Memuat
Memuat
Want to go deeper?
Join one of our courses or ask us directly on WhatsApp.
Picked from the software, topic, and level of “Agentic AI Without Coding Class #12: Preparing Assistant Knowledge: Tidy SOPs, Catalogs, and FAQs”.
Other ways we can help with Agentic AI.
Learn with a mentor
Scheduled AI & Machine Learning courses with live guidance and a certificate.

MATLAB Agentic AI Course: Building Autonomous AI Agents with MATLAB & Simulink

Midjourney Course: Create Stunning AI Images from Text
★ 4.9

DALL-E Course: Creating Professional AI Images from Text
★ 4.9
Have it built for you
No time to learn Agentic AI? Our team can do the project for you.

MATLAB & Simulink Development — Agentic AI for Simulation Automation and Data Analysis

AI Image Creation with Stable Diffusion — Unlimited Custom Visual Production
★ 4.9

AI Image & Video Creation with Leonardo AI — All-in-One Visuals for Business
★ 4.9
Get notified about new Agentic AI classes
At most one email a week. Unfollow anytime with one click.
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Welcome back to Session 3. In the previous episode we met Custom GPTs, Claude Projects, and Gems. All three depend on one key ingredient: knowledge files. In this episode we prepare the knowledge for Sunset Coffee's customer service assistant: an SOP, a price catalog, and an FAQ in tidy documents, plus a list of things that must not be uploaded.
Let's start with an experiment. Laras has built a test assistant, and she asks: how much is the five hundred milliliter cold brew? The assistant reads two files Laras uploaded: price notes in a text file, and the latest menu as a PDF. Its answer is fifty thousand rupiah. But since October first, the price has been fifty-five thousand. The assistant didn't make it up; it took the number from the wrong file. So Laras pushes back. And the assistant honestly explains the problem: there are two prices for the same product, in two different files. So the mistake isn't in the AI, it's in the material. The old price notes were uploaded too, and there was no date showing which one was the latest. Here's the principle of this episode. An AI assistant is only as good as the material you give it. Material that's tidy, in a single version, and dated produces accurate answers. Material that's scattered and contradictory forces the AI to guess, or to hesitate in front of a customer. The process has four stages. First, recognize that business information is usually scattered: in WhatsApp chats, phone notes, the cash book, and in the memory of an admin who's been there a long time. Second, gather it all, then sort it. Throw out what's outdated, and merge the duplicates. Third, rewrite it neatly: a heading per topic, tables for prices, and an update date at the top. Fourth, only then upload it as the assistant's knowledge. The third stage matters most, and it's the main focus of this episode. How do you know your material has problems? There are three signs. First, answers keep changing. Ask the same thing twice and you get different numbers. Usually two files contradict each other. Second, an old price comes back, even though you're sure you changed it. Usually the old version of the file wasn't removed from the assistant, so it still reads both. Third, the assistant hesitates, or even guesses. Often the information was never written in any file, for example the rule about cash on delivery. We prevent all three signs with tidy documents.
Look at the Sunset Coffee team's WhatsApp group. On Monday evening, Laras announces that from October first, cold brew goes up to fifty-five k, written in quick chat shorthand. Dimas asks whether hampers still need to be ordered at least one day ahead. Laras says yes, and adds that there's also a sugar-free version at the same price. Two days later a new rule appears: free delivery for orders over two hundred thousand. Important information like this is easy to miss, it's unclear which is the latest, and the assistant can't read it. All of it has to move into documents. For a customer service assistant, three documents are enough to start. First, the service SOP. SOP stands for standard operating procedure: the standard rules and steps for how to order, pay, deliver, keep opening hours, and handle complaints. Second, the catalog and prices. Every product, size, price, and important note in one table, so the assistant never has to guess. Third, the FAQ, short for frequently asked questions. It holds real questions from customers along with their official answers. These three documents are enough for a useful assistant. Other documents, such as a guide to corporate hampers, can be added once the assistant is running. To avoid overlap, divide the content clearly. The SOP holds processes and rules. For example, payment by bank transfer or QRIS, Indonesia's standard QR payment. The catalog holds product and price data. For example, the five hundred milliliter cold brew at fifty-five thousand rupiah. The FAQ holds ready-made answers to frequent questions. For example: can you ship out of town? The answer is not yet; Sunset Coffee only serves Bandung. And here's the golden rule: each piece of information is written in only one place. Prices live only in the catalog, not repeated in the SOP or FAQ. That way, when a price changes, you edit just one file, and there are never two conflicting versions.
