OpenAI’s Dot Is an Enterprise-Style Agent That Can Order Dinner

Introduction

OpenAI’s Dots platform arrives with the familiar trappings of consumer AI: a friendly blob-like avatar, a customizable name, and an interface that lets users watch an agent work in a separate window. In practice, however, Dots feels less like a personable shopping companion and more like workplace software with a cute front end.

The system operates through a virtual computer and can use applications such as Blender and GIMP. With permission, it can also reach a user’s own computer through the ChatGPT desktop app. A phone-based calling feature lets users talk through tasks while away from their desk. OpenAI is initially making Dots available to higher-tier subscribers, including the $100-per-month Pro plan used in the hands-on test. That pricing and rollout say as much about the product’s intended audience as its friendly visuals do: this is primarily a productivity tool.

What the testing showed

• Dot nearly completed an internet installation appointment and even found a promotional discount in email, but it became stuck at a human-verification step that required holding a mouse button.
• After the user took over the browser, the process reached a payment screen asking for bank account information. Dot paused, but the tester chose not to enter sensitive details into the virtual browser.
• Other consumer tasks produced mixed results. Dot missed a free coworking-space option, failed to log into an Ikea account because of a looping security check, and could not access a local restaurant’s ordering page.
• The agent performed much better on tasks inside the user’s own environment. It redesigned a personal website based on spoken feedback, combined video files, and resized the result for social media.
• Once connected to the desktop, Dot could click through a website backend and recreate the prototype page. The workflow worked, but it required a substantial set of permissions.

Why the distinction matters

The results highlight a basic problem with general-purpose agents: understanding an instruction is only one part of completing it. The agent must also survive login flows, fraud prevention systems, bot detection, payment requirements, and websites that were never designed for automated browsers. Shopping and account management are especially difficult because they combine technical friction with privacy and financial risk.

Dots’ strongest use cases are therefore not necessarily the most consumer-facing ones. Website maintenance, media preparation, and other tasks inside a user-controlled environment have clearer boundaries and are easier to review or undo. In that sense, “Codex for regular people” is a useful description. Dots is less a universal online concierge than a system that can take a defined piece of someone’s work and carry it out across software.

The broader promise is a new division of labor: a user describes the goal in natural language, while the agent operates a computer and returns a result. For that promise to become dependable, OpenAI will need to improve permission controls, site compatibility, sensitive-data handling, and the points where human confirmation is required. Cute avatars may make the technology approachable, but reliable boundaries will determine whether it can become everyday infrastructure.

Source: The Verge AI