Building a knowledge base: the most common mistakes
The knowledge base is the collection of business information an AI phone assistant draws its answers from: opening hours, services, prices, rules. Reputable systems answer only from it and invent nothing. The most common mistakes when building one: too little content, too much prose, out-of-date prices and missing emergency rules — all four are avoidable.
Why the knowledge base decides the quality
An AI phone assistant is exactly as good as the content the business puts into it — the language model supplies the conversational ability, not the knowledge of the business. The principle behind it: good systems answer questions only from the stored knowledge base; if the answer is not there, they ask a follow-up question or offer a callback instead of guessing. Why that principle matters and how providers implement it is explained in "Does the AI make up answers?". For the business that means: the hour it invests in its knowledge base keeps working in every single call.
Mistake 1: too little content
A common mistake is an almost empty knowledge base — greeting set up, three lines about the business, done. The assistant can then take down a request, but it can hardly answer a question: every price question, every call-out question, every "do you also do …?" question ends in an offer to call back. That disappoints callers and gives away half the benefit. The rule of thumb for the start: write down the ten questions the business is asked most often on the phone — the team knows them by heart — and answer exactly those first.
Mistake 2: too much prose
The counterpart to the empty system is the website pasted in wholesale: pages of marketing copy, company history, brochures in flowing text. The problem is not the volume as such but the form — in flowery prose the system finds the specific answer less reliably, and what it reads out sounds like a brochure rather than information. The format that works reliably is the question-and-answer pair: one specific question, one short answer of one to three sentences that can be read aloud. An invented example: "What does the call-out cost?" — "Within Mannheim €25, outside by the kilometre." That way the answer is written exactly as the assistant should say it on the phone.
Mistake 3: out-of-date prices and details
An out-of-date knowledge base is worse than an incomplete one — because the assistant gives the wrong information with full conviction. The classic case is prices that have never been updated since setup: the caller is quoted the old price and later holds you to it. The same goes for opening hours after a change, services that have been discontinued and old holiday arrangements. The practical consequence: everything that can change belongs with a maintenance date attached — and prices that fluctuate often are better phrased as a range with an offer to call back ("guide price X, binding in the callback") than as an exact figure.
Mistake 4: missing emergency and exception rules
The most expensive gaps are the rules for the exceptional case. What counts as an emergency, and which number is the call put through to then? What does the assistant say during the company holiday? How does it handle job applicants, sales reps, suppliers? Without these rules the system treats a burst pipe like a request for a quote — formally correct, in practice damage. The emergency rules in particular belong in the system from the start — in the knowledge base or in separate rules, depending on the provider — not only after the first incident. The appointment rules count too: which requests are bookable, for how long, with how much notice — details in "AI appointment booking: how it works technically".
The maintenance rhythm: building it once is not enough
A knowledge base is not a project but a stock of content that needs a rhythm. What has proven itself is a three-tier cadence: Weekly in the first weeks, go through the transcripts — every question the assistant could not answer is a candidate for a new question-and-answer pair. Monthly, check the details that change: prices, hours, promotions. As the occasion arises, follow up immediately when something changes — a price increase, a holiday, a new service. The effort drops quickly: after the first few weeks the core stock is complete, and maintenance shrinks to minutes a month. In fairness, though, this holds too: a business that cannot find those minutes will not be happy with any provider — the knowledge base then goes out of date with every system.
Frequently asked questions
How big should a knowledge base be at the start?
The ten most common phone questions with short answers, plus emergency and appointment rules, are enough to start with. After that it grows from the real calls — the transcripts show what is missing.
Can I simply have my website read in?
As a starting point yes, as an unchecked knowledge base no. Good systems read the website in and turn it into suggestions — services, frequent questions, opening hours — that you review before accepting them. Marketing copy, though, is rarely written to be read aloud and rarely contains the specific answers. So add whatever is missing as question-and-answer pairs.
What happens if a question is not in the knowledge base?
Reputable systems do not guess. They say so openly, ask a follow-up question or offer a callback. The request then lands in the business's inbox together with the transcript.
Who in the business should maintain the knowledge base?
Ideally the person who otherwise sits by the phone — they know the real questions and the right answers. What matters is clear responsibility, not a particular role.
How quickly do changes take effect?
With common systems immediately, or within a few minutes of saving. After important changes, check with a short test call.
Status & sources
As of 2026-11-15.
- Recommendations based on the state of systems common on the market, August 2026
- Example prices in the text are invented example data
- scoco.ai product information: as of 15 November 2026