As AI Breaks Down Technical Barriers, How Much Should We Learn?

01｜How many professional skills does one person need?

AI is making it possible for us to take part in more work that once belonged to different specialties. Software development is a field I know well, so I will use it as an example to explore this question.

What role do you work in today?

Software Engineer, Business Analyst, System Analyst, or Product Manager?

If you had to develop and run a software product on your own, how many roles would you need to take on beyond the work you already know?

For an OPC (One Person Company), all of this work needs to get done in some way. With AI, automation tools, and external services, one person now has opportunities to take part in work that previously required several people.

But just how many kinds of professional expertise does a software product involve?

Looking at this map, you may recognize some areas you know well, others you have only touched on, and some you know nothing about.

AI is lowering the barriers to working across different specialties.

But does being able to complete a task mean we also need to understand all the knowledge behind it?

02｜If we can do the work, do we need to learn it all?

Over the past six weeks, I have spent a great deal of time on my personal RAG and GraphRAG projects.

These technologies are already closely related to my own field. Even with AI, I still need to keep learning, experimenting, and understanding why things fail before I can form my own judgment about certain technical choices.

What concerns me even more is this: if learning new technology within my own field takes so much time, what happens when I venture into other specialties?

Can we really learn this much? And do we need to learn it all?

Of course, perhaps the premise of the question is wrong.

If specialized Agents can handle requirements analysis, system design, development, and testing, why must I master all of that expertise myself?

Multi-Agent collaboration already exists in real software development tools. Although I have not yet completed an entire project using multiple Agents, it has prompted me to rethink the relationship between learning and execution.

Perhaps we do not need to learn how to do everything ourselves to accomplish work that once required several specialists.

But that raises another question.

03｜If AI proposes the options, who judges them?

Imagine a future in which each of us operates like the owner of a small company.

We could delegate requirements, development, and testing to different Agents. They would propose solutions, explain their reasoning, and even review one another, while the person sets the goals and makes the decisions.

Suppose AI offers three options for a knowledge retrieval system:

• Vector retrieval
• Hybrid retrieval
• GraphRAG

It can explain the advantages and disadvantages of each in detail so that we can compare them.

But what if there is a better fourth option?

Someone familiar with the technology might quickly spot what is missing.

Does that mean someone unfamiliar with the field cannot make a reasonable judgment?

I do not think that necessarily follows.

Even without knowing what vector retrieval is, we can start with basic questions:

What is it? When is it useful? What methods does hybrid retrieval actually combine? Why does GraphRAG need a graph? Is there a simpler way to implement this?

By continuing to ask questions, working through explanations, looking for evidence, and experimenting when necessary, a person may gradually develop a sufficiently reliable judgment.

Perhaps what matters is not only how much expertise we already have, but also whether we can identify the core of an unfamiliar problem, question its assumptions, and explore it actively.

There is more than personal experience behind this line of thought.

In its Future of Jobs Report 2025, the World Economic Forum reports that 69% of surveyed employers consider analytical thinking a core skill. Creative thinking, adaptability, curiosity, and lifelong learning are also valued.^[1]

In a study involving 319 knowledge workers, Microsoft Research found that generative AI use is changing how people engage in critical thinking: some of the work shifts from gathering information and executing tasks toward verifying information, integrating results, and supervising tasks.^[2]

The study found that people who trusted AI more to handle a task were less likely to report actively checking and questioning how AI handled it and what it produced.^[2]

This reminds me that easier access to knowledge through AI does not necessarily mean we will think more deeply.

So, will foundational thinking skills become more important than knowledge of specific technologies?

I do not think we can draw that conclusion so easily.

Professional experience built up over time makes it easier to spot hidden risks, judge how serious a problem is, and recognize solutions that look reasonable but will not work.

Critical thinking can help us enter an unfamiliar field, while professional experience may help us recognize more quickly which questions are worth asking.

Perhaps AI changes how different abilities create value, rather than whether professional experience has value at all.

04｜How long will today's AI boundaries last?

