At 6:30, AI Started a Shift for Me

Every morning at 6:30, a scheduled workflow starts working for me.

It scans the 30 most recent items I have saved, removes anything I have already used, and compares the remaining ideas against my last 60 articles with vector similarity. If the score crosses the threshold, it stops. Only then does it move on to checking sources, drafting, making images, and translating.

I originally built it to save myself some repetitive work. Watching it take over the process one step at a time changed how I think about AI and jobs. What loosens first may not be an entire occupation such as “programmer” or “writer.” It may be the tasks inside that occupation that can be specified, checked, and delivered through a screen.

AI can compress a source into a summary. It can sort 30 candidates according to fixed rules. Given a title, tone, boundaries, and sources, it can even produce a complete first draft.

I used to click through pages and check these things by hand. Now, when the inputs and acceptance criteria are clear enough, a machine can run through them in one pass.

AI tends to replace individual tasks before it replaces an entire occupation.

The International Labour Organization's 2025 study of occupational exposure to generative AI examined nearly 30,000 tasks. It estimated that about one quarter of workers worldwide are in occupations with some degree of exposure.

Its main conclusion, however, was that jobs are more likely to be transformed than entire occupations are to disappear.

The reason is not mysterious. Every job contains some tasks that can be standardized, and others that require missing context, exception handling, coordination, or a person who is ultimately accountable.

That is also why “move into offline work and you will be safe” sounds more convincing than it really is.

Whether AI can reshape a job does not depend only on whether the worker sits at a desk or visits a physical site. The more important question is whether the work can be divided into clean units: what goes in, what must come out, and how an error is detected.

If the job can be explained in three sentences, the cost difference between outsourcing it to a person and delegating it to a model will eventually be calculated.

The price of a piece of copy, an image, or a block of code used to be tied, at least loosely, to production time. Buyers could not see how many attempts or revisions were involved, so “this will take several days” still acted as a rough cost estimate.

Now there is a tool on the other side of the screen that can produce a first draft in minutes. Even when that draft is unreliable, it pulls the price anchor downward. Saying that the work was made by hand does not automatically move the quote back up.

The remaining difference is whether the deliverable actually works in its specific setting. Someone still has to notice two contradictory requirements, uncover constraints the client never stated, and redo the work when the result fails. Those are the costs outside the first draft.

It is also easy to label everything left over as “judgment” and use the word to comfort ourselves. Judgment becomes valuable only when it leaves an auditable result: which flawed requirement was changed, which mistaken release was stopped, which failure condition was caught early, and whether the process can be reconstructed when something goes wrong.

If all someone can say is “I understand the business better,” but they cannot show any of those traces, the claim is not much more valuable than a polished self-assessment generated by a model.

Anthropic's 2026 research, based on real product usage, is similarly restrained. It did not find a systematic rise in unemployment among highly exposed occupations. It did find a signal worth watching: hiring of younger workers in those occupations may already be slowing.

That is not evidence of a “job apocalypse.” It looks more like an early change in price. A company may not eliminate an entire role first. It may simply hire one fewer junior employee and hand the practice tasks once given to that employee to AI.

As standardized deliverables get cheaper, the first thing to disappear may be the rung people used to build bargaining power.

Back to my 6:30 workflow. It can read, sort, write, and stop when a rule is triggered.

But I still have to decide why those rules exist, whether a saved idea is merely an emotion or a judgment fit for publication, whether the evidence is strong enough, and whether a flawed draft should be withdrawn.

There is nothing mystical or heroic about that work. It is simply difficult to compress into one reusable checklist.

So I will not turn off the AI because it can work for me. I will keep handing over the steps I can describe clearly.

What makes me uneasy is a different question. Once the machine has completed every item on the checklist, will I have anything left to contribute besides another checklist? Can I still explain why the work matters, what “done correctly” means, and who owns the outcome when it goes wrong?