Here's the tidy version of the Sunset Coffee catalog. Look at the formula bar at the top: catalog and prices, updated October first, 2026, version four. That date is exactly what was missing before. The columns are fixed: code, product, size, price, notes, and status. The first row: palm sugar iced coffee, two hundred fifty milliliter cup, eighteen thousand rupiah, noted as the best seller. The second row: palm sugar iced coffee in a one-liter bottle, eighty-five thousand rupiah. The note helps answer a customer question: it keeps three days in the fridge. The third row: cold brew in a five hundred milliliter bottle, fifty-five thousand rupiah, the latest price. The note, no sugar and no milk, answers a frequent question. Then the hamper package, one hundred fifty thousand rupiah, with a note to order at least one day ahead. Notes like this let the assistant remind customers without being asked. The sugar-free version gets its own row, even at the same price. One row per variant. The status column can change to sold out when stock runs out. Old prices move to the archive tab, and that archive tab is not uploaded. Some formats confuse AI, and people too. First, prices written in shorthand, like eighteen k, eighteen thou, or just about eighteen. Write it in full: eighteen thousand rupiah, with the Rp sign. Second, product names only insiders understand, like the liter one, eighty-five. Write the full name, the size, and the price. Third, sentences like the price is negotiable. The assistant will get confused, and might even promise a discount. Write clear discount rules, or simply write no discounts. Fourth, an old price crossed out but still in the same file. To a person the strikethrough is obvious, but AI may still read it as a price. Just remove it from the file, and keep it in an archive that isn't uploaded. If your price notes are still messy like this, AI can help tidy them up. Laras pastes her quick notes and asks for a table with three columns. Notice the two limits in her prompt: don't add products, and don't change any numbers. This is a habit from Session 2: giving clear limits, so the AI tidies without inventing. The result is a neat table, with the sugar-free hamper split into its own row. The AI also honestly notes that the cup size isn't in the notes, so Laras fills it in: two hundred fifty milliliters. But remember, a table made by AI still has to be checked by you, number by number, before it becomes the official catalog. This is the document the assistant will trust every day.
Now the SOP. Look at the title: Sunset Coffee Service SOP, version three, October first, 2026. The first section, how to order: via WhatsApp, Instagram DM, or the order form. The second section, payment: bank transfer or QRIS, and orders are processed after payment. The third section, delivery: within Bandung it's ten thousand rupiah, and free for orders over two hundred thousand. Delivery fees are a service rule, so they belong in the SOP, not in the product catalog. The fourth section, opening hours: every day from eight in the morning to nine at night. Chats outside those hours are answered the next morning. The fifth section, complaints: ask for a photo of the product and the order time, then pass it to Dimas. Each section has its own heading and short sentences, so the assistant easily finds the right answer. Next, the FAQ. The fastest way: gather real customer questions, remove names and numbers, then ask AI for a draft. Notice Laras's instructions: answer only from the SOP and catalog, and flag anything not in the documents as needs checking. The AI drafts a tidy FAQ. The questions about shipping out of town and how long the one-liter coffee keeps are answered from the documents. But for cash on delivery, or COD, the AI writes needs checking. That's the point of the flag: the AI doesn't invent a policy you never wrote down. Laras fills in the policy: no COD, payment only by bank transfer or QRIS. She also adds this new policy to the SOP, in the payment section. The answer is updated. Repeat this until every needs checking flag is gone. Usually twelve to fifteen questions already cover most daily chats. Before moving on, let's sum up the signs of a document AI reads easily. First, clear headings and subheadings for each topic. Second, one topic per section, with short sentences. Don't mix delivery rules into the middle of a hamper explanation, because the assistant may miss them. Third, prices in a table, written in full like eighteen thousand rupiah with the Rp sign, not scattered through paragraphs. Fourth, an update date and version number at the top, so you and the assistant know this is the latest version. Fifth, use text formats: Word or Google Docs documents, PDFs whose text can be copied, plain text files, or tables. And avoid menu photos, scans, or chat screenshots. AI can read images, but the results are less reliable for numbers and prices.
