Execution (1): Ideas and Assets#

This is where the execution part begins. It follows, in order, the process of building, running, and promoting a product on your own. The first step is “what to build,” and this chapter is about how to answer that question.

Why now is a good time for individuals#

Every time a new technology has emerged, new opportunities have opened up. With the spread of the internet, the arrival of the smartphone, and the rise of Bitcoin, each period of change reshaped the industrial landscape and produced new leaders. Now it is the era of AI. Many opportunities are open, and I think this is a better time than ever for an individual to take a shot.

One of the key elements of economic activity is capital, and capital moves toward wherever value is. Many people worry that “AI will replace humans,” but more precisely, AI is replacing or restructuring parts of what humans used to do, and “the shape of work” is changing. The people who create something valuable in that current will be the ones who attract capital.

The whole sequence of coming up with a business idea, planning and developing a product, launching it, running marketing, and analyzing data to improve the product has become far more accessible than before. What has changed most is the cost of trying. In a small organization of the past, there were few people and development took a long time, so the core strategy from the outset was “which one thing do we pick and pour our strength into?” Today it is much easier to build a small feature or an MVP quickly, drop it if there is no response, and try another direction. Focus has not disappeared; rather, the moment at which you choose and focus has been pushed later. You can cast several fishing lines and then concentrate on the spot where the fish bite.

The premises I relied on while sorting out ideas are below. They are also a summary of what the fundamentals part covered.

• AI understands queries in natural language and generates responses in natural language, handles more than 100 languages, and is strong at translation.
• AI generates images / video / music, understands and analyzes images, and converts between speech and text. However, precise control over what it generates is still limited.
• AI is trained on vast data and can answer many questions, but it does not know unpublished information and has a hallucination problem: it can be convincingly wrong. RAG, fine-tuning, and function calling are ways to handle information it does not know, and conveying context well matters.
• AI can act as an agent that calls tools repeatedly and carries a job through to the end, and coding agents have already become the main hands that build products.
• High-performance models are available via API, there are many open-weight models, and at this very moment services that solve all kinds of needs with AI keep appearing nonstop.

Identify and leverage your assets#

When coming up with business ideas, it is just as important to first identify the assets you already have and carefully examine how you can use them as it is to imagine “what will take off?” The assets here may be tangible, visible ones (a business, products, inventory, space, equipment, distribution channels, and so on) or intangible, invisible ones (knowledge, experience, content, a network, trust, personal abilities, character / taste, repeatable work know-how, and so on). It may sound “obvious,” but ideating with a vague sense of them in the back of your mind and listing your assets item by item and concretely checking their strengths / constraints / room to expand produce very different results.

Tangible assets are relatively easy to identify and analyze. If you already run a business, the question connects directly to “how do I apply AI to this business to lift its performance?” If you run a farm, you might think of demand forecasting, work-schedule optimization, or automating customer responses / order handling; if you run a restaurant, you can look for automation points in menu design / cost management / review analysis / reservation handling. If you sell products, you can go further and concretely design how to use AI for content that makes the product look more appealing, marketing copy per target segment, ad experiments, customer inquiries, summarizing reviews and extracting improvement points, and more. What matters is not “let’s do something with AI,” but finding the points of application based on where the bottlenecks are in the structure you already have (channels, customers, products, operations).

Intangible assets, by contrast, are harder to identify and analyze. Because “where they can be used” is not obvious, you have to consciously draw them out and give them shape. If you teach, you can think at once about how to reach more customers with your lectures (repurposing content, expanding promotion channels, messaging per customer segment) and how to make better teaching materials with less effort (generating outline drafts, generating examples / exercises, automating material updates). If you are “someone with the flair to be a YouTuber who gave up because filming and editing were too much,” you can try again by using AI directly or by dividing planning / scripts / thumbnails / editing among AI-based services. In other words, intangible assets gain value when you find the points where AI lowers the “execution barrier.”

In the end, starting from existing assets often puts you in a better position than digging up a completely new idea from scratch and executing it. An “asset” already carries traces of value and validation in some form, whereas a new idea has not yet been proven to hold real value on its own. So when thinking about business ideas, the effective approach is to first write down your assets concretely, and then narrow down to realistic ideas using questions like “whose problem can this asset solve, and what problem?” and “how much can AI improve cost / time / quality?”

One more thing to note here. Now that anything can be built quickly with AI, others can follow at a similar speed. Features have become easy to copy. So when choosing assets, it is wise to weight the ones that are not easily taken away even when copied: very specific customer problems, know-how embedded deep inside a particular industry’s actual workflows, trust with customers, and structures where repeated use and data accumulation happen. A moat comes less from the feature itself than from relationships, context, and accumulation.

Producing digital content#

If you have no clear tangible assets to build a business on, producing digital content is worth recommending as one of the realistic business paths an individual can choose. Digital content can be started without much initial capital, and once made, the output is easily distributed over and over to create value. Above all, the barrier to entry for “the ability to make content” is falling fast thanks to advances in AI, which makes it an increasingly favorable option for individuals.

Digital content takes many forms. Beyond traditional content like writing, illustration, and video, it includes “content in the form of software,” such as applications and games. Right now, rather than fixing a role first (“I’m going to be a blogger / a YouTuber / a developer”), you can start by deciding what “content” you want to make. As AI begins to assist production across writing, image generation, video editing, code generation, and more, much of the content-production process is turning from a matter of “complete mastery” into one of “proper instructions and iterative improvement.”

