ChatGPT로는 못 하는 것들 — AI Canvas로 미디어 모니터링 Agent 만들기

"Can't I Just Ask ChatGPT?"

This is the most common question I hear when introducing AI Canvas. "ChatGPT and Gemini both have web search now — why not just ask them?"

That's fair. They do have web search, and they will give you an answer about recent news. But try this test first.

Type this into ChatGPT:

"Give me 1,000 news articles about Samsung Fire & Marine Insurance."

The response is always the same. You get a summarized narrative organized by category — not the raw articles. ChatGPT can't deliver the full original text of 1,000 articles due to volume and copyright constraints. This isn't a flaw in ChatGPT. It's a structural limitation of the chat interface itself.

AI Canvas's Crawling node works differently. It calls the Naver News API directly and collects article titles, summaries, links, and dates as row-level structured data — a proper table. 100 articles or 1,000, it's all in rows. The LLM doesn't read someone else's summary. It analyzes the actual article data, row by row.

The difference between ChatGPT web search and AI Canvas is "reading a summary vs. analyzing the source data."

The Fundamental Difference Between Chat AI and AI Canvas

The Complete Workflow

Here's what we'll build — a Samsung Fire & Marine Insurance media monitoring Agent.

News collection is the entry point. Collected articles are classified as positive/neutral/negative by an LLM. Results are visualized as a pie chart, and negative articles are isolated for a separate risk report. The full summary report is sent automatically by email.

Step 1. Collect News — Crawling Node

What this node does: Automatically collects the latest articles from Naver News for a specific keyword.

From the left sidebar, find the Crawler under the Service Integration category and Drag & Drop it onto the canvas.

There are only three settings to configure.

Tip: Use a more specific keyword like Samsung Fire earnings to get more targeted articles.

Once collected, you'll have a structured table with columns like title, description, link, and postdate. This is the real input source you'll pass to the LLM — articles actually published today, not guesses from a model's memory.

Step 2. LLM Sentiment Classification — Prompt Node

What this node does: The LLM reads each collected article and classifies it as positive, neutral, or negative.

Drag & Drop a Prompt node from the Service Integration category and connect it to the Crawling node. Enter the following:

{{title}}

{{description}}

-----
From Samsung Fire & Marine Insurance's perspective, classify this news article as positive, neutral, or negative.

##Rules##
- Answer with only one of: positive, neutral, negative

{{title}} and {{description}} are automatically filled with the actual article content from the previous node. The LLM reads this real text to make its judgment — no room for hallucination.

Set the model to gpt-4o-mini. This is a simple classification task, so a lightweight model keeps costs low.

This is the key difference: Asking ChatGPT to "analyze Samsung Fire news" means the LLM reaches into its memory to guess. In AI Canvas, the LLM reads the actual article text collected moments ago. Completely different processes.

Step 3. Data Visualization — Pie Chart Node

What this node does: Automatically visualizes the positive/neutral/negative breakdown as a pie chart.

Drag & Drop a Pie Chart from the Visualize category and connect it to the Step 2 Prompt node.

Run the node and a pie chart renders instantly inside the canvas. You'll see: out of 100 Samsung Fire articles in this period, X% positive, Y% negative.

No spreadsheet, no BI tool needed. The chart is built inside the pipeline as the data flows through.

Step 4. Report Writing and Auto Email Dispatch

What this node does: The LLM summarizes the full dataset into an analysis report and sends it to a specified email address automatically.

Add two nodes:

① Report Writing Prompt Node (connected from Crawling node)

{{dataset}}

---
Based on the above data, write a media monitoring analysis report for Samsung Fire & Marine Insurance.

{{dataset}} inserts the entire data table from the Crawling node into the prompt. The LLM reads all 100 articles and summarizes them. Use gemini-2.5-flash — it handles long-form data well.

② sendEmail Node (connected from the Report Prompt node)

{{date()}} automatically inserts the execution date. Run it daily and the subject line becomes "Samsung Fire Media Report - December 20, 2025" automatically.

Tip: You can add a separate risk-focused report using only negative articles. Filter with a selectRows node where output_response = negative, then connect to a separate Prompt. This is exactly what the example canvas does.

Step 5. Schedule — Automatic Daily Execution

What this feature does: Runs the workflow automatically at a specified time. Set it once and the report arrives in your inbox every morning.

Click the Schedule icon in the left sidebar.

Save and you're done. Every morning at 8 AM, that day's Samsung Fire news is automatically collected, analyzed by the LLM, and delivered to your inbox. No one needs to do anything.

vs. ChatGPT: To repeat the same prompt in ChatGPT every day, you type it manually every day. There's no schedule. In AI Canvas, one configuration runs indefinitely.

Step 6. Collaboration and Canvas Cloning

Collaboration: Share your workflow with teammates.

Click Share in the top right → enter email addresses → set permissions (view or edit).

Teammates click the invite link and work on the same canvas together. No code, no server, no IT request required.

Canvas Cloning: Reuse the workflow for different keywords.

Click Clone at the top of the canvas. An identical canvas is created. Change the Crawling node keyword to Hyundai Marine & Fire or DB Insurance and a competitor monitoring workflow is instantly ready.

No need to reconnect nodes or rewrite prompts. Clone → change keyword → run. Three steps.

Final Checklist

Before calling your media monitoring Agent complete:

• Crawling node keyword is set to the company you want to monitor
• Prompt node contains {{title}} and {{description}} variables correctly
• Prompt node and Crawling node are connected with an edge
• sendEmail node has a recipient email address entered
• Schedule is set with the Crawling node as the start node
• You've run it once manually and confirmed the email arrives correctly

Closing Thoughts

ChatGPT and Gemini are excellent tools. But their fundamental nature is text generation. You can't control the input source, can't repeat execution automatically, and can't structure or automate the output.

AI Canvas uses LLMs as components in a pipeline. Data collection, filtering, classification, visualization, delivery — the entire flow runs automatically inside one workflow. And all of it without a single line of code.

After a decade of leading enterprise AX projects, the most important lesson I've taken away is this: what matters in AI adoption isn't which model you use — it's how reliably you can ground the AI in real input data. The workflow in this tutorial is the simplest possible expression of that principle.

To clone this canvas and try it with your own keywords → View the Canvas

Enterprise inquiries and demo requests → Request a Demo