Greg Eisenberg — OpenAI DevDay: Dots, Agents & $100B Opportunities

In one sentence

The new era of AI is defined by "Dots" (a personal agent platform) and a shift from selling raw tokens to selling specialized, high-value workflows that solve specific customer problems.

Why read this (and who it's for)

This summary is for founders, developers, and "idea junkies" who want to build businesses in the AI space. It solves the problem of "analysis paralysis" regarding how to actually make money from LLMs by identifying the specific infrastructure (Dots, Decisions API, Agents API) and business models (Free-to-Paid, Workflow-as-a-Service) that will define the next wave of billion-dollar companies. After reading, you will be able to identify "workflow" opportunities and evaluate them using the Trigger-Decision-Action-Feedback framework.

The big idea

Greg Eisenberg argues that the most significant shift in the AI landscape is the transition from a "chat" interface to a "personal agent" platform. With the launch of Dots, OpenAI is providing a front door to 1.2 billion weekly active users. This means the opportunity is no longer just about making a better chatbot; it is about building a "plugin" or "mini-app" that can be discovered and hired by a user's personal agent to perform a specific job.

The core thesis is that because OpenAI is now providing the managed infrastructure (via the Agents API and Decisions API), the competitive advantage for startups shifts away from "who has the best model" toward "who has the best vertical data, brand, and specialized workflow." Success in this era comes from identifying niche categories that are too small for OpenAI to build as native features but large enough to support a profitable, specialized business.

Key ideas

The "Dots" Platform & Plugin Economy

Dots is OpenAI's personal agent platform, designed to compete with Meta's Muse and other agentic platforms. It serves as a massive distribution channel where users ask for help (e.g., "get me a permit"), and the platform selects the best plugin to do the job. Why it matters: It creates a "plug-and-play" app ecosystem where your business can be hired by a user before they even know your brand name. In the book: Eisenberg notes that while there were "mixed reviews" on initial plugins, OpenAI is doubling down on them to make ChatGPT the primary app ecosystem for AI.

The "Sign in with ChatGPT" Opportunity

Users can now sign in to third-party apps using their ChatGPT login and token allowance. Why it matters: This removes the "inference tax" barrier for startups. You can now offer a "free" core app because the user is using their own plan, allowing you to focus on building a massive top-of-funnel, brand, and data set before upselling to premium features. In the book: Eisenberg compares this to the "Sign in with Facebook" era, which birthed multi-billion dollar companies like Yelp by providing a seamless way to access a user's social graph and identity.

Decisions API vs. Jev

The Decisions API allows developers to focus an LLM's intelligence on a specific set of user-defined questions with finite, predefined answers. Why it matters: It is a tool for classification, routing, and choosing an agent's next move. It allows for more reliable, "closed-loop" logic in complex workflows. In the book: Eisenberg highlights the potential for "UI design" applications—using a decisions API to generate high-converting app store screenshots or direct-response websites.

Managed Agent Infrastructure (Agents API)

The Agents API supports "computer use," allowing agents to interact with software to complete tasks, while providing managed infrastructure for tool search and context compaction. Why it matters: It allows developers to stop worrying about the "plumbing" of agentic behavior and focus on the "vertical harness"—the specific industry knowledge and data that makes an app useful. In the book: Eisenberg describes this as a "managed Conex harness," where the winner is the one who understands a vertical better than anyone else.

Selling Workflows, Not Tokens

The business model for the agentic age is moving away from charging for "usage" and toward selling "measurable business results." Why it matters: You want to own the "proprietary information," "specialist workflows," or "network of humans" that the AI uses to get the job done. In the book: Eisenberg lists the five pillars of value: proprietary information, specialist workflows, access to outside systems, collaboration/auditability, and a network of humans/suppliers.

Frameworks & tools

The Trigger-Decision-Action-Feedback Framework

Use this 4-step mental model to brainstorm and validate any AI business idea.

