From Skills to Income

Income comes from solving a problem someone values enough to pay for. This book teaches you to manage AI Workers and to build your own. Those skills earn money when they produce work people rely on: work that is correct, saves time, reduces errors or improves a decision.

A certification is evidence that you passed an assessment of these skills. A portfolio is evidence that you have done real work with them. Customers and employers pay for the results.

Four ways to earn

The four routes overlap. Many people move between them, or follow two at once.

What each route needs first

A job needs evidence. A certification gets your application read. Portfolio work that shows a real task, the AI Worker you built or managed for it, and how you checked the result gets you hired.

An FDE needs to work well inside someone else's business. You learn how the client's work really runs, find the step worth automating, and ship a small pilot their team trusts. Knowing the client's process matters as much as knowing the platform. An FDE also learns, inside real organizations, which problems are worth a product.

Client work needs a clear scope and a fair price. A project can be priced by its deliverables or by the value it creates, not only by the hour. Plan for the support that follows, because AI Workers need maintenance as models and tools change.

A founder needs a problem that many people share and will pay to solve. The product needs maintenance too, but one maintained product can serve many customers. You do not need every skill yourself. A domain expert who knows the problem and an AI builder who knows the platform can build it together.

The first three depend more directly on your own involvement, and they are where most readers will start earning. The roadmaps on the previous page lead to them. The rest of this page explains how a product built on an AI assistant works, because it is the newest route.

The idea in one line

The interface is the storefront, your server is the cash register, and the AI does the reasoning.

Here is what that means. Many people already use an AI assistant at work.¹ That assistant can be connected to outside tools. If you build one of those tools, you can charge for it.

What you sell is the part the AI does not have: your knowledge, your method or your data. The AI reads what your tool returns and does the thinking.

Read the figure from left to right. Copying the interface does not get anyone past your server, and only your server hands over the part worth paying for.

Two ways customers use your product

Both ways serve individuals and teams alike. What differs is how people work.

Through conversation. A connector links the AI assistant to your server. The customer adds it, signs in, and then simply asks the AI for help. The AI calls your server, and your server decides what that customer has paid for.

A connector can be added on its own. You can also package it in a plugin, which bundles skills, connectors and subagents into one package, so the customer sets everything up at once.² There are no screens to build, but the customer has to know what to ask.

Through a guided app. Many people will never write a good request to an AI. They need a screen with buttons, tables and a guided workflow. On Claude, such an app can be an Artifact that calls Claude.³ On Claude and on OpenAI, your server can also show its own screens inside the chat.⁴,⁵

The app sits on top of the same server and calls the AI to do the work each time someone uses it. On Claude, an Artifact can use connectors on the Pro, Max, Ultra, Team and Enterprise plans, and each person connects their own apps, even in a shared Artifact.⁶

Both ways use the same server. Build the first, and the second is an addition, not a rewrite. But reusing the server does not solve distribution. You still have to get the product to each customer, set it up and keep it updated.

Setup is an important obstacle to adoption. Before anything works, each customer usually has to add your connector and sign in. In a company, an owner must first turn on connectors for the organization. Each person then signs in separately, unless the connector uses one shared credential.⁷

On Claude Team and Enterprise, an administrator can authorize a connector once for everyone through enterprise-managed authentication. It works only if your connector supports it.⁸

For someone who has never done this, setup is a real hurdle, and it falls on the non-technical users a guided app is meant to reach. Write a short setup guide. Then test, in a real customer workspace, how a customer receives the product, turns it on and gets updates.

Why the value stays on your server

Anything that runs on the customer's side can be read and copied. That includes the app's code and the plugin's instructions. A paywall written into either one can simply be deleted.

So the check that decides who has paid, and the material worth paying for, both stay on your server. Customers pay on your own website, through a card processor such as Stripe or a local payment gateway, and your server records who has paid. The plugin and the app never handle card details.

When a customer uses the product, they sign in, your server checks their subscription, and only then does it send the next lesson, rubric or result. This is why you sell access to your server, not the plugin or the app itself.

A server stops casual copying, but it does not make your material leak-proof. A paying customer can read everything you send them, and could collect it piece by piece. So send only what each step needs: one lesson, one rubric or one result at a time, with a limit on how much one customer can ask for. Protect the material, not the page.

Who pays for what

Who pays for the AI's reasoning depends on the platform and on where the reasoning happens. On Claude, when someone uses an Artifact that calls Claude's built-in AI, the usage counts against that person's own plan, not the creator's.³

That covers only the AI inside the Artifact. Any model your own server calls is billed to you. For an AI that you run yourself, such as an open-source model, you pay for the computers it runs on.

Where the customer's plan does pay, the customer pays twice: once for the AI and once for you. Organizations that already pay for an AI workspace may accept this more easily than individuals do, but that is a guess to test with real customers.

There is a catch. Some of your material may be too valuable to send to the customer at all, such as a scoring method that is your whole advantage. Then your server has to do that work itself and return only the answer. If that work needs an AI model, you pay for it.

Keeping something completely secret and having the customer pay for its reasoning do not go together. Decide which material is worth protecting at your own cost.

Your other costs are real too: hosting, security, sign-in, payment fees, support and keeping your material up to date. And where your product uses the customer's AI limits, a careless design can use them up. Customers will blame you, not the AI vendor.

