Seeing Work as Tasks, Not Jobs: How to Carve Out What You Hand to AI

Ask whether a task can be handed off, not whether a job will be replaced

Looking at work by job title makes it hard to judge what AI can take on. A manager's job, a designer's job, and an office job are each bundles of many tasks. Splitting them into tasks lets you find the parts that can be handed over.

The view comes from a 2003 paper by economists David Autor, Frank Levy, and Richard Murnane. It made the task, not the occupation, the unit of analysis, and asked what computers do in place of people and what they leave to them.

The classification in the paper

The paper sorts tasks along two lines:

• Routine versus non-routine
• Cognitive versus manual

Routine tasks are ones you can write down as explicit rules. The paper argued that computers readily substitute for them, while non-routine tasks, such as problem solving that resists rules or interaction with people, tend to be complemented by computers. Using data, the authors showed that the share of routine tasks fell.

Estimates of automation's effect differ depending on whether you count by occupation or by task. A 2016 working paper by OECD researchers, for instance, reported that when tasks within jobs are considered, the share of workers at high risk of replacement looks smaller than in occupation-level estimates. Treat any such figure as a rough guide.

The 2003 paper was not about modern AI. It analyzed the computers of its time. The task view still applies, but the line between routine and non-routine moves with technology, so do not treat it as fixed.

Carving out tasks to hand over

Start by listing a week of your work down to units of ten to twenty minutes. Not "sales," but "summarize an inquiry," "draft a reply," "schedule a visit."

Then ask these questions about each task:

• Can you describe the input and the output in words?
• Can you write down what a good result looks like?
• If it comes out wrong, can a person catch and fix it?
• Does it rest on trust with another person or a judgment you are accountable for?

If the first three are yes and the last is no, the task is a candidate. Summaries of inquiries and drafts of routine replies are easy to hand off on the condition that you review them before use. Price negotiation and decisions about how to apologize carry responsibility, so you keep them.

Handing it over well

For each task, write the input, the form of the output, the standard it should meet, and anything it must not do. Pairing one instruction with one task makes checking easier. Rather than passing a large job in one piece, cut it small and review each output before connecting the next step. Then you can see where a mistake came from.

What an inventory of tasks tends to miss

When you list tasks, thinking time, waiting time, and judgment are easy to leave out. Before "write a reply," for example, you review the message and recall the person's situation. If you hand over only the writing, the AI lacks the information it needs, and the output falls short.

Before handing off a task, write down the information you use in your head to do it: past exchanges, pricing rules, preferences in wording. That becomes the material for your instruction.

Reviewing after you hand it off

Carving out a task is not a one-time job. If checking the output takes as long as doing the work yourself, handing it off has little value. Each month, look at the time spent checking and the amount of correction, then decide whether to improve the instruction or take the task back.

Work also changes. When new kinds of requests arrive or your business shifts direction, rebuild the task list.

The smaller the unit of work, the quicker it is to judge whether it can be handed off. When in doubt, ask whether you could give a new hire this task with only this explanation. If you could, there is a chance you can give it to an AI tool. If the explanation runs too long, either the task is still too big or your standards have not yet been put into words.

What to do next

1. List about twenty tasks from one week of work, each ten to twenty minutes long.
2. For each, write the input, output, and standard of a good result in one line.
3. Set aside tasks that involve accountability or trust as ones you keep.
4. From what remains, pick one task that is easy to check.
5. Write the instruction for it and try it for a week, reviewing every output yourself.