OpenAI DevDay: Dots, Agents & $100B Opportunities

OpenAI’s Dots is presented as a personal-agent platform that could compete with Meta’s Muse and “Instinct.” Greg Eisenberg’s central idea is that an agent could interpret a request, route it to a business’s plugin, and complete the task before the user knows which business is involved. Plugins are mini apps, and OpenAI is doubling down on them as a way to turn ChatGPT into an AI app ecosystem. Sam Altman has reported mixed reviews of plugins without explicitly calling them a failure; Eisenberg also cites Altman’s report that ChatGPT has 1.2 billion weekly active users.

Eisenberg sees a substantial business opportunity in agent-driven transactions: personal-agent platforms could handle billions of dollars in transaction volume, with the possibility of trillions in 10 to 20 years. A related opportunity is analytics that shows businesses whether platforms such as ChatGPT, Muse, and Gemini discover their products, choose them for a given intent, invoke them successfully, and complete the task. Such tools could test realistic prompts, identify problems in metadata or tool definitions, track task completion, and evaluate plugin updates. Eisenberg compares the potential category with Mixpanel, Amplitude, and Google Analytics, and points to Profound as an example of a company focused on AI discovery.

OpenAI’s Decision API converts developer-provided text or images into finite answers, supporting tasks such as content classification, request routing, and choosing an agent’s next action. It is in limited preview, and Eisenberg says its quality relative to “Jev” remains to be seen. He sees possible applications in selecting higher-converting app-store screenshots and improving direct-response websites or apps. The Agents API, meanwhile, brings computer use, Codex multi-agent capabilities, tool search and calling, and context compaction into developers’ applications, with OpenAI managing the underlying infrastructure. Eisenberg sees this as an opening for vertical AI products built around specific work, including CAD cleanup, scientific research monitoring, Shopify catalog cleanup, video editing, and franchise-contract review.

ChatGPT sign-in could let users access their existing token allowance on other websites, creating a possible route to distribution and monetization. Eisenberg suggests offering a free core app through ChatGPT sign-in, then charging for advanced features such as enterprise use, team sharing, or additional data. His reasoning is that users with an existing ChatGPT plan may be less reluctant to try another AI app if they can use their current allowance. More broadly, he argues that AI companies should sell workflows and measurable outcomes rather than tokens, differentiating through specialist processes, proprietary information, access to outside systems, collaboration, and networks of people or suppliers.

Eisenberg frames agent businesses around a trigger, a decision, an action, and feedback: an event starts a request or autonomous process, the agent chooses what to do, an application takes action, and the result is recorded. He expects strong businesses to own key parts of this loop, describing the opportunity in terms of controlling two of the stages while emphasizing trigger, action, and feedback. One route is to build a brand and distribution, bring users into an app, prove its value, and then offer related services, potentially extending into email or advertising. Another is a “real-world API” connecting platforms such as ChatGPT or Dots with vetted specialists, including permit expeditors and customs brokers, and earning a transaction fee.

The opportunity to connect agents with real-world services is paired with a constraint: some companies may hesitate to use frontier AI when their data is sensitive. Eisenberg cites private intelligence and codec security cloud among the launches in this area, pointing to security as a concern for adoption rather than assuming every workflow can move immediately onto a general-purpose agent. Across the examples, his thesis is that the largest opportunities come from combining agent capabilities with a specific workflow, reliable execution, and a way to measure the outcome.