A few days ago, I posted a short note to my social feed. The last line was:

I may be much closer to replacing myself than I thought.

The point was simple: I used to believe that AI would not affect ordinary people very much in the short term. The truly unimaginable changes felt like an eight-to-ten-year problem. Now I have changed my mind. I think the serious impact may arrive in three to five years, possibly sooner.

This is not a grand prediction.

The thing that changed my mind was a specific working moment. In a real project, AI completed a set of code changes through an issue-driven workflow: more than 1,500 lines changed, more than 40 files touched, and more than 20 commits. My role was not to write all of that code by hand. My role was to verify the result in a test environment and decide whether it could move forward.

This does not prove that AI can ship production systems without supervision. It does not prove that AI can solve every hard problem. It proves something narrower, and more uncomfortable:

When task boundaries, repository constraints, test gates, and human review exist, AI can already take over a real stretch of execution.

The uncomfortable part was not how many lines AI changed. It was that my position in the workflow had changed.

The painful part is not that AI can chat

We have become used to AI chat.

It can answer questions, explain concepts, draft text, and help edit code. When most people see those capabilities, the reaction is often: useful, but not enough. It is unstable. It makes things up. It needs supervision. It is still far from replacing real work.

That judgment was reasonable for a while.

But what I recently felt was not that AI can produce smoother sentences or score higher on another benchmark. The real shift is that AI is entering workflows that are deliverable, reviewable, and traceable.

That is different from chat.

In chat, AI is an answerer. You ask, it responds. It gets something wrong, you correct it. You ask another question, it starts again.

In a production workflow, AI becomes a worker. It receives an issue, understands the goal, breaks down the work, edits files, commits changes, leaves evidence, and waits for verification. It does not need to ask at every step what to do next. It does not only give advice. It creates a result that can be inspected, reverted, and held to a standard inside a real project.

That is the line that matters.

The 1,500 lines are not the point

When people hear “1,500+ lines, 40+ files, 20+ commits,” they usually ask: Was the quality good? Were there bugs? Could it ship directly to production?

Those questions matter.

But they are not the main point. In real work, human-written code also has bugs. Human pull requests need review. Human delivery still needs tests. The question is not whether AI is perfect in one pass. The question is whether AI can now take over a stretch of execution that previously required continuous human effort, context switching, and attention.

My answer has changed.

I used to think of AI as an assistant. It could help me search, generate snippets, or modify local pieces of code. But the project still moved because a human decomposed the task, wrote the code, organized the commits, and prepared the evidence.

Now I am seeing a different pattern:

human defines the issue
AI executes the work
human verifies the boundary

The human has not disappeared. But the human has moved.

If the task is clear enough, the boundary is explicit enough, and the verification method is concrete enough, AI is no longer only “helping me think.” It can begin to “execute for me.”

That is the point:

AI may not first replace a whole person. It may first compress the execution layer inside a person’s job.

Once the execution layer is compressed, many kinds of ordinary work get repriced.

Ordinary people are not at risk because AI is smarter

I increasingly think the real risk is not that “AI is smarter than me.”

That statement is too abstract, and people naturally resist it. They will say AI lacks common sense, does not understand the business, cannot be responsible, and still needs a human to clean up mistakes.

All of that is true.

But the labor market does not only reward who is smarter. It also rewards who can complete a reviewable task faster, cheaper, and more consistently.

Many ordinary jobs are not about inventing a great idea from nothing every day. They include a lot of work like this:

  • Turn a request into executable tasks.
  • Modify files according to existing rules.
  • Move information between systems.
  • Revise based on feedback.
  • Run checks and organize results.
  • Write summaries and hand the work to the next stage.

These jobs require experience, and they should not be dismissed. But they also contain many repeatable actions.

AI will not first appear by replacing a whole person. It will replace stretches of execution. Today it is a set of code changes. Tomorrow it may be a competitor research report, a page draft, a data cleanup, a test pass, or a batch of issue fixes.

When those actions connect, a person may suddenly realize that the space they used to occupy has become smaller.

That is why I no longer believe this is only an eight-to-ten-year problem.

I am not saying every industry or every role changes on the same day. I am saying the mechanism of impact has appeared: AI does not need to replace a whole person at once. It can keep compressing reviewable stretches of execution.

Human value moves toward the boundaries

This does not mean humans have no value.

The opposite is true. The more AI can execute, the more human value moves from “doing every step by hand” to “defining what should be done, what counts as correct, and where the work must stop.”

In that project, I was not useless.

I still had to judge whether the issue was clear, whether the risk was acceptable, whether the change matched the product truth, whether the test environment could verify it, what should not go to production, and where human confirmation was mandatory.

But that is no longer the old version of “I write the code.”

It is closer to this:

I define the task boundary
I design the verification path
I judge the risk location
I decide whether the work moves to the next stage

These actions still matter. They may matter more than before.

The problem is that not everyone will naturally move to that position. Many people will keep tying their value to “I can perform this execution step by hand.” But if AI compresses that step to one-tenth of the time, or if someone who knows how to use AI can coordinate several workers at once, the old value gets recalculated.

That is why the phrase “ordinary work” matters. Most people will not lose to a model in isolation. They will lose to a new workflow: fewer waits, faster execution, clearer verification, and the same or smaller team.

The dangerous part is not noticing the workflow has changed

I do not want this to become a fear article.

“AI will take jobs” is too easy as a slogan. It either creates anxiety or resistance, and then nothing changes.

The more useful point is that AI’s impact on ordinary people may not arrive as one dramatic event.

It will arrive inside workflows.

You will see one colleague stop writing every line by hand and instead use issues to drive AI through a batch of changes. You will see a small team hire fewer people for repetitive execution because one person can define tasks and use AI to generate, check, and revise. You will see an independent builder finish in a day something that used to take a week to reach a reviewable state.

Each change may not look shocking by itself.

Together, they change one question:

What makes an ordinary person hard to replace inside a job?

I do not have a clean answer.

But I am increasingly sure the answer cannot simply be “I can execute.”

A more realistic answer may be:

I can define the task
I can judge the result
I can own the boundary
I can turn AI execution into a deliverable workflow

If a person remains only in the execution layer, their position becomes more fragile.

Why I am writing this down

I am not writing this because I have found the final AI work method.

I am writing it because the experience hit me.

When I saw AI complete a real set of issue-driven changes and my main job became test-environment verification, I realized my old timeline was too conservative.

Not eight to ten years.

Maybe three to five.

Maybe sooner.

But the change is not as simple as “AI replaces people.” More precisely, AI first changes the execution layer of work. Then it forces people to answer harder questions: What am I responsible for? Where do I have judgment? Can I turn a vague request into a task that AI can execute, verify, and deliver?

If not, the danger comes quickly.

If yes, the human role may not disappear. It may move upward.

But that move does not happen automatically.

It starts by admitting one thing: AI is no longer only a smart answerer in a chat window. It is becoming an executor inside workflows.

And a lot of work sits directly inside the execution layer.

If you want to see how I am turning this issue-driven AI delivery pattern into something reusable, see marlinBian/issue-driven-ai-engineering.

It is not the full evidence for the private production change described above. It is the public-safe home for the method, templates, and control-plane experiments.