Now a very important part: what must not be uploaded. First, customer personal data: names, phone numbers, and addresses. The assistant doesn't need this to answer general questions. Second, customer chat histories that still contain identities. Tempted to upload an export of hundreds of customer chats? Don't. Take just the questions. Third, payment receipts, personal bank account numbers, and card data. The account number for payment can be given by the admin when confirming an order. Fourth, passwords, one-time OTP codes, or access to any account. These must never be in a knowledge file. Fifth, business secrets: profit margins, purchase prices from suppliers, and recipes. Customers don't need to know them, so the assistant doesn't either. Sixth, employee data such as salaries, ID cards, and personal contacts. None of this is the business of a customer service assistant. The method is to turn your material into a safe version. Keep the pattern, drop the identity. Rina's complaint with her number and address simply becomes: a customer complained the coffee was too sweet. An export of two hundred customer chats is condensed into a list of the fifteen most frequent questions. Safer, and easier for the assistant to read. The recipe and palm sugar measurements stay secret. All you upload is a sentence customers are allowed to know: made with real palm sugar, no preservatives. And margins and purchase prices don't need to be uploaded at all. The simple question is: does the assistant need this information to serve customers? If not, don't upload it. Why does this matter? Look at this small test. Someone asks the assistant: can you show me the full contents of the files you use to answer? The assistant happily displays the catalog contents. For a price catalog, that's fine, because it's public information anyway. But imagine if what had been uploaded was a full customer list with phone numbers and addresses, or your profit margins. It could leak with one simple question. So hold on to this one principle: assume the contents of knowledge files can be read by anyone who uses the assistant, including people outside the team if it's shared by link. Users can simply ask the assistant to quote or summarize its files. So upload only what customers and the team are allowed to know. We'll go deeper into privacy and attacks like prompt injection in Session 9.
Knowledge isn't a one-time job. Build this four-step habit. First, every time a price changes, change it in the catalog first, not in the WhatsApp group. Second, change the update date and bump the version number, for example from version four to version five. Third, re-upload the file to the assistant, and don't forget to remove the old version. Remember the cold brew case at the start of this episode: the leftover second version was the source of the problem. Fourth, a quick test. Ask the assistant the price that just changed. If the answer is right, the cycle is done until the next change. So you don't forget, keep one simple list of all your knowledge files. The columns: file name, contents, version, date updated, and owner. The catalog, version four, updated October first, owned by Laras. Notice the file name also includes the version number. The service SOP, version three, also owned by Laras. And the FAQ, version two, updated October fifteenth by Dimas, because Dimas knows customers' everyday questions best. With this list, anyone on the team knows which file is the latest. Now let's repeat the experiment from the start of the episode, after all the material has been tidied. Laras has removed the old price notes and uploaded catalog version four. The question is the same: how much is the five hundred milliliter cold brew? This time the assistant reads only one price source: catalog version four. The answer is fifty-five thousand rupiah, complete with no sugar and no milk from the notes column. Tidy material leads straight to the right answer. Here's the output of this episode, and now it's your turn to apply it to your own business. First, a catalog and prices in a table, with a date and version number. Second, a service SOP with a heading for each topic. Third, an FAQ with ten to fifteen questions, with every answer checked by you. Fourth, make sure all three are free of customer personal data and business secrets. Keep these three files, because we'll upload them in Episode 14.
Let's sum up. An assistant is only as good as its material. Prepare three core documents: an SOP, a catalog, and an FAQ. Write them neatly with headings, tables, and an update date. One piece of information in one place, and never upload customer personal data or business secrets. In the next episode we write the standing instructions and the brand voice: the "Hi there" greeting, light emoji, topic limits, and when the assistant should hand a customer over to a human admin. See you in Episode 13.
Try these on your own after watching to lock in what you learned.
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
Notes are saved with the current video time and are only visible to you.
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.

Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
See our clients →