Picked from the software, topic, and level of “Agentic AI Without Coding Class #49: Sample Final Projects: Online Store CS, Order Logging, and a Content Generator”.
At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Welcome back. Your automation map, your way of measuring impact, and your thirty-day plan are ready. Now it's time to decide on your final project. In this episode we break down five sample final projects: automatic customer service for an online store, an order logger, a content generator, an office admin assistant, and a weekly sales report agent. After that, we choose the project that fits you best.
Step 1: Five examples, one pattern
These are the five sample final projects participants choose most often. First, automatic online store customer service: it answers customer questions twenty-four hours a day, and passes complaints to a human. Second, an order logging system: orders from WhatsApp, Instagram, and forms land neatly in one spreadsheet. Third, a social media content generator: from ideas to draft captions, with approval before anything is posted. Fourth, an office admin assistant for professionals like Dewi: summarizing emails, managing schedules, and preparing draft replies. Fifth, a weekly sales report agent: data in Sheets turns into a report that's ready to read every Monday morning. We read each example with the same pattern. First, the problem: what often goes wrong, runs late, or gets missed. Second, the flow: drawn briefly on a canvas, with up to six nodes. A good final project is actually simple. Third, the tools: you've already learned all of them in Sessions 2 to 9. There are no new tools in this episode. Fourth, the results: what changes and can be measured. The result numbers shown in this episode are illustrations. You measure your own numbers the way we did in episode forty-seven. As you watch, try picturing your own business or job in each example. Which parts are the same, and which parts would you need to change?
Step 2: Example 1: automatic online store CS
Example one, automatic online store customer service. The problem: Sunset Coffee gets dozens of chats with the same questions every day, and night chats go unanswered. The flow starts with an incoming WhatsApp chat. The chat goes to the CS AI Agent, which holds fixed instructions and the Sunset Coffee brand voice, just like the assistant you built in Session 3. Then there's a checker node: does this need a human or not? For example, a complaint, or a question that isn't in the FAQ. If not, the answer is sent straight back through WhatsApp. If so, it escalates to Dimas. This is the escalation flow from Session 8. Below the agent are the catalog and FAQ tool in Sheets as its knowledge source, and memory so the agent remembers what was said earlier in the conversation. This whole flow is only seven nodes. If you finished the Session 8 practice, most of it is already built. The final project just tidies up the knowledge, the escalation rules, and the measurement. The results show up in WhatsApp. At ten at night, a customer asks whether there's a no-sugar hampers for a gift. The agent answers from the catalog: yes, the no-sugar version is Rp150,000. The customer orders two packages and asks about delivery. The agent calculates correctly: two packages come to Rp300,000, and since that's over Rp200,000, delivery in town is free. Then the customer raises a complaint: yesterday's coffee tasted a bit off. The agent doesn't argue, and it doesn't promise compensation on its own. It apologizes and passes the complaint to Dimas. The result: questions answered twenty-four hours a day, and the admin only handles the cases that truly need a human. Let's sum up example one. The first result: customer questions are answered twenty-four hours a day, including at night. Second, in this illustration about seventy percent of chats are resolved without the admin stepping in. The rest are exactly the chats that matter for a human to handle. Third, complaints are handled by a human in under an hour during working hours. The tools used: the assistant from Session 3, the agent with memory from Session 7, and WhatsApp with escalation from Session 8.
