Software has been automating the wrong half for fifty years

Cron is fifty years old

Scheduled execution has been a solved problem since the mid-seventies. If the difficulty of recurring work were about remembering to start it, we would have finished with this a long time ago. And yet a very large number of people still assemble the same report by hand every Friday, using software that could have triggered itself since before most of them were born. That gap is worth explaining rather than shrugging at.

What each wave actually solved

Look at the sequence and a pattern shows up quickly. Every generation of tooling picked off one real dimension of the problem and shipped it, and the thing left over was the same thing every time.

• Schedulers answered when. Genuinely useful, and it turned out that timing was rarely the reason a job was slow.
• Workflow builders answered in what order, which works beautifully as long as every input takes the same path through the graph.
• Robotic process automation answered how to click things, which mattered when the systems had no APIs and matters less now that they do.
• Chat interfaces answered how to ask, which lowered the barrier enormously and left you responsible for starting every conversation.

The leftover was always the same

In every case, what did not get automated was the decision about what matters on this particular run. Which of this week's twenty developments belong in the update. Whether this exception is the boring kind or the kind someone needs to hear about today. Whether the thing that looks minor is actually the thing. That call has to be made fresh each time, on inputs that never repeat, and none of the four waves above could make it.

The tell is the last step of the workflow

Here is the diagnostic. Go and look at any carefully built automation inside a company and find its final step. A remarkable number of them end in something called human review, or approval, or sanity check. That step is not quality assurance on the automation. That step is where the actual work happens, and the automation is the part that fetched the materials. Nobody planned it that way. It is just where the machine ran out of things it could do.

What changed, precisely

Not that the work got simpler. Two capabilities had to arrive together, and they did not arrive at the same time. The first is being able to look at a pile of unstructured input and make a defensible call about what deserves attention. The second is being able to start on your own, keep context between runs, and put the result where the work lives. Judgment without initiative is a better chatbot, and you still have to open it. Initiative without judgment is cron with a language model bolted on, forwarding everything it finds and leaving you to sort it.

Where to start, given all that

The natural first candidates are the recurring jobs where a call has to be made every time and the stakes of any single run are contained. Weekly summaries, triage, monitoring, standing research. These are exactly the jobs that resisted the previous fifty years of tooling, and they are also the ones where handing over the judgment pays back immediately, because the alternative is you making the same call every week forever.