AI AgentsAI WorkflowAutomationLearn

Should AI Do This or Should I?

Should AI do this, or should you? A five-question checklist, a build / buy / delegate / keep scorecard, and a five-task exercise.

Oct 28, 2026

Five questions and four lanes — build, buy, delegate, keep human — for deciding what AI should touch. Run the checklist, score the lanes, then take the five-task exercise through your own week.

Start here, verbatim from the stream: recording a 30 minute video, transcribing it, and then taking those ideas and building them into something useful beats having AI create content from a short prompt. The machine multiplies what you feed it — so feed it the good stuff, and decide deliberately what it gets to touch.

The 5 questions

Q1 — Think or remember? (Shukla's test)

Does the task need someone to think, or just remember? Repeating the same steps — reminders, moving a lead into a sheet, notifications — is automation: simpler, cheaper, no surprises. Understanding — drafting, judging, personalizing — is where AI belongs.

Q2 — Is it irreversible?

Messages sent in your name, relationships touched, things published: a human decides, always. My Friday rule, verbatim: I'm gonna go try to do five or 10 of these things, see exactly what happened, fine tune it, work out the kinks, then start adding it. Small supervised cohort first; scale only after trust.

Q3 — Frequency vs upkeep

Daily, stable, boring → automate or buy. Rare, fiddly, identity-shaped (like Ghost Studio) → think hard before building, because you will support it forever: if you build these super apps that support yourself, they can be hyper powerful — at the same time, you have to support them.

Q4 — What does failure cost?

A wrong connection request is cheap and fixable. A wrong message to a first-degree contact is not. Low-risk: let AI run. High-risk: a review queue where you approve before anything sends.

Q5 — Joy and judgment

If doing it yourself keeps your taste sharp — writing, editing, talking to customers — keep it. The premium skill of the next years is quality control of AI output, and you can't control what you've stopped practicing.

Scorecard — try it now

How scoring works: read each lane’s test, then score one real task below. Nothing is saved — your notes are the app.

Build when: core to how you work + no mature tool + worth supporting. Test: would you still fix it at 8am when it breaks? Yes → build.

Quick check — name one task and pick its lane:

The 5-task exercise

List five things you did this week. Run each through Q1–Q5, assign a lane, and publish one result: what moved, and what you took back. That loop — review, learn, hand more off once trusted — is the whole game: it's not set up and forget it. It requires constant monitoring and human intelligence.