1. Trigger: Identify the starting point. What happens that makes a user go to ChatGPT? (e.g., "An invoice becomes due," "Inventory runs low," or "It's tax time.")
2. Decision: What is the logic? What does ChatGPT need to decide? (e.g., "Which vendor should we pay?" or "How should we categorize this expense?")
3. Action: What is the actual execution? This is the "Submit" button. (e.g., "File the permit," "Update the Shopify catalog," or "Send the email.")
4. Feedback: How do we know it worked? (e.g., "Was the permit approved?" or "Did the customer reply?")

Note: Eisenberg argues the strongest businesses will own at least two of these three: Trigger, Action, and Feedback.

Memorable examples & stories

The "Free" App Strategy

By using "Sign in with ChatGPT," a founder can offer a core app for free because the user pays for the inference. The Lesson: Focus on building the brand, the distribution, and the data set first; only charge for "enterprise stuff," "team sharing," or "extra data."

The Real-World API (The "Last Mile")

While everyone is building digital tools, there is a massive opportunity in "Real-World APIs"—connecting a digital agent to a human specialist. The Lesson: If a user asks a personal agent to "clear a customs shipment," the winner is the one who can connect that request to a vetted human customs broker.

The "Profound" of Agent Analytics

Just as Mixpanel and Google Analytics became the standard for the web, there is a massive opportunity to build the analytics layer for "Agent Discovery." The Lesson: There is a huge need for tools that measure "activation share by customer intent," "tool definition problems," and "invocation-to-completion" rates for plugins.

Key quotes

"You can be hired before the customer knows your name." Why it matters: This highlights the power of the "Dots" distribution—if your plugin solves the job, the platform will surface you.

"You want to build these products that are too niche to become major OpenAI features, but you have enough of an audience that you can charge... something for upselling and for that workflow." Why it matters: This is the "Goldilocks Zone" for AI startups: too small for the giants, but big enough for a profitable business.

"Sell workflows instead of tokens." Why it matters: This is the fundamental shift in how you should think about your business model in the agentic era.

How to apply it

This week

• Audit your "Vertical": Identify one specific, repeatable workflow in a business you understand (e.g., "Shopify catalog cleanup" or "Contract review for franchise agreements").
• Map the Framework: Take that workflow and map it out using the Trigger-Decision-Action-Feedback framework.
• Identify the "Action": Determine what the "Submit" button looks like. Does it require a real-world human? Does it need to connect to a specific API?

Ongoing

• Focus on "Managed" Infrastructure: When building, use the Agents API to handle the "plumbing" (tool calling, context compaction) so you can spend 100% of your time on the user experience.
• Build for "Sign in with ChatGPT": Design your product to be "free" to start, leveraging the user's own token allowance to lower the barrier to entry.
• Think "Real-World": Whenever you think of a digital task, ask: "Is there a physical/real-world component (a permit, a shipment, a physical inspection) that an agent can't do alone?"

Common pitfalls

• The "Token Trap": Trying to build a business based on selling "access" to a model. In the age of Dots and Agents, the model is a commodity; the value is in the workflow.
• The "Too Big" Trap: Trying to build a tool that is so broad that OpenAI will eventually just turn it into a native feature. You must find the "niche" that is too specific for a general-purpose model.
• Ignoring the "Action": Many people focus on the "Decision" (the LLM's logic) but fail to build the "Action" (the actual integration with the real world).

Connects to

• The "Platform" Play: Similar to how Facebook Login created a massive ecosystem for developers, "Sign in with ChatGPT" is the new gateway for AI app distribution.
• Vertical AI: This reinforces the idea that the "moat" in AI is not the model, but the proprietary data and the specific vertical expertise.

If you remember only three things

• The Distribution Shift: "Dots" is a personal agent platform with 1.2 billion users; your goal is to be the "plugin" they hire.
• The Business Model: Use "Sign in with ChatGPT" to offer a free core app and focus on selling high-value, niche workflows.
• The Framework: Use Trigger-Decision-Action-Feedback to identify where your business actually provides value.

Internalize it

• Reflection questions
  • Which "boring" vertical do I understand better than anyone else?
  • If I were to build a "plugin" for a