Step 3: Example 2: an order logging system
Example two, an order logging system, this time drawn in Make. The problem: orders come in from many channels, and Dimas copies them by hand. There are frequent typos, and sometimes orders get missed. The first trigger is the order form. The second trigger is a WhatsApp chat or an Instagram DM containing an order. Both feed into one AI step that pulls the key data out of free text: the name, product, quantity, and delivery time. This is the AI step you learned in Session 6. The cleaned-up data is added as a new row in the orders sheet, like the table update tool in Session 7. Finally, the customer receives a summary of the order to double-check. The customer becomes a checker too, so logging mistakes are caught quickly. The result is one order table. Rina came from the form, Budi from WhatsApp. Putri from Instagram, and Joko ordered five cups of iced coffee on WhatsApp. The source column tells Laras which channel is busiest. Look at Andi's row. His message wasn't clear, so the AI didn't guess. The product and quantity are filled with question marks, and the status is needs checking. Better flagged than logged wrong. The expected results: every order in one place, fewer logging mistakes, and Dimas only needs to check the flagged rows. The time saved, as always, is measured yourself with a one-week log. A summary of the tools in example two. First, the form and chat triggers from Session 6. Second, the AI step that extracts data from free text, also from Session 6. Third, the tool that adds rows in Google Sheets, from Session 7. Then two safety rules. Missing data is flagged as needs checking, not guessed. This rule is written clearly in the AI step's instructions. And a confirmation to the customer as a double check. This project suits participants who are only just getting comfortable with workflows, because it doesn't need an agent making complicated decisions yet.
Step 4: Example 3: a social media content generator
Example three, a social media content generator. This is Nadia's problem. She's a freelancer who manages accounts for several small businesses, and every Monday she runs out of time preparing content for all her clients. The flow is triggered by a schedule, every Monday at eight. The researcher agent looks for trends and ideas that are relevant this week. The writer agent turns them into four draft captions, using the prompt templates from Session 2. The reviewer agent checks prices, facts, and brand style. This is the researcher, writer, reviewer agent team from Session 9. Finally, the drafts go to Nadia on Telegram for approval. No content is posted without a human's approval. For other clients, Nadia just duplicates this workflow and swaps in their brand style guide and product list. One pattern, many clients. The results arrive on Nadia's Telegram. At eight oh six, the bot reports: drafts for four Sunset Coffee posts are ready. Nadia just replies approve, or revise with a number and notes. Nadia sees that draft three mentions a discount Laras hasn't approved. She replies: revise three, don't mention a discount. Within a minute, draft three is fixed: the promo is replaced with a story about palm sugar from local farmers. Nadia replies approve. The four posts are scheduled from Tuesday to Friday. The result: drafts ready every Monday morning for every client, and Nadia only spends her time checking and deciding. Let's sum up example three. Every Monday, there are four draft posts ready for each client. Nadia's time per client, in this illustration, is about forty-five minutes of checking and revising, instead of half a day writing from scratch. And zero posts without approval. The tools are the prompt templates from Session 2, plus the agent team and human in the loop from Session 9. If you already built the content pipeline in Session 9, this project just needs polishing.
Step 5: Example 4: an office admin assistant
Example four, an office admin assistant for Dewi at Maju Bersama Co.. The problem: about sixty emails come in every day, and Dewi's mornings disappear into sorting them. Look at the inbox: emails from customers, the directors, suppliers, finance, and a newsletter all mixed together. One of them is from Pelita Bookstore, a bookstore: their order of fifty reams of A4 paper hasn't arrived, and they're asking about the delivery schedule. An email like this needs a quick reply, but the data has to be checked first. Without an assistant, Dewi has to open this email, check the order status, look at her calendar, then write a reply. A small job, but one that repeats dozens of times a day. This is where the assistant gets to work. Dewi uses an Admin Assistant project in Claude that holds fixed instructions and the office SOPs, like in Session 3. She asks for an email summary, a draft reply to Pelita Bookstore, and a meeting slot with the directors. Through the connector from Session 5, the assistant reads Gmail. Then it reads Dewi's Google Calendar to find a free slot. The result is a tidy morning summary: the directors' request, Pelita Bookstore's question, three invoices due, and a Thursday ten o'clock slot. Notice that the assistant is honest that the delivery date needs to be confirmed with the warehouse. It doesn't make up a date. Dewi asks for the meeting invite to be created as a draft, not sent right away. The result: email sorting that usually takes an hour gets done in a few minutes, while decisions and sending stay in Dewi's hands. The key to example four is a clear split of tasks. The assistant reads and summarizes emails, and Dewi just reads the summary. The assistant writes draft replies, then Dewi edits and sends them herself. The assistant creates meeting invites as drafts, and Dewi checks and sends them. And promising a delivery date is not the assistant's job at all. That's Dewi's call, after confirming with the warehouse. A split like this applies the safe access permissions from Session 5.
