Every Unanswered Question Becomes Company Knowledge
How a company knowledge base grows itself: when the AI agent can't answer, a specialist replies once and the company keeps the knowledge for good.
Friday, 11:40 p.m. A prospect messages you: “Can you ship on net-30 terms if this is our first order?” Your AI agent checks the knowledge base. There’s a policy on payment terms for existing customers. Nothing about first-time buyers.
Most companies end up in one of two bad places from here. Either the bot confidently makes something up (“Yes, of course!”) and your CFO cleans up the mess on Monday. Or it says “Please contact a sales rep,” the prospect waits all weekend, and buys from whoever answered first.
PRANUS takes a third path. The agent says plainly: “Let me check with a specialist and get back to you.” The question goes to people, the customer gets an answer, and your company knowledge base gets a new entry. Next time, the agent answers on its own. Here’s how the loop works.
Why knowledge bases stop growing after launch
Every knowledge base is incomplete on launch day. That’s fine. The problem is that it barely grows afterward.
Here’s the usual pattern: a customer asks something new, the rep asks in the team chat, someone replies, the rep forwards it to the customer. Problem solved. But the knowledge stays buried in a chat thread. A month later another customer asks the same thing, and the whole dance repeats.
Every one of those questions is a signal: there’s a gap here. Catch those signals and the knowledge base grows by itself — exactly where customers need it.
How it works, in seven steps
1. The customer writes, the agent searches
The customer writes on any channel; messages land in CONNECTUS. The AI agent searches the MEMUS knowledge base.
2. The knowledge base answers honestly, in parts
MEMUS doesn’t give one yes-or-no for the whole question. It breaks the answer into parts: what’s covered with a quote from a document, what’s partly covered, and what’s missing. For the payment-terms question: “Terms for existing customers — covered, with a link to the policy. Terms for first-time customers — not in the knowledge base.”
The agent replies with what’s confirmed and says, about the rest: “I’ll confirm this with a specialist and get back to you by noon Monday.”
3. The question goes to people
A card appears in the Questions section of CHATUS, the company messenger: the customer’s question, a summary of the conversation, what the knowledge base already covers, and a deadline. The specialist doesn’t have to read the whole thread.
While the question is with people, the bot stays quiet on that topic. If the customer adds details, they’re attached to the same card.
4. Specialists answer
There are three ways to answer, and you can mix them by topic:
| Mode | How it works | Best for |
|---|---|---|
| Queue | Every specialist in the area sees the card; the first one clicks Take it | Routine questions, on-call rotations |
| Direct | The question goes to a named expert. If they don’t pick it up in time, it moves to the shared queue | Narrow topics: legal, engineering, finance |
| Panel | Several specialists discuss, an AI assistant drafts a summary, people mark which messages go into the final answer | Complex or disputed questions |
Panel questions can run in blind mode: specialists don’t see each other’s answers until approval. The first confident reply doesn’t anchor everyone else, and disagreements surface right away.
5. The answer is approved and sent
A designated owner approves the answer, and it goes to the customer. The system shows two previews:
- what the customer receives — personal, with their details;
- what goes into the knowledge base — generalized, with no names, phone numbers or contract numbers.
A “don’t add to knowledge base” checkbox covers one-off and personal cases.
If three more customers asked the same thing in the meantime, they’re attached to the same card (“3 more customers are waiting for this answer”), and all of them get the approved reply.
6. The knowledge becomes a candidate
The answer doesn’t enter the knowledge base as fact right away. It goes in as a candidate, and the knowledge base owner reviews and approves it. The rule is simple: the knowledge base holds only what’s been confirmed.
The new entry goes where the gap was: a question from the AI sales agent fills in the sales knowledge base; one from the support agent goes into support’s.
7. Next time, it’s instant
Approved knowledge becomes a ready answer. When a similar question comes in — even phrased differently — the agent finds it in a fraction of a second, without calling a large model. Faster and cheaper.
What keeps the loop from stalling
Question-answer-knowledge loops break on small things: a question nobody picked up, a customer forgotten. So every question has a deadline:
- As the deadline approaches, whoever took the question gets a reminder.
- Once it passes, the question escalates to the team lead.
- The customer isn’t left hanging: the agent writes, “Our specialist is still looking into this; we’ll reply by…”
Human operators can feed the loop too. In a customer conversation, they select messages and click Send to Questions. That way, knowledge from live conversations with people — not just with agents — makes it into the base.
Knowledge with an expiration date
An answer that was right in March may be wrong by September: terms change, a new policy comes out. In MEMUS, knowledge can carry an end date, after which it drops out of search results automatically. Retired knowledge isn’t erased; it’s closed, so you can see what used to apply and when it stopped.
And if a specialist’s answer contradicts an older document, the answer is recorded as a correction to that document, and search returns the corrected version.
What the company gets
Customers get an honest answer — not a fabrication, and not “contact us on Monday.”
Specialists answer each question once. The second time it comes up, the agent finds the answer itself.
Managers see where the company’s knowledge has holes: which topics keep going to people, who answers quickly, and where questions sit.
AI agents get more accurate every week — not because they get smarter, but because the knowledge base grows exactly where customers ask.
Setting it up: a checklist
- Define question areas: sales, logistics, finance, support.
- Assign owners and a response target for each (for example, 60 minutes during business hours).
- Pick a mode for each area: queue, direct or panel.
- Decide who approves knowledge for the base — it doesn’t have to be the person who answers customers.
- Connect the knowledge base and escalation to your AI agents in the agent console.
Where the starting knowledge comes from
The question loop fills gaps, but you don’t have to start from zero. During implementation, the agent builds a starting knowledge base from your documents and chat history — see Sales scripts and a knowledge base from your chat history. And for why the agent is able to say “I don’t know” instead of improvising, read AI that says “I don’t know”.
The bottom line
A knowledge base shouldn’t be a project you’ll “finish later.” Every unanswered question is a free hint about what’s missing. PRANUS turns those questions into answers for customers and knowledge for the company: the agent asks, people answer once, and the knowledge base grows on its own. It’s all set up during 14-day implementation.
Want to see the question-answer-knowledge loop on your own use case? Request a demo.