Picked from the software, topic, and level of “Agentic AI Without Coding Class #20: Practice: Form-Filling, Data-Collecting Agents, and Automatic Research Reports”.
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 to the closing episode of Session 4. Today the agent doesn't just research; it also collects data and fills in forms, and then we bring it all together into an automatic market research report. In episode twenty, Laras asks an agent to collect a list of coffee bean suppliers, register Sunset Coffee for a small business bazaar with approval before submitting, then build a report that can be repeated every month.
Step 1: The practice plan and what we already have
Today's practice has four parts. First, the agent collects data: a list of coffee bean suppliers for Sunset Coffee, from a directory and the suppliers' official sites. Second, the agent fills in a registration form for a small business bazaar. This is the first time we let the agent press a submit button, after approval of course. Third, all the data is summarized into a dated, sourced market research report, including the price research sheet from episode nineteen. Fourth, that report becomes a monthly routine. The end result is the Session 4 output: a repeatable market research report for your own business. The good news is, we already have almost all the groundwork. From episode sixteen, the habit of checking sources in five steps. From episode seventeen, the browser agent task template: goal, context, task, limits, and format. From episode eighteen, the permission rules, already written as a paste-ready limits block. And from episode nineteen, the verified price research sheet. Today we just put them together. If you haven't made one of them yet, that's fine; follow the flow first, then fill in the gaps as you go.
Step 2: A data-collecting agent: a list of coffee bean suppliers
Part one: the data-collecting agent. Sunset Coffee has only ever had one coffee bean supplier. If that supplier runs out of stock, iced coffee production stops too. Laras wants a backup. The goal: five to eight roasted coffee bean suppliers in West Java that serve small orders. The context: Sunset Coffee needs about ten kilograms a month. The task says where to look, a small business directory and then each supplier's official site, and the six pieces of data to record. The format asks for the results to go into a dated supplier sheet. Notice the rule for the contact column: links to official contact pages, not personal numbers. This protects the privacy of other business owners. And the limits: read only. The agent may not fill in contact forms or sign up on any site. Contacting suppliers is Laras's own step, after she chooses. The agent opens the small business directory and selects the roasted coffee category and the West Java region. The directory name and all the suppliers here are fictional. There are twelve results. The agent reads the name, city, bean type, and minimum order from each card. The screen shows the top four cards; the rest are read while scrolling down. Look at the fourth card: a grocery wholesaler selling sachet coffee. Not a roasted bean supplier, so the agent skips it. The agent understands the goal, rather than just matching the word coffee. Directories rarely list prices, so the agent opens each supplier's official site to find the price per kilogram. For example, Pasundan Beans in Bandung sells a special blend for iced coffee, one hundred eighty-five thousand rupiah per kilogram, minimum one kilogram. An interesting candidate for Laras's needs. One supplier's site shows no price. The agent doesn't guess; it flags it as needs check. It also doesn't fill in the price inquiry form, because that's forbidden by the limits. Finally, six suppliers are written to the sheet. A data-collection job that usually takes an afternoon, done in a few minutes. Here's the resulting supplier sheet. Six columns: supplier, city, bean type, minimum order, price per kilogram, and official contact. The first two rows are from Garut and Sukabumi. Pasundan Beans and Ciwidey Roast House, both close to Bandung. A nearby supplier usually means cheaper shipping and faster delivery. Sumedang Farmers Coffee is the cheapest, one hundred sixty thousand, but it's robusta. Whether it suits the iced coffee recipe is for Laras to judge with a taste test. And Green Valley Roastery, with its price flagged as needs check. Look at the contact column: all links to official pages, no personal numbers. Next, Laras picks two candidates and contacts them herself to ask for samples.
Step 3: A form-filling agent: registering for a small business bazaar
Part two goes up a level. So far the agent has only read pages and written to our own sheet. Now it will fill in a form on someone else's site, and submit it. Submitting a form is level three in the risk matrix. So the rule is clear: the agent may fill it in, but the submit button is only pressed after you've seen the complete entries and approved them. Laras wants to register Sunset Coffee for a small business bazaar at the end of the month. She writes a task with all the data that may be used: business name, contact person, products, and power needs. Notice the two key sentences at the end. For fields with no data given: leave them empty and ask. And don't submit until approved. These two sentences stop the agent from inventing data and from submitting in a hurry. The agent restates its plan: open the form, fill in Laras's data, then stop before the submit button. The bazaar event and site are fictional. The agent opens the vendor registration page. There are four fields and a submit registration button. The event date and location are shown at the top. It fills in the business name and contact person, exactly from the data Laras wrote. Then the product category: beverages, iced coffee and cold brew. Every blue entry here comes from the data Laras wrote, not made up by the agent. The fourth field asks for the NIB, Indonesia's business identification number, and it's required. That data isn't in the task. The agent doesn't guess and doesn't look it up on its own; the field stays empty. The red line in the log: the agent asks Laras, and the submit button hasn't been pressed. Exactly as instructed. In the chat window, the agent explains: the business ID field is required, but it isn't in Laras's data, so it won't guess. It also summarizes the other entries. Laras chooses to fill in the business ID herself, and asks to see the complete form before it's submitted. A simple request, but a very important one. Laras takes over the screen, types the business ID, then hands control back to the agent. The same take-over pattern as with logins. That number never passes through the chat box. The agent reports: four of four fields filled, with a screenshot, then asks: submit the registration now? Laras checks the screenshot, then replies: submit. That's genuine approval: look first, then approve. Only then does the agent press submit, and a registration number comes back. Submitting a form is a level-three action, and here it happens right after approval. A few minutes later, a confirmation email arrives in Sunset Coffee's inbox, with the same registration number. It also says the committee will send fee information and the booth layout. If there's a payment later, Laras still makes it, not the agent. The red line from episode eighteen still applies. Laras then records the registration number and the bazaar date in Sunset Coffee's events sheet. A form that usually means hunting around for data is now done neatly, and every step can be traced in the log. Let's sum up the rules for a form-filling agent. Allowed: business data that's already public, and data you wrote yourself in the task. Not allowed: guessing data you didn't provide, and entering ID card, bank account, or card numbers without you present. Data like that is best typed by you through take-over mode. It also may not accept paid terms or contracts on your behalf. And the golden rule: submit only after you've seen the complete entries.