Step 6: Example 5: a weekly sales report agent
Example five, a weekly sales report agent, the main project from Dewi's automation map. The problem: every Monday, Dewi spends about two hours putting together a report for the directors. The flow is triggered by a schedule, every Monday at seven in the morning. The schedule triggers an analyst AI Agent, whose instructions tell it to compare sales with targets and explain the reasons. The agent uses an AI model as its brain, a sales sheet tool for this week's data, and a targets sheet tool with targets by region. The result is sent as a draft to Dewi's email, not straight to the directors. Look at the status at the top: last execution Monday at seven, succeeded. The agent's instructions also ask it to name its data sources and flag numbers that look odd, rather than guessing the reasons. This is the way to stop AI from making things up that we covered in Session 9. Monday at seven oh three, the draft report is already waiting in Dewi's inbox, labeled draft. It's concise: total sales are up six percent, Jakarta and Surabaya are around target, and Bandung is at eighty-seven percent because two large customers postponed their orders. All these numbers are examples. The agent also flags something that needs attention, the A4 paper stock, and names its data sources. Dewi just checks it, corrects anything if needed, then forwards it to the directors. A report that used to take two hours now takes about fifteen minutes to check. Let's sum up example five. The draft report is ready at seven oh three every Monday, before Dewi arrives at the office. Dewi's checking time is about fifteen minutes in this illustration, compared with about two hours putting it together herself. And no report goes to the directors without a check. The tools are the scheduled workflow from Session 6 and the agent with Sheets tools from Session 7. The same pattern fits Sunset Coffee's weekly sales recap, the quick win on Laras's map.
Step 7: Choosing your project
Now it's your turn to choose. Online store customer service fits if you get lots of chats with repeat questions. Its level is intermediate, and the sessions to review are 3, 7, and 8. The order logger fits if your orders come from many channels. Its level is easy to intermediate, with Sessions 6 and 7 to review. The content generator fits if you post regularly but are often late. The sessions to review are 2 and 9. The office admin assistant fits if you're drowning in email and schedules. It's the easiest, because it only needs an assistant and connectors from Sessions 3 and 5. And the weekly report fits if your data is already in Sheets but the report is still put together by hand. The sessions to review are 6 and 7. A teacher like Mr. Hendra can use the same pattern for a weekly recap of student grades. Not sure which to choose? Use three conditions. First, the project is near the top of your automation map, with the highest impact times ease score. Second, it can go live within thirty days using tools you already know. If you'd still have to learn a new tool, save it for next month. Third, its impact can be measured with one of the three measures: work hours, costs, or customer response. If all three conditions are met, that's your project. It doesn't have to be one of the five examples. A simple project that actually gets used is worth far more than a complicated one that never goes live. Once you've chosen, write a project card like this one. The problem block: which task, how many hours a week, and what often goes wrong or runs late. The flow block: the trigger, the AI step, the action, and the approval point, written with arrows on one line. The tools block: an assistant, connector, workflow, or agent, plus the session where you learned it. The result block: the impact target within thirty days, and how you'll measure it. And the limits block: which actions need a human's approval, and which data must not leave. This card becomes the outline of your presentation in the last episode. Keep this project card in the same document as your automation map, impact dashboard, and thirty-day plan. Together, those four make up your complete final project package.
Done!
Let's recap. All five sample projects follow the same pattern: a clear problem, a short flow, tools from earlier sessions, and results you can measure. Choose a project near the top of your map that can go live within thirty days, then write its project card. In the next episode, the last episode of this class, we polish the project with a go-live checklist, practice a five-minute presentation, and go over the certificate criteria. See you in episode fifty.
Practice exercises
Try these on your own after watching to lock in what you learned.
Choose one of the five sample projects, or your own, that best matches the top of your automation map.
Fill in the project card: problem, flow, tools, target result, how you'll measure it, and security limits.
Draw its short flow in n8n or Make, six nodes at most, then run it once with sample data.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,286 episodes
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.
