The tech world is abuzz with agentic AI, and recently, I put two major players to the test: OpenAI’s ChatGPT Work and Anthropic’s Claude Cowork. My mission? To see how these intelligent agents handle real-world desktop automation, specifically organizing a messy folder of PDFs. While both performed admirably, one key difference left me with a significant concern regarding safety.
ChatGPT Work recently launched, offering capabilities both in the browser and as a desktop app. Unfortunately, this new desktop version has replaced some of the familiar, beloved features of the older ChatGPT Desktop. Setting aside my initial disappointment over these changes, I dove into testing Work’s new agentic functionalities.
Putting AI to the Test: The PDF Challenge
Letting an untested, potentially “hallucinatory” AI loose on my personal files is a daunting thought. To mitigate risk, I replicated an early, controlled test I previously ran with Claude Cowork: organizing the PDFs in my “Downloads” folder. This folder typically holds files for three days before my Mac utility, Hazel, sorts them into a general “PDFs” folder.
For this crucial test, I created a temporary copy of my “PDFs” folder, ensuring the AI would only interact with a sandbox environment. This protective measure was key, as I worried about the AI potentially going rogue and accessing other parts of my system. Thankfully, it stayed within its designated boundaries.
My test folder contained 447 PDF files, a significant increase from the 308 files used in my earlier Cowork test. The goal was for the AI to move beyond simple file-type sorting and intelligently organize these documents based on their actual content. This is where agentic AI can really shine, transforming a tedious task into an automated process.
ChatGPT Work’s Initial Impressions and Discoveries
My journey began by launching ChatGPT Work and pointing it to my isolated test folder. I started with a simple query: “What can you tell me about this folder?” Work’s response was quick and comprehensive, detailing the number of PDFs, their total storage, page counts, main themes, and even encrypted files. This initial assessment was very promising.
One notable advantage for Work emerged immediately: it identified a significant number of duplicate files. Claude Cowork had missed these duplicates during its earlier run, some of which had existed even back then. This indicated Work’s ability to not just compare file sizes and dates but also delve into content analysis, discerning identical files despite dissimilar names. I then instructed Work to remove these duplicates, which it executed promptly.
Renaming and Reorganizing: A Mixed Bag
While Work excelled at duplicate detection, it initially missed a minor detail that Claude Cowork had caught: the presence of many generically named files. When prompted, Work diligently analyzed and proposed new, more descriptive filenames. I appreciated its reasoning process, explaining how it identified generic names and what it chose to rename. This provided valuable insight into its decision-making.
The presentation of Work’s renaming proposals was initially jumbled, requiring a few additional prompts to reformat it into something readable. Once rectified, the AI demonstrated solid name derivations, adding genuine value to the files. Following my approval, Work proceeded with renaming the files, demonstrating its capability for advanced file manipulation.
Next, I tasked Work with organizing the renamed files into appropriate folders. It began by developing a basic taxonomy, recognizing the diverse nature of the files and planning a multi-pronged approach. This agentic problem-solving, breaking down a large task into smaller, manageable steps, was impressive.
Work successfully generated a robust set of categories, mirroring the quality of those discovered by Claude Cowork in my previous test. While there was a slight formatting hiccup with indentation, it didn’t detract from the effectiveness of the categorization. I gave Work the go-ahead to proceed with organizing, which it did efficiently, creating a well-structured hierarchy.
The Major Red Flag: Permission to Operate
Here’s the critical concern that makes me hesitant about ChatGPT Work: it never asked for permission before modifying any files. Throughout the entire process of moving and renaming hundreds of documents, despite being set to “Ask for Approval” mode, Work proceeded without explicit confirmation. This is a fundamental flaw, as this setting is supposed to ensure user oversight before any major changes.
In contrast, Claude Cowork consistently prompts for approval before making significant alterations like moving or renaming files. This crucial difference in user control is what currently gives Claude a considerable safety advantage for high-stakes desktop automation. Without this explicit approval step, users are ceding significant control to the AI, which can be risky.
Performance and Cost Analysis
The entire file reorganization project with ChatGPT Work took 1 hour, 13 minutes, and 6 seconds. While this felt slow in the context of agentic AI, it was undoubtedly faster than a human could have accomplished the same task. In comparison, Claude Cowork often feels snappier, especially when tackling large directories or complex code.
Using my $20/month ChatGPT Plus subscription, which grants access to Work and Codex, this project consumed approximately 11% of my monthly AI usage capacity. This translates to roughly ten such projects before hitting the usage limits. Considering that each project could save me around 90 minutes of tedious manual work, that’s potentially 15 hours saved per month. For $20, saving 15 hours of “mind-numbing administrivia” is a compelling value proposition.
The Verdict: Choose Wisely (For Now)
Aside from the glaring lack of permission requests, ChatGPT Work’s overall quality and capabilities are comparable to Claude Cowork’s. Work excelled at finding duplicates, while Cowork was better at identifying generically named files. Both ultimately achieved the core task of reorganizing my documents effectively.
However, my current recommendation is clear: if you are undertaking high-stakes file manipulation and demand precise oversight, opt for Claude Cowork. Its consistent permission requests provide an essential layer of safety and control. Once OpenAI addresses this critical permissions issue in ChatGPT Work, the choice might simply come down to personal preference between the two AI ecosystems. Until then, caution is warranted with Work for sensitive tasks.
Source: ZDNet – AI
