Sales Scripts and a Knowledge Base From Your Chat History
How to build sales scripts and a knowledge base from chat history and calls: objections, FAQs, what not to promise — step by step.
A new sales rep starts on Monday. Her manager sends her a link to “Sales Script v3 (OLD)” from two years ago and says, “Read this, then ask Anna.” Anna is the best rep on the team. She knows how to handle “your competitor is cheaper,” what to do when a customer asks for net-60 terms, and why you never promise two-day delivery. All of that lives in Anna’s head — and in three thousand customer chats.
Your company’s knowledge is already written down. It’s just sitting in conversations and calls instead of documents. Here’s how PRANUS builds sales scripts and a knowledge base from chat history, and which parts of the job stay with people.
Why the knowledge base never gets written
Writing a sales script from scratch takes a sales manager days per pipeline. A company-wide knowledge base takes weeks and is rarely finished: everyone’s busy, and “How to answer the warranty question” is never anyone’s top priority.
Yet the answers to all of those questions have already been given — hundreds of times, in real conversations with customers. Some worked. Some didn’t. The job isn’t to invent a script. It’s to find what already works in your own history.
How the knowledge factory works, step by step
The knowledge factory is part of PRANUS implementation. It takes customer conversation history from CONNECTUS and turns it into draft documents for the MEMUS knowledge base.
Step 1. Inventory and estimate
The factory checks which channels are connected and how many conversations they hold, and shows a cost estimate before it starts. Every run has a hard spending cap: hit it, and the factory stops and asks you. It starts with a pilot on 50 conversations, then processes the rest.
For a sense of scale: in our own pilot, run on chats from a logistics company in our own group, about two hundred conversations cost $2–3 in AI usage.
Step 2. Cleanup
Personal data — names, phone numbers, ID numbers — is stripped out. Ads, spam and empty threads are filtered. Long threads (on messaging apps, a customer can stay in one chat with you for months) are split into episodes by pauses and topic changes, not cut off.
Step 3. Facts with a source
From every conversation the factory extracts facts: which service, what price, what the customer asked, how the rep answered, how it ended. Each fact links back to the conversation it came from, with a date.
Step 4. A map of knowledge areas
Before writing anything, the factory proposes a list of topics: services and pricing, delivery, paperwork, objections and so on. You confirm or edit it. Each area then becomes its own document.
Step 5. Draft documents
Sales gets seven documents:
| Document | What’s inside |
|---|---|
| Services and pricing | What you sell, at what price, on what terms |
| How we communicate | Tone, forms of address, standard phrasing |
| Sales flow | Conversation stages from first question to payment |
| Objections | What customers say and which answers led to a deal |
| FAQ | The most common questions and how to answer them |
| What not to promise | Deadlines, discounts and terms reps must never offer |
| Documents and process | Contract, invoice, onboarding steps |
Support gets five more on the same principle. A price list doesn’t become one giant table: each service is a short section — service, price, terms — so an AI agent can find the exact answer.
Step 6. Questions for the owner
The most valuable thing the factory produces isn’t the drafts — it’s the contradictions it surfaces. If reps say one thing in some chats and something else in others, the factory doesn’t pick a side. It asks you, with three buttons: “That’s correct,” “Remove from the knowledge base,” “Agent checks with a specialist.”
A real example from our pilot: the extracted facts said a border-crossing slot is booked in the morning, while the written policy said bookings are accepted until 11 p.m. the same day. The factory flagged it as a question. One answer, and the knowledge base no longer carries a contradiction an AI agent would have repeated to a customer.
Step 7. Testing on real questions
The factory sets aside a sample of real customer questions up front and doesn’t use them while building. Afterward it runs them through the finished knowledge base the same way an AI agent would search. Where the base falls short, it gathers more facts on those topics and tries again. A different model grades the answers than the one that wrote the documents.
Step 8. Approval and publishing
A person approves the drafts. Only then do they go into the knowledge base, where employees and the AI sales agent can use them.
Objection handling that actually closes deals
The usual approach to a script is collecting the “right” answers to objections. The factory goes further: it links each objection, the rep’s answer and the deal outcome from the CRM. The Objections document keeps the answers that more often led to a closed deal — not the ones that just sound good.
Here’s what an entry looks like:
Objection: “Other carriers are cheaper.” What worked: clarify the route and timeline, spell out what’s included in the price (insurance, paperwork), and compare the total cost, not the base rate. What didn’t: offering a discount right away — customers then tended to ask for another one.
That’s an illustration of the format, not a ready-made script. Every company’s answers are different.
Calls become scripts too
Chats are half the story. Calls are the other half. In CRMUS, every call longer than seven seconds is transcribed and summarized. Then the conversation is tagged:
- topic and purpose of the call;
- outcome: agreed, call back, thinking it over, lost (with the reason);
- customer objections, quoted word for word, and how the rep responded;
- questions the rep couldn’t answer;
- what the rep did well or poorly: qualified the volume, stated the price, booked a next step, or kept interrupting.
Each call is tagged once, and reports and scripts are assembled from those tags quickly and cheaply. The Suggest Script button drafts a stage script from successful and unsuccessful calls plus established sales literature. Scripts are versioned, and every subsequent call is scored against the script from 0 to 100, so a manager sees exactly where conversations go off track.
What stays with people
The factory never publishes on its own. You make three decisions:
- Confirm the topic map — what the knowledge base should cover.
- Resolve the contradictions — what’s actually true.
- Approve the drafts — read and edit them before people and agents start relying on them.
That’s usually a few hours of a manager’s time instead of weeks of writing from scratch.
A knowledge base that doesn’t go stale
A one-time export is out of date within a month: prices change, a new service launches. So the factory keeps running. Once a week, new conversations add facts to the right sections. If a new fact contradicts an old one, you get a question — not a silent overwrite. Documents are updated as new versions instead of piling up as duplicates.
And when a customer asks something the knowledge base doesn’t cover at all, the agent doesn’t make up an answer. It asks a specialist, and the answer becomes new knowledge. That loop is covered in Every unanswered question becomes knowledge. How the AI sales agent learns from all this is in How our AI sales agent learns.
Checklist: get your history ready
- Connect the work accounts where customer conversations actually happen.
- Mark personal chats as Personal — they stay out of the analysis and the CRM.
- Gather your current price lists and policies; the factory checks chats against them.
- Name the person who will resolve contradictions.
- Set a spending cap for the first run.
The bottom line
You don’t need to invent scripts and a knowledge base. You need to find them in your own history. The knowledge factory does the grunt work: reads thousands of conversations, extracts facts, ties answers to deal outcomes and surfaces contradictions. People decide what’s true and approve it. It’s part of implementation; the details are on the knowledge transfer page.
Curious what the factory would find in your chats? Request a demo and we’ll show you real drafts and owner questions.