What should I put into the AI assistant's knowledge?

Your services under names customers understand, your prices with what they include and exclude, working hours and locations, your return, warranty and payment policies, the objections your sales team hears daily with your approved answer to each, the questions customers actually ask, and the limits of what the assistant must never answer.

Where do I get the content if I have no documentation?

From three sources you already have: your website for the basic structure, your past WhatsApp conversations to extract the real questions in the customer's own wording, and half an hour with your best salesperson to write down how they handle objections. You never need to start from a blank page.

What corrupts a knowledge base and makes replies wrong?

Four things: old price lists nobody deleted, which turn replies into a coin toss between two prices; three different answers to the same question; internal jargon the customer does not understand; and written promises nobody honours, which become complaints. Polluted knowledge is worse than incomplete knowledge.

How do I keep the knowledge current after launch?

With three management rules: one named owner responsible for it, a rule that any price or policy change goes in the same day, and a half-hour monthly review where you read the month's conversations, add the questions the assistant could not answer, and delete whatever is no longer true.

Your Company Knowledge: What the AI Assistant Actually Answers From

Part two opened the box and showed how the assistant works: it reads the customer's message, searches your company's knowledge, then composes a reply grounded in it. And it ended on a sentence that governs everything after: the quality of the replies is a direct reflection of the quality of the knowledge you feed it. This part is about the side you control completely.

The idea to take away is simple and unforgiving: an AI assistant does not create knowledge, it distributes it. If your company's knowledge is scattered, contradictory, and living in three employees' heads, the assistant will distribute that confusion to your customers faster. If it is organised and approved, the assistant will hand your best answer to every customer within the same minute.

What goes into your company's knowledge?

Start from a practical question: if your best salesperson sat down to train a new hire, what would they tell them in the first week? That is exactly the content. In practice it is seven things:

Your services and products under names the customer understands, not your internal names, with a description of what each one solves.
Prices and precisely what they include. A price alone creates a dispute; a price with what it covers and what it does not closes the dispute before it starts.
Working hours and locations, how a customer reaches you, plus public holidays and exceptions.
Policies: returns, warranty, delays, payment, cancellation. After price, this is what customers ask about most.
The real objections your sales team hears daily: you are expensive, a competitor is cheaper, why do you take longer? Each with the answer you approve.
The questions customers actually ask, not the ones you assume they ask. These are not invented at a desk; they are extracted from your conversations.
The limits of the knowledge: what the assistant must never answer, and when it hands the customer to a human.

Notice that six of the seven already exist in your business today. The problem is not that you lack the knowledge; it is that it has never been written down in one place.

Where is the raw material? Three sources you already have

Your website. The fastest starting point. Your services pages, about page and FAQ give you the basic structure: who you are, what you offer, how you describe yourself. Most serious assistants will read your site link and fill in the fundamentals within minutes.

Your past WhatsApp conversations. This is the most valuable source and almost nobody uses it. Open the last hundred conversations and read them with a neutral eye, and you will come out with the real questions ranked by frequency, phrased the way the customer phrases them. You will also see the questions that embarrassed your staff because nobody had a single agreed answer. That list beats any document written from scratch.

Your sales team's head. Part of your company's knowledge was never written at all: how your best employee handles a price objection, and what difference they explain between you and a competitor. Sit with them for half an hour and write down what they say. That one session moves the assistant from answering to selling.

What corrupts the knowledge?

Loading everything in is not a virtue. Polluted knowledge is worse than incomplete knowledge, because the assistant does not know which of two facts is the right one and will pick confidently. Four common contaminants:

Old price lists nobody deleted. The most dangerous one. Two prices for one service means your reply to a customer is now a coin toss, and half your customers get a price you would not accept.
Three answers to one question, written by three people at three different times. Pick one and delete the rest.
Internal jargon. The package name in your accounting system means nothing to a customer. Write in the customer's language, not your staff's.
Promises nobody honours. If you write that you reply within an hour and you do not, you have programmed the assistant to manufacture complaints instead of deals. Put nothing in the knowledge that you do not actually deliver.

The discipline: who owns it and when it gets reviewed

The biggest reason projects like this die is not technical. The knowledge is entered once with enthusiasm, then prices change and it does not, and within two months you stop trusting the assistant and switch it off. The cure is three plain management rules:

One owner. A specific person, by name, responsible for the assistant's knowledge. Responsibility spread across everyone is responsibility held by nobody.
The same-day rule. Any change to a price, a policy or working hours goes into the assistant's knowledge the same day, exactly as you would update your website.
A short monthly review. Half an hour: read the month's conversations, pull out the questions the assistant could not answer and add them, delete whatever is no longer true. That half hour is the difference between an assistant that improves and one that erodes.

And start small. Do not wait for complete knowledge before going live: twenty common questions with approved answers, plus a correct service and price list, cover the overwhelming majority of your daily messages. The rest gets built from real conversations, which is far more accurate than anything drafted at your desk.

The Origami view

That order of operations is why training in Mahir, Origami's AI assistant for WhatsApp, starts from your website link: you give it the link, it reads your site and fills in your company's details, then you add your services, prices and frequently asked questions with answers you approve yourself. It does not answer from general knowledge, only from the knowledge you entered, and it runs on your own existing company number with no new number and no extra app on your employee's phone — linking is a single one-time QR scan. The trial is free, so you can build the knowledge and watch it work before committing to anything.

Conclusion

Your company's knowledge is not a technical file. It is a commercial asset that used to be scattered and is now written down. Organise it once, give it an owner, review it monthly, and you get an assistant that replies the way your best employee replies. But one uncomfortable question remains: what if the customer asks about something that is not in that knowledge? What stops the assistant from confidently inventing an answer? That is part four.

Sources

Saudi Data and AI Authority (SDAIA) — guidance on AI ethics and data governance.
Saudi Personal Data Protection Law and its implementing regulation — SDAIA, on handling customer data inside conversations.
Official WhatsApp Business documentation from Meta — messaging and business account policies.

Your Company Knowledge: What the AI Assistant Actually Answers From