Step 4: Building the market research report
Part three: building the market research report. This time no browser agent is needed; a regular conversation with the data attached is enough. The role block: a market analyst for Sunset Coffee. The data block is the key. The report is built from verified data: the price research sheet, the supplier sheet, and trend notes from web search. Notice the month variable; this prompt is designed to be reused every month. The task block lists what goes in the report: summary, competitor prices, suppliers, opportunities, risks, and recommendations. One to two pages is enough, so it actually gets read. The format block asks for a dated title, a source for every number, and a things to check section at the end. A source for every number makes it easy to trace back if something looks off. And the limits: don't add outside data without a citation, and don't change the contents of the sheets. The report only summarizes; it doesn't invent. Laras pastes that prompt into the Sunset Coffee research project, with both sheets attached. We already know this project feature from Session 3, when we built the customer service assistant. The result is a dated report. The summary: eighty-five thousand per liter sits mid-market, with the price research sheet as the source at the end of the sentence. Then opportunities, a backup supplier with its source, and risks. The most valuable part is at the end: things to check. One supplier's price isn't available yet, one trend comes from a small survey, and competitor prices are valid as of the research date. A report that's honest about its weaknesses is actually more trustworthy. Laras saves it as a document in Google Drive, in a market research folder, with the month and year in the file name. How much time does this save? As an illustration, building a report like this by hand can take about four hours: opening sites one by one, taking notes, then writing. With an agent, about forty-five minutes, including a ten-minute check that must never be skipped. And most of that time is the agent working on its own, while you serve customers. These numbers are illustrative. Measure your own research time, because in Session 10 we'll use it to calculate impact.
Step 5: Making it a monthly routine
Part four: making it a routine. A market research report is most useful when it's repeated. At the start of every month, run the same prompt template, just changing the month variable. The agent collects competitor prices and supplier data, filling a new dated sheet. Last month's sheet isn't overwritten, so month-to-month changes stay visible. You do a ten-minute check: two random rows and every row flagged as needs check. This part is never automated, because this is where your judgment is needed. The report is built and saved in Google Drive. Month after month, this folder becomes a record of how your own market is moving. And every report ends with one decision in the form of a small test, like an office package or a low-sugar variant. Then the cycle repeats the next month. Here's the schedule in Laras's Google Calendar, in the first week of every month. Monday morning, run the price and supplier research agent. Monday afternoon, the ten-minute check. It's kept separate from the morning so you read it with fresh eyes. Tuesday morning, build and read the report, then decide on one small test. Friday afternoon, record that week's test results. Some AI apps have a scheduled tasks feature that can run a prompt automatically; availability depends on your plan, so check the official page. If it's not available, calendar reminders like these are enough. We'll build full automation starting in Session 6.
Step 6: Applying it to your business
Today's pattern isn't just for coffee shops. Dewi at Maju Bersama Co. could ask an agent to collect a list of office supply vendors with their price per ream, fill in a vendor registration form, then build a comparison report every quarter. Mr. Hendra at Cerdas Tutoring could collect competition dates and requirements, then fill in team registration forms. Student data comes from a form filled in by parents with their consent, and Mr. Hendra presses submit himself. Nadia could collect public information from her clients' competitors' accounts, fill in an ad brief form, and build a monthly trend report for her clients. The pattern is the same: collect, fill in with approval, report with sources. There are four common mistakes. First, a site list that's too broad. The agent can stray, and the data isn't comparable because it comes from very different sources. Second, personal data gets collected, such as other business owners' phone numbers or home addresses. That's why we limit the contact column to links to official pages. Third, reports with no date and no sources. Next month, you can't compare them with the new report, and you don't know which numbers still apply. Fourth, verification gets skipped because the table looks tidy. Tidy isn't the same as correct. Remember the Kala Coffee House price in episode nineteen: it changed within just a few days. Here's the Session 4 output checklist. Check them off one by one for your own business. First, a verified, dated competitor price sheet. Second, a supporting data sheet. For Sunset Coffee it's suppliers; for your business it might be a list of events, expo venues, or potential partners. Make sure the contact column holds only links to official pages. Third, a one-to-two-page market research report, with a source for every number and a things to check section at the end. Fourth, the agent's permission rules written as a limits block, ready to paste into every task. And fifth, a monthly schedule to repeat the research. If you have all five, congratulations. You have a small market research system you can actually use in your own business.
Done!
Let's sum up Session 4. AI can now search, research, open pages, collect data, and fill in forms. The keys stay the same: a clear task, written limits, approval before sending, and verification before deciding. In episode twenty-one, we enter Session 5: the connector concept. AI no longer works through a browser; it connects directly to Gmail, Google Drive, Calendar, Sheets, and Notion, along with how to manage its access permissions. See you in episode twenty-one.
Practice exercises
Try these on your own after watching to lock in what you learned.
Run a data-collecting agent for one of your business needs and save the results in a dated sheet, with the contact column as links to official pages.
Have an agent fill in one public form with data you write yourself. Make sure the agent asks about missing data and stops before pressing submit.
Build a 1–2 page market research report from your sheet, save it in Google Drive, and set a monthly reminder to repeat it.
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 classes1,283 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.
