ChatGPT at Work: AI Agents for Non-Coders, Promises and Limitations

OpenAI aims to make AI agents a common office tool with ChatGPT Work, offered starting at $20 per month. The company promises entire tasks completed autonomously for non-engineers, but actual usage reveals complex permissions, fragmented settings, and functional limits. The stakes are both commercial and technical: to extend AI beyond coding to increase token consumption, in the face of industry rivals already at work. Initial use cases exist, while the audience remains far from that of ChatGPT online.

Access Frictions and Observed Functional Limits

User tests describe a confusing permission setup to link an agent to a cloud drive, with repeated errors during a simple read-only access and insufficient model assistance. A mobile alert ultimately required full access to function. Several key settings are only accessible on the web, forcing users to switch between the web and mobile applications. Connected to Google Calendar, the tool can create events but not new calendars. Feedback considers the tool to be of little relevance for low-effort tasks, comparable to the work of a very poor intern.

Limited Adoption Compared to ChatGPT Online

OpenAI does not specify the usage distribution between Work and Codex, but the combined application has only 20 million users, while the company claims to have over a billion users interacting with ChatGPT online. For now, the tool is positioned for routine, data-intensive coordination tasks, with OpenAI employees generating weekly reports and transforming spreadsheets into planning tools. Launched last month and accessible from $20 per month, Work still has room to grow beyond this base.

Guided Interface, Buttons, and Skeuomorphism for "Discoverability"

ChatGPT Work adds buttons to select projects and plug-ins, while maintaining a single input inspired by skeuomorphism to aid adoption. Internal debates exist about the utility of these buttons if a single query suffices, but Andrew Ambrosino defends their inclusion at this stage to ensure discoverability, while anticipating their eventual removal. He argues that upstream products are necessary to hope to reach a billion users. The strategy aligns with tools that have popularized "vibe coding" and aims to make functions similar to OpenClaw accessible through natural instruction. It takes into account that most users do not use a CLI and that an agent must work with legacy tools and sites. Ambrosino considers experimentation vital; he himself has granted broad access to his personal services and accepts a risk he claims he has not yet had to face.

The Economic Equation: Tokens and Industry Rivals

Agents operating over long durations consume more tokens, which increases the value per user for OpenAI. Extending usage beyond coding, which remains a narrow segment, is deemed crucial for recouping investments in training and computation. In the field, specialized players like Harvey in law and Clay in sales are advancing with an agnostic approach to the model, choosing the most effective AI available. Analysts, including Christian Catalini, believe that without key complementary assets to industrialize AI, value will be captured elsewhere.

In Daily Use: Memos, Dashboards, and Graphs

Venture capital investors report assembling communications and analyses into investment memos with agents, while operations teams build dashboards and custom visualizations. Sam Altman uses it to organize his vacations; an OpenAI engineer obtained relevant graphs from technical Slack exchanges. For Akshay Nathan, who leads product engineering, employees face a deluge of information scattered across systems like Salesforce, and the value of ChatGPT is to make this access truly usable. Work thus connects the agent to email, the browser, and multiple SaaS platforms to contextualize its actions.

Reported Use Cases in Testing: Calendar, Analysis, and Monitoring

One user test indicates that the agent was able to extract a preschool calendar from an email and add it to Google Calendar. However, the same tester refuses to grant access to sensitive emails, drafts, or bank accounts, while believing that the tool would gain in utility with more trust. He mentions a self-updating financial dashboard on publicly traded companies, a queryable database on space launches previously built in Python, and a weekly email update on AI research. He intends to continue his experiments.

Technical Promise and Scope: From Harness to Complex Projects

On the technical side, each LLM relies on a harness that filters information, allows tools, and formats responses; it is this harness, enriched with instructions, that transforms the model into an agent for long-term tasks. ChatGPT Work, derived from Codex, thus aims to provide non-engineers with the ability to undertake complex projects, beyond simple question-and-answer interactions. OpenAI presents a broad mission of democratization, with a product launched last month and accessible from $20 per month, targeting office jobs where digital flows dominate. The ambition remains to extend beyond the developer sphere, with OpenAI product managers asserting that the tool can autonomously and securely conduct entire tasks.