At some point this year, probably on an ordinary Tuesday, you looked at a task on your desk and wondered whether you should still be doing it by hand.
Maybe it was the monthly compliance report built from three spreadsheets. Maybe it was the fourth file noting of the day, or the OPD (Outpatient Department) roster, or the minutes of a meeting that ended an hour ago and is already fading.
You know AI could probably help. What you don't have is a way to decide. Not a vague feeling. A procedure.
That is what this post gives you. We call it the Kill It or Keep It test: five questions you ask about any recurring task, and four honest verdicts that come out the other side. It takes about three minutes per task. By the end of this post you will have run it on your own work.
Why you never got a decision procedure
Almost everything written about AI and jobs comes in two flavours.
The first is hype: AI will do everything, adapt or perish. The second is reassurance: AI can never replace human judgment, relax. Both are useless on a Tuesday afternoon, because both talk about your job. Jobs are the wrong unit of analysis.
Jobs do not get automated. Tasks do. Your job is a bundle of forty tasks, and AI's usefulness varies wildly across the bundle. The officer who dismisses AI because "it cannot understand governance" is right about six of his tasks and wrong about twenty. The manager who fears AI will replace her is afraid of something that will actually just eat her most boring Thursday hours.
So the first move is to stop asking "is my job safe" and start asking "which of my tasks should move." That question has an answer you can compute.
The five questions
Take one recurring task. Ask these five questions about it.
1. How often do you do it?
Weekly or more, score it high. Below monthly, score it low. Automation has a setup cost: learning the prompt, building the habit, checking the output until you trust it. A task you do twice a year rarely pays that back. The compliance report you build every single week pays it back forty times over.
2. Do the inputs arrive in a knowable shape?
Some tasks feed on structured material: documents, forms, precedents, data files, last month's version. Others feed on ambient context that lives in your head, in corridor conversations, in fifteen years of knowing how your department works. AI is excellent with the first kind and blind to the second. A grievance reply built from the complaint letter and your standard positions is structured. Reading the room in a tense review meeting is not.
3. Can you verify the output faster than you could produce it?
This is the most underrated question in the entire test, so slow down here. The best AI handoffs live inside an asymmetry: producing the draft takes you two hours, checking a draft takes you ten minutes. Wherever that gap is wide, AI fits. Wherever checking is as slow as doing, be careful. If you cannot judge whether the output is right without redoing the work, you have not saved time. You have added a review layer to a task you still effectively do.
4. What breaks if the output is wrong?
Be honest about the blast radius. A clumsy internal summary costs a little embarrassment. A wrong figure in an audit response, a misstated clause in a legal filing, a dosage error in a clinical document: these cost careers and worse. High stakes does not disqualify a task. It changes the verdict from "hand it over" to "hand over the effort, keep the checking."
5. Must a human own it?
Some tasks carry accountability that cannot move, no matter what the technology does. A government officer's signature on a file noting is statutory. A doctor's discharge decision is clinical responsibility. A manager's appraisal of a person is a commitment made by a person. For these, the question is never whether AI replaces you. It is how much of the surrounding effort AI can absorb while the ownership stays exactly where it is.
The four verdicts
The test's name says kill or keep, but honest scoring produces four buckets, not two.
Kill
High frequency, structured inputs, quick to verify, low stakes. Meeting minutes. Status summaries. Reformatting. First drafts of routine correspondence. Hand these over completely and stop thinking about them. This is not laziness. It is the correct allocation of a scarce resource, which is you.
Shrink
Structured and verifiable, but the stakes are real. Audit preparation. RTI (Right to Information) draft responses. Grievance replies. Budget notes. AI produces the 80 percent draft, you verify line by line and sign. The typing goes. The signature stays. For most professionals in government, healthcare and finance, this is the biggest bucket, and it is where the wage premium for AI skills actually lives.
Split
The task is really two tasks wearing one name. A board deck is (a) deciding what to argue and (b) producing slides. File noting is (a) judgment about the case and (b) drafting prose in the prescribed format. Splitting means you notice the seam, kill the (b), and keep the (a). Most resistance that sounds like "AI cannot do my job" dissolves here, because the part AI cannot do was never the part that ate your hours.
Keep
The value is the human. Negotiation. A difficult conversation with a patient's family. Building trust with a new stakeholder. Mentoring. Anything where the deliverable is the relationship. Keep these without guilt, and notice something important: the hours the other three buckets free up should flow here, because this is the work that compounds your career.
The three traps
Run the test naively and you will fall into one of three holes. Watch for them.
Trap one: killing a load-bearing task. Sometimes the memo was never the deliverable. Writing the memo was how you came to understand the problem, and the understanding is what your boss actually pays for. Before killing anything, ask: what do I learn by doing this manually, and where else will I learn it? If there is no other place, the task is load-bearing. Shrink it instead of killing it, and stay close.
Trap two: keeping out of identity. People protect the tasks they are best at, because being the person who writes the cleanest notes or builds the tightest decks is who they are at the office. If a task scores kill on all five questions and you still refuse, the reason is not in the workflow. It is in the mirror. That is a real thing to work through. It is not a scoring error.
Trap three: automating a process that should not exist. If the weekly report exists because someone asked for it in 2019 and nobody ever cancelled it, the verdict is not "kill it with AI." The verdict is kill it, full stop. Automating a pointless process makes it permanent. The test is also an audit of your workload, and some tasks fail the audit entirely.
Run it this week
Here is the exercise we give in the first week of our bootcamp, and it works on paper as well as it works in a classroom.
List your ten most recurring tasks. Real ones, from your actual week, in your actual words: "prepare the Monday review file," not "stakeholder communication." Run each through the five questions. Write a one-letter verdict next to each: K (kill), S (shrink), P (split), KP (keep).
Most professionals who do this honestly find two tasks to kill, four to shrink, two to split, and two to keep. That is roughly ten hours a week sitting in the first three buckets. The professionals earning the 56 percent AI wage premium are not smarter than you. They ran some version of this test, and then they acted on the verdicts.
The follow-up question that separates the two groups: when the hours come back, where do they go? If freed time just absorbs into more email, the test changed nothing. Decide in advance which KEEP task gets the reclaimed hours.
If you want company while you do it, the Garage Labs community is full of professionals running exactly this exercise on file notings, hospital rosters and audit files. And if you want to go from verdicts to actually building the AI workflows that execute them, Cohort 50 of our AI Accelerator Bootcamp starts August 1st: ten weeks, live and mentor-led, no tech background needed, twenty seats.
Kill the typing. Keep the judgment. That is the whole trick.