A voice agent is only as good as the answers it can give. A knowledge base for AI agent use is everything the agent is allowed to draw on: opening hours, prices, policies, the steps for a booking and what to say when something is unclear. When a caller asks whether the clinic accepts a particular insurance panel, the agent looks for that answer there. If the answer sits in three documents that disagree, the agent will guess or go vague.
Build the knowledge base from the questions callers actually ask. Write each answer for speech, keep one source of truth for each topic, name an owner and a review date for every answer, and test the answers on real questions before go-live. Show staff where each answer came from so they can check it.
Which questions should a knowledge base for AI agent cover?
Do not start with a list of everything your business does. Start with the calls. Your existing FAQ is a starting point, but a FAQ for voice AI needs shorter answers than a web page. Your call notes, your front-desk log, the staff inbox and the questions people type on WhatsApp are all useful sources. Count how often each question comes up, and write answers for the top of that list first.
For example, a dental clinic may find that most questions are about booking, parking, first-visit paperwork and whether a treatment is covered by a plan. Those are the answers to write first. Rare questions can wait until the calls show a real gap.
Keep a running list of the questions the agent could not answer. That list tells you what to add next, and it is usually the most valuable document in the project.
Write answers for the ear
Text on a website is built for scanning. A caller hears it once and cannot skim back. An answer that reads well on a web page can sound stiff, long or confusing when a voice reads it out.
The rules are simple. Lead with the answer, keep each answer to a few sentences, spell out numbers and avoid lists of more than three items. Remove anything that only works on a screen, such as "click here" or "see the table below".
| Written for a web page | Written for a call |
|---|---|
| Our opening hours are Monday to Friday, 9 am to 6 pm, and Saturday 10 am to 2 pm, excluding public holidays. | We are open Monday to Friday, nine to six, and Saturday until two. We are closed on public holidays. |
| Parking is available at the rear of the building, subject to availability. | There is parking at the back of the building. |
| Refunds are processed within 14 working days of receipt of the returned item. | Refunds take up to fourteen working days after we receive the item back. |
The figures in the last row are illustrative. Use the policy your business actually has, and check the wording with the person who owns it.
One source of truth for each answer
Most wrong answers start with the same fact living in two places. The website says one thing, the staff handbook says another and an old FAQ from last year says a third. The agent cannot know which is right, so it picks one, or blends them.
Choose one owned document for each topic. Other places link to it or are removed. Each entry carries a date, so an old version is easy to spot. When two sources disagree, the owner decides and the other source is corrected the same day.
Good AI agent knowledge management is mostly this discipline, and it is worth more than writing more content. A shorter knowledge base with one answer per fact beats a long one that contradicts itself.
Owners, review dates and the update rhythm
Every topic needs a named owner, a person who answers for its accuracy, and a review date. A workable rhythm looks like this:
- Opening hours, prices and availability: update whenever they change, and check at least monthly.
- Policies, such as returns, cancellations and eligibility: review each quarter, and whenever the policy changes.
- Anything involving money, safety or a formal commitment: review before each change, with a second person signing it off.
- Gaps from the call log: review weekly, and turn each gap into an answer or a handoff rule.
Log every change with a date and a reason. The log lets you explain a wrong answer to a manager later, and it shows you when an answer keeps changing for no good reason.
Teach the agent to say when it does not know
No knowledge base covers everything. The agent needs a safe way out. When the answer is not in the knowledge, it should say so plainly, then offer to take a message or pass the call to a person.
Avoid instructions that push the agent always to give an answer. That push is one of the main causes of invented replies. The article on why AI voice agents make things up explains the layers that stop this on a live call. The rules that decide when the call must be passed on are covered in writing rules your agent can keep.
Grounded answers, drawn only from what you have approved, are the aim. A good unknown answer sounds like this: "I don't have that detail in front of me. I can take your number and have the team confirm it, or put you through to someone now." The caller gets an honest answer and a clear next step.
How do you test the answers before callers do?
Testing is where most knowledge bases fail, because the team tests the answers it wrote rather than the questions callers ask. Test with the wording callers really use, including the slang, the misspelt names and the second question that follows straight after the first.
- Take 50 real questions from your call notes, with the approved answer for each.
- Ask each question on a test line, using different wording.
- Score each answer as right, partly right, wrong or invented.
- Fix the source behind every answer that is not right, rather than the single reply.
- Run the same questions again after each change, and again after any update to the source.
Keep the test list. It grows with every new question you find, and it is the quickest way to catch a change that broke something that used to work.
Show staff where each answer came from
Staff who take a handoff need to know where an answer came from, especially when a caller disputes it. Show the source document and its date next to each answer in your internal tools. That also helps each owner see which answers need a review.
When a caller says the agent was wrong, the team can see the exact source and fix it, instead of guessing what the agent was told.
How VoiFlow handles this
VoiFlow enforces its rules outside the model, so the rules that must always hold do not depend on a single sentence in the instructions. Every call is recorded, transcribed, scored and replayable, so you can find the exact call where an answer went wrong and fix what the agent was given. The work page shows how promises and the customer's history carry from one call to the next.
Frequently asked questions
How big should a knowledge base for an AI agent be?
As big as your most common questions need, and no bigger. Start with the questions your call notes show most often, and add more only when the calls reveal a gap.
Should I upload my whole website?
Not as it stands. Website text is written for reading and often includes old offers. Pull out the answers you want the agent to give, rewrite them for speech and give each one an owner.
How often should the answers be reviewed?
Prices and opening hours whenever they change, and at least monthly. Policies each quarter. Anything involving money, safety or a formal commitment before each change, with a second person checking it.
What happens if the agent does not know the answer?
It should say so, then take a message or offer a transfer to a person. The missing answer goes on the list for the owner to write, check and add.
Start with one call type and a short list of answers. Then try a call to hear how they sound on a line, or contact us about setting up the first week. The seven-day go-live plan shows where the knowledge base fits in a launch.
