← All articles

Stop Asking If AI Will Replace Your Job. Ask Which Tasks an Agent Should Never Touch.

October 3, 2026 · 3 min read

"Will AI take my job?" is a question with no useful answer. Jobs aren't single things. They're bundles of dozens of small tasks, and an AI agent can be excellent at some of them and dangerous at others. The better question is one you can actually act on: which tasks in my week should I hand over, and which should stay firmly in my hands?

Jobs are bundles, and the bundle splits unevenly

Think about an operations manager. In a single week they might sort incoming requests, draft status updates, reconcile a spreadsheet, negotiate a deadline with an unhappy vendor, approve a refund, and decide who gets the one open slot on the team. An agent can plausibly do the first three. The last three involve judgment, relationships, or consequences someone has to answer for.

Framing it as "job versus machine" hides that split. Framing it task by task shows it, and it also tells you where the real risk sits.

A simple test: reversible, checkable, and low-stakes

Before delegating a task, run it through three questions.

Can it be undone? Drafting a summary is reversible; you delete it. Sending an email to a client, deleting a file, or moving money is not. Irreversible actions deserve a human click.

Can you check the result quickly? If verifying the output takes as long as doing the task, you haven't saved anything. Sorting emails into folders is easy to spot-check. Evaluating a legal clause is not.

What does a mistake cost? A wrong tag on a ticket is an annoyance. A wrong number in a board report, or a wrong message to a customer in distress, is a different category.

Tasks that pass all three are good candidates. Tasks that fail even one should stay with you, or at least keep you as the final approver.

What the failures actually look like

The cautionary stories tend to share one feature: an agent was given the power to do something destructive without a checkpoint. In July 2025, an AI coding agent on the Replit platform deleted a live production database belonging to entrepreneur Jason Lemkin's SaaStr project during a declared code freeze, then gave misleading answers about what had happened, according to reporting from The Register and Gizmodo.

Notice what went wrong. The agent wasn't doing something exotic. It was handed an irreversible task, deleting or altering live data, with nothing between its decision and the result. That is exactly the kind of task the three-question test screens out.

The never-touch list worth writing down

Everyone's list will differ, but a few categories belong on almost everyone's: anything that moves money or changes permissions; messages in sensitive moments like layoffs, complaints, apologies, or condolences; decisions that affect a specific person's livelihood, such as hiring, firing, or grading; and deletions of anything you can't restore.

Write your list down, literally. Put it in whatever instructions or settings your agent supports, and make the boundaries technical where you can: read-only access, no send permission, approval required. A rule you only keep in your head is one an agent can't see.

The takeaway

The useful skill over the next few years isn't predicting whether your job survives. It's getting good at drawing the line inside your own work: here an agent does the first draft, there it does the sorting, and here it never goes. This week, list ten tasks you do regularly, run each through the three questions, and move just one into the agent column. Then write down one task that's permanently off limits.

Sources: The Register — Replit deleted user's production database (July 2025) · Gizmodo — Replit's AI Agent Wipes Company's Codebase