There is an even larger uncertainty: we do not know how much work AI will be able to do next.

Looking back at ChatGPT, its scope has changed considerably—from text conversations to image understanding, voice interaction, data analysis, and web search, then Deep Research, ChatGPT Agent, and ChatGPT Work, which can carry out tasks across files and applications.

Years mark typical stages of capability expansion, not when a technology first appeared; capabilities have expanded through model training as well as through external tools and execution environments.

These changes do not mean AI can already do all work independently, nor that it automatically updates its model parameters every time it uses a tool.

But they remind me not to treat what AI cannot do today as a professional barrier that humans will retain forever.

For example, we may currently believe that reviewing complex system designs requires specialists. But what if future Agents can reliably check designs, compare alternatives, and even identify risks that other Agents have missed?

Knowledge we believe we must master ourselves today may be acquired and used differently in the future.

That does not make learning meaningless.

It simply means that if AI's capabilities are still changing, it is difficult to decide what to learn by predicting what it will never be able to do.

05｜How might foundations, breadth, and depth change our value?

I am beginning to think of individual capability as three interconnected parts.

Foundational thinking skills help us understand unfamiliar problems, examine assumptions, ask questions, and build judgment through exploration.

Breadth of knowledge helps us understand connections between specialties and discover more possible approaches, rather than staying within a familiar technical field.

Depth of experience helps us understand complex problems, practical constraints, and failure modes, and recognize when more careful validation is needed.

These three do not conflict.

Consider a risk I have been thinking about: if my model provider stops offering its service, could I keep using my application? Familiarity with deployment helps, but I have not yet tested multiple providers, local deployment, or a hybrid approach in practice.

The two choices may produce the same results today while carrying different long-term risks. But preparing alternatives also has a cost. Whether it is worthwhile depends on the specific risks and business needs.

In some situations, asking the right questions and obtaining trustworthy evidence may be enough; we do not need to master every underlying implementation detail.

In other situations, however, a lack of expertise and practical experience can make it difficult even to verify whether a result is reliable.

So I do not believe everyone should become a generalist who masters every technology. Nor do I believe deep experience in one field alone will necessarily be enough.

As AI lowers the barrier to basic execution, differences between people may shift toward their ability to solve more complex problems across fields.

Foundational skills, breadth of knowledge, and depth of experience may together shape how far a person can go with AI.

But there is still insufficient evidence to say whether this will actually widen capability gaps between people, or affect long-term earnings and career opportunities.

And individual capability is not the whole story of organizational productivity.

06｜When teams have more AI employees

Even if one person can use AI to take on several professional responsibilities, that does not mean all work should be assigned to one person.

Organizations still need to consider delivery speed, quality, cost, and accountability for the results.

An experimental study published in 2025 examined product innovation tasks at Procter & Gamble. It found that, on these specific tasks, individuals using AI could perform at the level of two-person teams working without AI.^[3]

This suggests that AI may indeed change the benefits of traditional collaboration. It does not establish that one person can replace an entire software delivery team.

In its Work Trend Index 2025, Microsoft introduced the concepts of Human-Agent Teams and the Work Chart: organizing people and Agents more flexibly around outcomes, rather than allocating work only through fixed job roles.^[4]

Future software teams may include more “AI employees” taking on specialized responsibilities.

AI allows individuals to take on more kinds of professional work, but that does not mean the division of labor will disappear.

One person can coordinate several Agents, while a team can use Agents to work in parallel and combine complementary expertise. Different projects may need very different configurations.

Perhaps what changes is not the existence of specialization, but how work is redistributed between people, and between people and AI.

For us as individuals, that brings us back to the original question:

As AI continues to lower professional barriers, perhaps we do not need to learn everything. We need to know what is worth mastering ourselves, what we can delegate to AI, and how to accomplish things with other people.

I do not yet know how these changes will ultimately affect professional value and team size.

This article is a reflection at this stage of my thinking. I hope to keep testing these ideas through further practice with multi-Agent collaboration, engineering review, and the effects of AI capabilities on individual judgment.