There is one thing that scares every business owner about putting AI in front of their customers.
What if it makes something up?
Not theoretically. Practically. What if the AI promises free same-day delivery to Abuja when you only deliver in Lagos? What if it quotes a warranty period that does not exist? What if it confirms a meeting time you never agreed to?
A made-up answer from your AI is not a glitch. It is a promise from your business. Customers screenshot. Customers hold you to it. Customers leave bad reviews when you walk it back.
We obsessed over this problem for a year. This is the stack we built.
Layer One: Your Knowledge Base Is the Only Source of Truth
Most chatbots answer from "general knowledge." That is exactly the problem. General knowledge is the open internet. The open internet has wrong prices, wrong addresses, wrong hours, and confident lies.
Beakle does not work that way. Every reply is grounded in your Knowledge Base — the FAQs, policies, product details, and notes you wrote.
When a customer asks "do you deliver to Ibadan?", Beakle searches your KB first. If you wrote "Lagos delivery only, ₦2,000 within Mainland, ₦3,500 to Island," that is the answer. Exactly that. Not a paraphrase, not an embellishment.
If the KB has nothing on Ibadan, Beakle does not invent. It escalates. Which brings us to Layer Four. First, the next safety net.
Learn how to seed your KB properly in What the Knowledge Base Is.
Layer Two: Negative Entries — Telling Beakle What NOT to Say
Sometimes the bigger risk is not what to say, but what to never say.
A typical case: a fashion seller does not do refunds, only exchanges. But the AI, trained on the polite-customer-service patterns of the open internet, has a strong bias toward saying "of course, we can refund you." Even when nobody told it that.
Negative KB entries fix this. You write:
"Do NOT offer refunds. We do exchanges only, within 7 days, with receipt."
Beakle treats this as a hard rule. It is more powerful than any prompt instruction. We learned this the hard way: telling an AI "do not say X" in a system prompt gets ignored maybe 5% of the time. Telling Beakle through a negative KB entry gets ignored zero percent of the time, because the entry is retrieved and pinned to the reply context every single turn.
Layer Three: FAQ Verbatim — Bypass the AI Entirely
For some questions, you do not want AI rewording at all. You want the exact, identical text every single time.
- Your bank account details.
- Your business address.
- Your opening hours.
- Your warranty terms.
These are the answers that get screenshotted. The answers that show up in disputes. They cannot vary.
Beakle's FAQ verbatim mode flags specific KB entries as "send this text exactly." When a customer asks the matching question, the AI does not paraphrase, summarize, or "make it friendlier." It sends your words, character for character. The AI is not in the loop for these replies. It cannot hallucinate what it never wrote.
This single feature has prevented more disasters than any other layer.
Layer Four: Confidence-Based Escalation
Even with the best KB, customers ask things nobody anticipated. "Can I pay in installments via Carbon?" "Do you ship to Cotonou?" "Is this compatible with my 2018 Camry?"
When Beakle reads a question and cannot find a confident match in your KB, it does not guess. It hands the conversation to you.
This is the single most important defense in the stack. AI that knows what it does not know is honest AI. We spent more engineering time on the confidence threshold than on any other part of Beakle, because getting this calibration right is the difference between a reliable assistant and a liability.
Customers see a polite hold message. You see a clear escalation in your dashboard with the question, the context, and the reason Beakle paused. You reply. Beakle learns. The customer never gets a confident-sounding lie.
Deep dive: How Confidence-Based Escalation Works.
Layer Five: The Owner Correction Loop
The first four layers prevent hallucination. The fifth layer kills it permanently, one correction at a time.
Here is how it works. Beakle escalates a question. You reply manually with the right answer. You then type:
/qa
Beakle reads your reply, extracts the question-answer pair, and adds it to your Knowledge Base. The next customer who asks the same thing — or anything close to it — gets your answer instantly. No escalation, no waiting.
Every business is different. The questions your customers ask in month one are not the ones they ask in month six. The KB you started with is not the KB you need. The correction loop closes that gap automatically.
After 30 days of /qa use, most businesses see their escalation rate drop by 60-70%. Beakle gets quieter. Your customers get faster answers. Neither of you sees a hallucination.
Why This Stack Matters
Other AI tools optimize for sounding fluent. We optimize for being right.
Fluency is easy. Any large language model can produce confident, grammatical English about anything. Truth is hard. Truth requires grounding, refusal, escalation, and a constant feedback loop with the person who actually runs the business.
This stack is why you can hand Beakle your customers and sleep. Every layer exists because of a real failure we saw and decided to make impossible.
When Beakle is unsure, it tells you. When you correct it, it learns. When your KB says no refunds, it says no refunds. When you wrote your bank details, those are the bank details that go out — yours, exactly, every time.
That is what AI for serious businesses should look like.