How it works on Claude and on OpenAI

If you will not build a product, skip this section and go on to "Where the book teaches this".

The same server works on both platforms, because both connect to outside tools through MCP, the Model Context Protocol: one standard plug for AI apps. Build your server to it once, and any AI app that supports MCP can plug in. Both platforms also support MCP Apps, MCP's shared way for a server to show its own screens inside the chat.⁹ What differs is how customers find the product, see it and pay for it. Each vendor sets these rules and changes them often, so this section is checked every month. Check them again before you sell.

On Claude

As of October 29, 2026, connectors work like this:

• For everyone, you can list your connector in Anthropic's connector directory, from any paid Claude plan. Anthropic checks it against its directory policy, which does not admit connectors that serve ads, or that move money without a person approving.¹²,¹³
• A connector that is not listed can still be shared with an install link.¹⁴
• On the free plan, a customer can use connectors from the directory and add one custom connector. Plugins need a paid plan.⁷,¹⁵,²
• A connector can show its own screens inside the chat, on every plan.⁴

Artifacts that use Claude are shared like this:¹⁶

• An Artifact that uses Claude or connected apps cannot be shared with "anyone with the link", on any plan.
• Inside a Team or Enterprise organization, you can share it with specific people or the whole organization, and on Enterprise with groups. Each person uses their own plan and their own connections.³
• People outside the organization can be invited by email, a feature in beta, and can open it, but Claude and connectors do not work for them. On Team and Enterprise, an owner must first allow invitations outside the organization.
• On Pro, Max and Ultra, only people you invite by email can open it, as of October 31, 2026.¹⁷ The help page says the parts that use Claude or connectors do not work for people invited from outside your organization, but not what that means on a personal plan. Test this before you sell to individuals this way.
• Legacy Artifacts, made in a chat before September 16, 2026, follow older rules.³

So today, an Artifact that calls Claude works fully for customers inside an organization. A connector reaches individuals too, on every plan, and it can bring its own screens.

On OpenAI

Building it. You build an app for ChatGPT with OpenAI's Apps SDK. The app is your MCP server plus, if you want one, an interface that ChatGPT shows inside the chat, such as interactive cards.⁵ So on OpenAI, the conversation and the guided screens can live in one app. As on Claude, an app can also go in a plugin, which brings apps and skills together.¹⁸

Signing in. Your server signs users in with OAuth 2.1, following the MCP authorization standard. ChatGPT sends a token with every call, and your server checks it before running any tool. If the check fails, your server returns an error that makes ChatGPT ask the user to sign in again.¹⁹ This is where your check of who has paid belongs.

Who pays. Using an app may count toward the user's ChatGPT plan limits, and toward limits the app's maker sets.²⁰ As on Claude, any model your own server calls is billed to you.

Reaching customers. For one company, an admin or owner on a Business, Enterprise or Edu plan adds your server as a custom MCP server, tests it and publishes it to that workspace only.²¹ Admins also decide which apps their workspace may use. In Business, many apps are on by default, and new Enterprise and Edu workspaces start with a selected set turned on.²⁰

For everyone, you submit a plugin to ChatGPT's directory as a verified individual or organization, with a demo account for the reviewers. Plugins that need an extra sign-up step during review are rejected.²²

Charging is the biggest difference. Inside ChatGPT, an app may sell only physical goods, with the checkout on your own website. It may not sell subscriptions, digital products, services or credits, show your plans or promote upgrades.²²

An app may let a user sign in to an account they already pay for, and use what that subscription includes. If a feature needs a higher plan, the app may say so and link to a page that describes your plans, but not to a checkout page. It may not offer a weaker version than your own service, or add a fee that applies only in ChatGPT.²²

So on OpenAI, customers subscribe on your website, outside ChatGPT, and the app only checks what they have paid for. This is the model on this page, with one difference: the app never sends anyone to a checkout for a subscription.

Design for both. When a signed-in user has not paid, return a short message and a link to your plans page, not to a checkout page. That one message works on both platforms.

In open source

MCP is an open standard, so open-source agents can connect to the same server. If you also run an open model yourself, you pay for the AI.

Part VI gives the full rules for all three, each with the date it was checked.

Where the book teaches this

• Part I and Part II teach what an AI Worker is and how to manage one. That is the expertise your customers and employers will pay for.
• Part III teaches the build: skills, plugins and connectors, the server behind them, and hosted agents, on Anthropic, OpenAI and open-source platforms.
• Part VI teaches the business: choosing your vertical, pricing, getting paid, charging for access to your server, and publishing to the Claude and OpenAI directories.
• Parts V and VII take this to the scale of companies and countries, where the challenge becomes return on investment and the capacity to produce and export AI work.

Why this matters

This model removes some of the costs that used to come before the first customer. You do not have to build and host a separate application for the interface, and where the customer's plan runs the AI, you receive no bill for its reasoning.

What remains is a modest server, your material, and the work of finding customers. That puts a product within reach of a training firm or a small consultancy anywhere in the world, and a country whose people can do this can export AI work, not only import it.

Start with a problem someone will pay to solve. Show useful results on a small scale, and earn the customer's trust before you grow. Whichever route you take, that is how skills become income.

Part 0. How AI Works comes next.