I Semi-Automated the Weekly Busywork I Always Do

The little things you repeat every week — cleaning up meeting notes, summarizing Slack, drafting emails, wrangling CSVs. Each one's no big deal, but together they quietly eat your time. This is exactly where handing work to an AI pays off.

One important caveat, though: "just automate all of it for me" almost always fails. What actually works is the "break it down, human checks the end" semi-automated shape. Here are the patterns I actually run, along with the tricks that make them work.

Example 1: Turn a weekly report into a pile of sub-tasks

Every week I pull a bunch of numbers together into a report. My first instinct was to one-shot it — "look at all of last week's data and write it up nicely." It flopped, hard. Things got missed, the numbers looked off, the format changed every time.

What worked was splitting one giant task into a collection of small sub-tasks. One "agent" owns section A, another owns section B, each with a narrow, specific instruction. Then I pour the results into a fixed template to assemble the single report.

It's the same as staffing a human team — "you just do this" beats "do everything," for people and for AI. The accuracy goes up, and you can verify each sub-task on its own. If one piece is off, you just regenerate that block.

Example 2: Summarizing Slack and long threads

That huge discussion thread, or the channel you couldn't keep up with. Just having it summarized into "what got decided in 3 lines + who's now on the hook for what (TODOs)" makes getting back up to speed wildly faster.

The trick is to fix the output shape. Lock a frame — "①conclusion ②decisions ③open questions ④action items (with owners)" — and make it pour into that every time. Let it summarize freely and the granularity shifts on every run, which makes it useless. Keep the shape consistent and you can drop it straight into your notes or task tracker.

Example 3: Draft the email — but you hit send

For support replies and emails, AI drafts, human sends is the sweet spot.

For the routine stuff, have it produce a reply that matches the tone of past threads. Now you're editing instead of writing from scratch, and the felt effort drops off a cliff. The key: never hand "send" to the AI. Anything touching money or customers especially — the last inch is always yours. Roughly 80% draft, 20% finish is the split least likely to blow up.

Example 4: Wrangling messy CSVs into shape

That CSV someone exported with columns all over the place — reshaping it to a fixed schema, or pulling just the rows you need and classifying them. Dull, but a classic time sink.

AI is great at this too. But it'll happily drop rows when the volume is high, so never take "did all of it" at face value. Reconcile the counts — N in, N out. Eyeball a few sample rows yourself. If the numbers don't match, suspect that some got silently skipped. Same theme: semi-automated, human does the final check.

The tricks, summed up

After a lot of trial and error, what works boils down to four things:

1. Break tasks down small — "just do this" beats "do everything." Sub-tasks are easier to verify.
2. Fix the output format — lock a frame and pour into it. Same shape every time means you can use it as-is.
3. Keep the human on the final review — semi-automated, not fully. Don't hand over the last push: send, confirm, commit.
4. Templatize and repeat — save an instruction that worked and reuse it. Don't reinvent it from scratch every week.

It's less "AI takes your job" and more hand the boring prep to the AI so you can focus on the judgment. Automating busywork is really just that division of labor. Pick the one thing you do every single week, and start by breaking it down small.