow-up. Every answer should end with a natural next step. After hours: “Want me to book you in?” After pricing: “Want me to get you a firm quote?” The knowledge base is not an encyclopedia. It is a tool for moving calls forward.
The categories every business needs
Most knowledge bases settle into the same set of categories. Use this as your starting checklist and add what is specific to your trade.
Hours and location. Regular hours, holiday hours, emergency availability, address, parking or arrival instructions. Include the exceptions. “Closed Sundays except for emergencies” prevents the most common confusion.
Services and pricing. What you do, what you do not do, and what it costs. The “what you do not do” entries are underrated. An explicit “we don’t do commercial refrigeration” saves everyone time.
Booking and policies. How to book, deposit requirements, cancellation policy, what happens if the tech is late, warranty or guarantee terms. Policies cause the most disputes, so these answers need to be exact.
Service area. Where you go, and just as importantly, where you do not. List the boundaries in terms callers use: town names and neighborhoods, not zip code ranges.
Common situations. The questions specific to your trade. A plumber: “what to do if a pipe bursts before we arrive.” A clinic: “what to bring to a first appointment.” A law firm: “what to bring to a consultation.” These are the answers that make the AI sound like it has worked in your industry for years.
Handling what the AI does not know
No knowledge base covers everything, and the fallback behavior is a feature you should configure deliberately. The standard options: take a detailed message for a callback, warm-transfer to someone who knows, or offer to text the answer once confirmed. Pick one default and make it consistent.
What the AI must never do is invent an answer. This needs to be an explicit instruction in the configuration, not an assumption. Test it during setup by asking the demo line something obscure about your business that is not in the knowledge base. The right response is some version of “I don’t have that information, but let me get someone who does.” The wrong response is a confident, specific, wrong answer. Our escalation guide covers this fallback behavior in detail.
Keeping it current
A knowledge base rots. Prices change, hours shift for the season, you add a service, you drop one. Every change that you would tell your staff about needs to reach the knowledge base too, ideally at the same time.
The practical system: assign one person ownership of the knowledge base, and make updating it part of the same process as updating the website. New price on the site, new price in the knowledge base, same day. Most vendor dashboards let you edit entries in plain language without technical skill, so the barrier is process, not ability.
Review the “I don’t know” log monthly. Every good system records the questions the AI could not answer. That log is a to-do list written by your callers. The questions that repeat become new knowledge base entries. Do this for three months and the AI’s answer rate climbs steadily, because you are filling exactly the gaps your callers keep finding.
Advanced: documents and website sync
Many vendors now let you feed the knowledge base from existing material: your website pages, PDF price lists, service brochures. This is a fast way to build the initial version, and it works well for factual content like hours and service descriptions. But treat imported content as a draft. Website copy is written to be read, not spoken, and it usually needs shortening and restructuring before it works on a call. Import, then edit for voice.
Some systems can also sync with your website automatically, pulling updates when pages change. Useful, but verify what it pulls. You want it syncing the hours page, not the blog. Configure the sources explicitly rather than pointing it at the whole site.
How to evaluate this on a demo
Ask the vendor to show you the knowledge base editor. You want plain-language editing, categories or tags, and a test mode where you can ask questions and see which entry the AI pulled the answer from. That traceability matters. When the AI gives a wrong answer in production, you need to find the entry that caused it.
Then run the adversarial test. Ask about your pricing, your hours, and your service area, then ask something you know is not in the base. The first three should be crisp and correct. The last one should trigger the fallback gracefully. If the vendor hesitates to let you test with your own material, that tells you something about how the feature holds up under real use.
The bottom line
A custom FAQ knowledge base is twenty to thirty well-written answers in the callers’ own language, each short enough to speak, each ending with a next step, backed by a fallback rule that admits ignorance instead of guessing. Build it from your real questions, assign someone to keep it current, and mine the unanswered-question log monthly. That is how an AI receptionist starts sounding like it works for you specifically, because it does.