Of course, there are still technical limits. Implementing the details in your head 100% faithfully remains difficult, and raising the quality of the output above a certain level still requires human judgment and editing. But what matters is not the limit’s “current position” so much as the speed at which it is crumbling. So the strategy at this point is simple: pick a topic you care about (or a problem you want to solve), choose a content format that fits it, and use AI to run the loop of building small, shipping fast, watching the response, and improving. This blog itself is that experiment, and how it is run is covered in the “Marketing and Growth” chapter.

Aggregating, processing, and reproducing information#

One area where AI is particularly strong is extracting the body of a web page, grasping the key points, and summarizing and restructuring it. In the past, pulling the information you needed from a particular site meant a developer building a crawler, fixing the code every time the page layout changed, and a person hand-editing the summaries. Now requests like “extract the title / body / price / table from this page” or “boil down what these articles all say in common to 10 lines” are far easier to handle with AI, and automating them to run repeatedly is possible too.

This flow resembles how internet services evolved in the past. After online commerce took off, “price comparison sites” appeared; after blogs took off, “meta-blogs” appeared. What they had in common was that they did not copy the original producers’ information as-is, but aggregated information from multiple sources and processed it into a form that let users decide faster. AI explosively raises productivity at exactly this processing stage: classifying, summarizing, comparing, organizing viewpoints, formatting.

To turn this idea into an actual business, it is important to go beyond plain summarization and be clear about “who the summary is for.” You can differentiate by collecting news only from a specific industry (gaming / commerce / real estate) into a morning briefing, or by emphasizing the points needed for decision-making from the perspective of a specific role (CEO / marketer / developer). Following basic principles like source attribution and copyright, and providing links / original text / evidence alongside, raises credibility as well. The key is not to use AI as a “replacement writer,” but to design a system that automates information aggregation and processing and reproduces it consistently. The weekly AI news brief I publish on this blog has exactly this structure: an agent reads the sources and writes the draft, and I review and publish it.

Building a learning or counseling assistant#

Today’s AI is close to a know-it-all that can offer an answer to almost any question, and at the same time a language genius that handles more than 100 languages. Beyond understanding and generating text, it is very good at converting language between forms: translation, speech to text, text to speech. This combination makes it possible to implement in software “a learning experience where a teacher is always right beside you.”

For example, an English-learning service may no longer strictly need a real English teacher. It can take a sentence the user says and correct grammar / pronunciation / naturalness of expression, rephrase the same meaning more formally / more casually, and keep a conversation going through role-play scenarios. Add speech conversion and you can automate “listen - speak - get feedback - repeat,” so a small team or an individual can quickly build and test educational content / learning apps.

Another direction is a “counseling assistant.” AI’s strengths in counseling are that it never tires, keeps the context by summarizing what the user says, and helps the user organize their thoughts through questions. The counseling domain, however, calls for caution. AI can be convincingly wrong, it cannot diagnose a user’s health, and treating problems that need professional intervention (depression / anxiety, self-harm urges, violence / abuse, addiction, and so on) lightly can be dangerous. If you build it as a service, you must make clear that “this tool is not medical or psychological treatment,” provide guidance for crisis situations, set privacy principles, and define the scope of responsibility for its answers.

Building an expert chatbot for a specific domain#

We now live in a world where building a chatbot service with the APIs offered by AI companies is very easy. What matters is not “can you build a chatbot,” but how you give users a reason to use your chatbot instead of ChatGPT or Gemini.

The key is to clearly compensate, within a specific domain, for what general-purpose models do poorly. RAG and function calling can be used aggressively for this. RAG retrieves internal materials such as your company’s documents, manuals, policies, and knowledge base and attaches evidence; function calling lets the model answer based on the results of “actual work” like database queries and integration with external systems. The structure borrows the general-purpose model’s way with words, but raises the accuracy and practicality of its answers with data and functions only you have. However convenient and pleasant a department store may be, someone who wants a special loaf of bread will seek out the shabby, century-old bakery with a line out the door.

The service I am developing at my company is an example of this direction. It is an expert chatbot that helps with ad analysis and execution, with a range of functions wired to a database we built ourselves. It can query campaign performance under specific conditions, analyze the causes, and propose the next action (budget adjustment, target change, creative tests) in one continuous flow. This “internal data + execution functions” combination is value that general-purpose services cannot readily provide, and users get not mere advice but results that actually move the work forward.

Another idea is a customer-service (VoC) chatbot. Beyond giving accurate product information and reducing frequently asked questions, it can be extended to structure and store customer complaints and opinions, and to alert the person in charge when certain keywords (delivery delays, refunds, product defects) pile up past a threshold. The competitive edge of an expert chatbot comes not from the “conversation” itself but from the experience of handling a specific domain’s information most accurately and seeing the necessary work through to the end. And as we saw in Chapter 2, if you expose your service as an MCP server, you can bring that experience into the agents users already use.

Summary#

• Since the cost of trying has dropped, run several hypotheses as small experiments before picking one, and focus where the signal comes from.
• Ideas start from your assets. Write the assets down as a list, find the bottlenecks, and find the points where AI lowers the execution barrier.
• Weight the moats that come from relationships, context, and accumulation over features that are easy to copy.
• If you have no tangible assets, digital content, information reproduction, learning / counseling assistants, and expert chatbots are realistic starting points.

The next chapter covers turning the chosen idea into an actual product, that is, how to develop a product together with agents.