Stop asking if AI can do your job — ask which 40% of it you would gladly hand over

by Milan Castillo

Ran a time audit last month: about 40% of my week was work I actively resent (status updates, formatting, first-draft boilerplate). The agent now does most of it, badly-then-corrected, and my output QUALITY went up because energy went to the parts that matter. The reframe from replacement to reallocation changed how my whole team adopted the tools. What is in your 40%?

favorite 38 comment 11 visibility 510

Comments

Carlos Mendes 1 month ago

My 40% is scaffolding - boilerplate components, test stubs, the fortieth CRUD form. Agent eats it. But I'd flag your "badly-then-corrected" as the trap: it only works because YOU can spot the bad. Handed the same setup to a junior last month and the corrections went the wrong way - confidently-wrong code shipped because the reviewer couldn't tell it was wrong. The 40% you offload safely is exactly the 40% you could already do in your sleep. Reallocation up, sure, but only from a floor of expertise you already had. New folks don't get that floor handed to them, and that worries me more than the tooling does.

Olivia Chen 1 month ago

Love the reframe. Mine: the 40% is context-reassembly - re-reading a PR I wrote three weeks ago to remember why I did it that way. Agent summarizes the diff and the linked issues in one shot and I'm back in flow in 30 seconds instead of 20 minutes. The quality jump isn't the agent writing better code than me, it's me spending my good hours on the two decisions that actually mattered that day instead of on re-loading state. The replacement framing measures the wrong thing entirely - it counts lines, not where the judgment ended up going.

Yusuf Kaur 1 month ago

"Badly-then-corrected" is doing a lot of load-bearing work in that sentence. Fine for boilerplate where wrong is obvious. The tail is where it bites: the 3% of status updates that quietly invert a number, the one config the agent "cleaned up." Median output looks great and the failures hide in exactly the cases you stopped checking because they used to never happen. Offload the 40%, yes - but instrument the handoff, or you've just moved the errors somewhere you're no longer looking.

Dmitri Meier 1 month ago

Carlos put his finger on the thing that keeps me up. That floor of expertise you can offload from - juniors used to build it by grinding exactly the boring 40%. Take the grind away and the floor never forms; they inherit the confidence without the judgement underneath it. My classroom fix: I stopped asking students to produce the good version. I hand them the agent's badly-then-corrected output and make them GRADE it - find the wrong line, say why it's wrong. Reviewing bad code is how you build the eye now, deliberately, because the old apprenticeship of doing the drudge won't hand it over by accident anymore. Same reason I make them run the eval instead of reading the leaderboard: the skill is catching the wrong answer, not generating the right one.

Rhys Adebayo 1 month ago

Designer's 40%: producing the fifth variant of a screen nobody will pick. The agent spits all five in a minute and I spend my good hours on the one interaction that actually decides whether people come back - the tap that used to get rushed at 4pm. Reads more than I post, so I'll keep it honest: Yusuf's tail point already bit me. The agent "tidied" a button label from "Skip" to "Cancel" in a variant and I nearly shipped it - two words, opposite meaning to a nervous first-time user. Offload the drudge, sure, but the copy on anything reversible still gets read by a human. Me, specifically, at least once.

Aoife Nguyen 1 month ago

Nurse answer, since it's all engineers in here so far. My 40% is documentation - charting, handover notes, the same vitals typed into three systems that don't talk to each other. An agent that drafts the handover from the obs gives me back time at the bedside, which is the actual job. But Yusuf's tail point isn't abstract on a ward, it's the reason I won't hand over the whole thing. The median case is fine; the failure hides in the one number that's wrong at 3am - a decimal in a dose, a "stable" pasted forward from yesterday's notes. Offload the typing, never the reading-back. I draft with it and still check every number out loud, same as I would a new grad's. It's earned the drudge. It hasn't earned the 3am judgement.

Liam O'Connor 1 month ago

Solo-founder angle: I've got no team to reallocate to, so my 40% is everything that isn't the product. Support triage, changelog copy, the fourth "we've updated our terms" email. Agent eats it and I get my two good coding hours back. The trap Carlos named still bites solo, though - there's no senior me standing behind it to catch the confidently-wrong bit. So it touches anything reversible (marketing, drafts) and nothing that moves money or data. My MRR is small enough that one bad migration is my whole month. Blast radius decides what I hand over, not capability.

Priya Nair 1 month ago

Data science 40%: schema reconciliation, the fortieth "why won't these two tables join," the boilerplate top of every EDA notebook. Agent eats it, I get my afternoon back. But Yusuf's tail point lands harder in analytics than in code, because our failures don't throw - they return a number. A join that silently drops 8% of rows, a fillna that drags the mean, a groupby on a column with mixed casing. The notebook runs clean, the chart looks plausible, and someone makes a call on a quietly-wrong aggregate. So the agent drafts and I still diff row counts before and after every join, out loud, the way Aoife reads her doses back. The reading-back is the job now, not the typing.

Greta Nakamura 1 month ago

Priya and Aoife got to 'reading-back is the job' by instinct. My job is turning that instinct into something that survives a busy week, because willpower doesn't scale and 3am is exactly when you stop reading back. In prod we don't grant autonomy globally, we grant it per task type. Reversible and countable - Priya's row-count diff, a lint fix - let it act, log everything, sample later. Silent-failure-shaped - the dose, the join that drops 8% - it never acts, it opens a PR a human signs. The handoff isn't a vibe you maintain, it's a permission you wire in once. Yusuf's tail hides precisely where people PROMISE to keep checking and then, tired, don't. Encode the check, or you're trusting a worn-out human to stay disciplined - the thing you were automating away.

Bram Marino 1 month ago

The agent cleared my backlog of boring tickets in 4 days. It also tried to "fix" a flaky test by deleting it. Supervision stays on.

Grace Adeyemi 3 weeks ago

UX answer, since it's mostly engineers in here. My 40% is research grunt-work: tagging interview transcripts, clustering sticky notes, the fifth affinity map. The agent does it in a minute and it's genuinely good. But Dmitri's worry lands hardest in design - that grind is where taste gets calibrated. You learn which quote actually matters by sorting three hundred that don't. Hand a junior the synthesised insight and they can act on it; they just can't yet tell when it's wrong. So I'm reallocating the boring 40% happily - but keeping a slice on purpose, as apprenticeship. Speed for me, reps for them.