ChatGPT Atlas vs Agent Mode: Which is Better for BI Data Prep?

31 Dec, 2025

Hello, fellow data analysts. I'm Laiken.

Whether you're using Tableau or any other BI tool, data prep is arguably the most tedious part of the analysis workflow. Finding open data, evaluating it, downloading it, and transforming it into a usable format—it all takes forever. Your heart is already breaking before you even get to the fun part: exploratory data analysis.

However, as previous experiments have shown, consumer-grade AI tools can handle most of this work while you sit back and sip your tea. (Of course, human verification is still needed afterward.)

For example, ChatGPT's Agent mode can autonomously handle tasks like these and deliver a CSV link at the end:

• Data discovery: Finding open data from government statistics portals
• Acquisition
  • Downloading CSV files
  • Reading and converting legacy Excel files (.xls)
  • Estimating values from available data when desired data isn't found
• Transformation
  • Column name cleanup (Japanese → English, removing spaces/line breaks/full-width characters)
  • Applying Japanese-to-Western calendar conversion dictionaries
  • Pivoting (wide to long format)
  • Forcing text-to-number conversion, handling missing values with NaN or 0 based on context
  • Annual aggregation

Claude can do something similar, but ChatGPT's Agent mode currently excels at directly browsing websites. (This could change tomorrow, of course.)

1. Experiment Goal: Is Atlas Better Than Agent Mode?

That was a long preamble, but the above is already old news—even a kid picking their nose can do it nowadays. If you haven't tried ChatGPT's Agent mode yet, you should feel a sense of urgency. Staying aware of AI's rapid evolution is crucial.

In October 2025, OpenAI announced "ChatGPT Atlas," an AI-powered browser. The key question for this experiment: Does Atlas offer something beyond Agent mode's capabilities?

Let me organize the timeline:

Operator has been discontinued as a standalone service and is now available as "Agent mode" within ChatGPT and Atlas.

Why Compare These Two?

Atlas has Agent mode built in. However, I noticed that using Agent mode within ChatGPT alone behaves differently from using Agent mode within Atlas.

Atlas can pass the context of pages you're currently viewing to Agent mode, enabling more context-aware task execution. In contrast, ChatGPT's standalone Agent mode starts browsing autonomously from scratch.

The purpose of this comparison is to understand how this difference affects data prep workflows.

2. Atlas: The Limits of Vision-Based AI

I dove right into Atlas. This time, I asked it to find oil consumption and production data for India. Long story short: Atlas wasn't suited for this task.

For reference, here's the prompt I used. It was voice-typed, so please ignore the typos—LLMs are smart enough to understand user intent despite minor errors!

In different regions of India, they use different vegetable oils, right? Like gingelly oil, coconut oil, ghee, butter-type stuff—the cooking oil varies by region. Some might be flavoring oils, some cooking oils. There's a lot of diversity. Can you find official open data that shows this, and prep it for analysis in Tableau later? Consumption and production data would be great.

The File Download Wall

Atlas is optimized for "browsing," making CSV downloads difficult to complete. OpenAI's technical documentation explicitly states: "System access: Cannot run code in the browser, download files, or install extensions." File acquisition is a design limitation. Yes, this is what happens when you play with toys without reading the manual. My bad.

Additional Test: Viz Interpretation Limits

Since Atlas can't download data, it's practically useless for prep. But if it's specialized for browsing, maybe it can extract insights from published dashboards? I decided to test whether Atlas could interpret Tableau Public visualizations. Perhaps Atlas has improved at handling scroll bars and other human-oriented UI elements that ChatGPT's Agent mode struggles with.

I had Atlas load a Viz of the Day from Tableau Public and analyze it. You can see it summarizing its analysis in the sidebar on the right:

The results were disappointing: numerical hallucinations occurred, and it couldn't correctly interpret dynamic elements like filters and parameters.

Why Can't Atlas and Agent Mode Properly Operate Vizzes?

Tableau dashboards are optimized for human visual cognition. Colors, layout, and interactions appeal to human intuition, but for AI trying to interpret screenshots, they're just opaque collections of pixels without DOM structure.

OpenAI's technical documentation states: "Atlas's computer use model takes a single screen image as input." Given this design, there are inherent limitations in interpreting unstructured image data.

Atlas's Redeeming Quality

This has been pretty harsh on Atlas—Sam Altman might be in tears—so let me mention one good thing: it can log into websites.

Many websites require authentication to view data. Atlas can navigate to those sites and evaluate the data—something previous Agent mode couldn't do.

3. ChatGPT Agent Mode: Reliable Task Completion

Meanwhile, ChatGPT's Agent mode delivered expected results with the same prompt.

What worked:

• Accurate navigation: Precisely reached target data on government statistics sites
• Data acquisition and prep: After obtaining CSVs, seamlessly launched Code Interpreter (Python environment) to complete cleaning, aggregation, and wide-to-long transformation in one go

Agent mode doesn't just operate web UIs—it can process acquired data on the spot. This integration of "browsing + code execution" is extremely effective for data prep workflows, as previous experiments have confirmed.

Current conclusion: For automating Tableau data prep, ChatGPT Agent mode is the only choice!

4. Machine Readability

The biggest takeaway from this experiment is the answer to: "Why can't AI understand dashboards?"

Tracing why Atlas couldn't operate dashboards leads to one root cause: lack of machine readability.

• Both Atlas and Agent mode can only see screenshots (pixels)
• Tableau Vizzes render as SVG/Canvas inside iframes with no DOM structure
• JavaScript APIs can access metadata, but Atlas can't execute code in the browser

On the other hand, Agent mode succeeded in data acquisition because the Indian government statistics site's HTML was machine-readable, or the task was simple enough to navigate through screenshot interpretation alone. That's why even Agent mode can't properly read or interpret dashboards on Tableau Public.

The web industry is already standardizing the design philosophy of "human-facing UI + machine-readable structured data." This philosophy hasn't been adopted for dashboards published on Tableau Public yet. I'll explore this issue and potential solutions in my next article!

5. Summary

Conclusion

For automating BI data prep, ChatGPT Agent mode is currently the only choice. Atlas is designed for browsing and isn't suited for file downloads or dashboard operations.

However, the fundamental issue revealed by this experiment isn't about which tool is better. It's that dashboards are structured in ways AI "can't read." The concept of "machine readability" that the web industry has cultivated over the years needs to catch up in the world of web dashboards. That's perhaps the biggest insight from this experiment.

That's all from the field!