The Intern That Never Sleeps
Some of the best AI I built is the kind nobody ever opens.
The best AI I built one quarter is the one nobody has ever opened. It arrives in the team channel every Monday with the competitive news already gathered, sorted, and summarised, then goes back to sleep until the following week.
Most AI you meet is a chatbot: you go to it, you ask, it answers. That is the pull model, and it is perfect when you have a question. But much of the work that actually drains a team's time is not a question anyone enjoys asking. It is a chore that recurs forever, like keeping half an eye on what competitors are doing across a few product lines. Nobody wants to do that by hand every week, so, mostly, nobody does it well.
That is a job for a different shape of AI. Not something you talk to. Something that turns up.
Push beats pull for the boring things
So I set up scheduled jobs that run themselves. On a cadence, each one wakes, gathers the relevant news for its product line, summarises it, formats it, and posts it straight into the channel where the team already lives. No login, no prompt, no "I really must check that this week." The work appears where you already are, finished.
The shift is small and it matters. A chatbot is a tool you have to remember to use, which on a busy week means you will not. A scheduled agent removes the remembering. The output is not necessarily better than what you would get by asking; it is simply that it happens, every time, which over a quarter dwarfs quality on any single week.
What is actually under it
The anatomy is unremarkable, which is rather the point. A trigger, a schedule. A gather step, pull the sources. A summarise step, the model's real job. A format step, make it readable at a glance. A deliver step, post it where people will see it. Each piece is mundane. Chain them together and leave them running and you have quietly retired a recurring task no human was doing reliably anyway.
The valuable insight here is not technical. It is noticing which of your recurring chores are mechanical enough to hand off, and being willing to spend an afternoon wiring up something you will then happily ignore for months.
The part that should keep you slightly suspicious
Here is the catch with unattended automation: it delivers a wrong answer exactly as punctually as a right one.
A person doing this weekly would notice if a source went strange, or an announcement got mangled, or the summary stated something that never happened. An automated job notices nothing. It will post a clean, well-formatted, entirely wrong digest at the appointed minute and feel no remorse, and because it looks like every other week's, people will believe it. The failure mode of automation is not that it stops. It is that it keeps going long after it should have stopped.
“The failure mode of automation is not that it stops. It is that it keeps going long after it should have stopped.” WHY SET-AND-FORGET NEEDS WATCHING
So you build in a little defence. Be fussy about sources, since the whole thing inherits their reliability. Make the output show where each item came from, so a reader can check rather than simply trust. And, above all, only automate the things where a bad week is survivable.
A rough rule for what to hand over
The test I use is two questions. How high are the stakes if it is wrong, and how easily would someone catch it? Low stakes and easy to catch, automate it and move on; a weekly news digest is precisely this. High stakes or hard to catch, keep a human in the loop, because the cost of a confident error outweighs the time you would save.
Automate the boring and reversible. Watch the consequential. Most of the skill is telling which is which, and the rest is the small discipline of trusting a machine to do a job while staying faintly suspicious of it. My weekly intern has never once been late. I still read it as though it might be lying, which is, I think, the correct relationship to have with an intern. ▮