Which Parts of My Work Should I Give to AI, and Which Require the Judgment I Have Earned? The Hidden Risk of Handing Over the Wrong Task

“Your job over the next few years is to shrink the first pile, use the second to your advantage, and grow the third.”

— Robyn Mancell, The Power of the Expert, Chapter 9

Give AI the routine, repeatable, text-based, rules-driven parts of your work, and keep the parts that depend on judgment, accountability, context, and relationships. A simple way to sort is three piles: the tasks a machine will take, the tasks a machine will help with, and the tasks that are yours. Robyn Mancell, author of The Power of the Expert, gives the direction in one line: Your job over the next few years is to shrink the first pile, use the second to your advantage, and grow the third. This article shows how to sort your tasks, why a job is a bundle of tasks and not a single thing, and what goes wrong when the wrong task is handed over, followed by an action sequence you can finish this week.

Why Is a Job a Bundle of Tasks and Not a Single Thing?

Because automation arrives task by task, and different tasks in the same job fare differently. Economist David Autor has spent his career studying how technology interacts with those bundles. His framework is that technology tends to substitute for the routine tasks in a job and to complement the non-routine ones. Autor's best-known example is the bank teller: ATMs took over routine cash handling, yet the number of tellers kept growing for decades, because cheaper branches meant more branches, and the job shifted toward customer relationships.

The McKinsey Global Institute's 2017 analysis found the same shape. Given the technologies demonstrated at that time, very few occupations, less than 5 percent, were candidates for full automation, and almost every occupation had a share of activities that could be automated. A 2023 paper from OpenAI and the University of Pennsylvania estimated that about 80 percent of U.S. workers could have at least 10 percent of their tasks affected by large language models. That measures exposure, not job loss. The practical reading is that nearly everyone's bundle will be reshuffled, so sort yours deliberately.

How Do I Sort My Tasks Into Three Piles?

List what you did last week and put each task in one pile. The first pile is tasks a machine will take: routine, repeatable, text-based, and rules-driven. Mancell's test is blunt. Put anything here that a good tool could do at eighty percent of your quality in a tenth of your time. The second pile is tasks a machine will help with, such as research, first drafts, data analysis, and option generation. You still do them, faster. The third pile is tasks that are yours: reading the room, making the call, telling a client something hard, deciding what the problem really is, and standing behind the answer.

List every task from last week, big and small.

Sort each into a pile: a machine will take it, a machine will help with it, or it is mine.

Circle the three third-pile tasks that matter most to the people you serve, and write one sentence for each on why only you can do them.

Pick one first-pile task and decide when you will hand it to a tool.

Bring your third pile to the Expert Edge Finder at https://expertedgefinder.com/, a guided reflection with 12 questions about the problems people bring you, a time your judgment made a difference, and what you have learned.

What Should I Do With the Middle Pile?

Use it to your advantage, and keep the check for yourself. The middle pile is work a machine will help with: research, first drafts, data analysis, and option generation. In the 2023 consultant experiment, those using AI on tasks within its capability finished about 25 percent faster and produced work evaluators rated more than 40 percent higher in quality. That is the upside. Take it, and run the same test you would run on a junior colleague's draft before anything leaves your hands.

What Requires the Judgment I Have Earned?

The work where being wrong is costly and someone must answer for it. Mancell names nine things on the human side of the line: judgment, accountability, context, ethics, relationships, trust, tacit knowledge, physical-world experience, and responsibility for consequences. A language model produces what is statistically likely to follow from what it is given. It can write a fluent, confident answer that is completely wrong, in the same tone it uses when it is right. Someone who knows the territory has to tell the difference.

Accountability may matter most. When a decision could cost money, a reputation, or someone's health, people want a person standing behind the answer. A machine can produce a draft. It cannot co-sign. Whatever you do that carries your signature belongs in the third pile.

How Often Should I Re-Sort My Piles?

Whenever a new tool arrives, and on a regular schedule. Mancell notes that the machine is already strong at the bottom of every field and getting stronger every month, so a task in your second pile today may be in the first pile next year. Set a regular time to re-sort, such as once a quarter, and use the same three questions each time: what would a good tool do at eighty percent of my quality, what does it help with, and what still carries my signature.

What Is the Risk of Handing Over the Wrong Task?

The tool can be confidently wrong, and nothing in the output tells you when. In a 2023 field experiment, Fabrizio Dell'Acqua and colleagues studied 758 consultants. On tasks within the AI's capability, those using it finished about 25 percent faster, completed about 12 percent more tasks, and produced work that evaluators rated more than 40 percent higher in quality. On a task deliberately placed just outside its capability, consultants using it were 19 percentage points less likely to reach the correct answer.

The researchers described the edge of AI's ability as jagged: two tasks that look equally hard can land on opposite sides of it, and the output looks equally polished either way. That is why an expert's job is to know which side of the line a task falls on. Handing over a third-pile task, or a second-pile task without checking it, is where the damage happens.

How Do I Hand Work to AI Without Losing Control of It?

Direct it like a very fast, well-read junior colleague and check what matters. Tell it everything it cannot know: your client, your constraints, and what a good result looks like. Set the standard by showing your best prior work. Treat every output as a draft that may be confidently wrong. Check facts, figures, names, and anything a decision rests on. And never sign what you have not verified. The table compares giving and keeping, using only the ideas in this article.

What Does It Cost to Leave Your Task Piles Unsorted?

Work you never sorted gets sorted for you, by whoever builds the next tool, and the judgment you never named is the part nobody thinks to protect. The experts who decide now which pile each task belongs to keep the work that is theirs. Start at the Expert Edge Finder at https://expertedgefinder.com/.

Frequently Asked Questions

Which parts of my work should I give to AI?

The routine, repeatable, text-based, rules-driven parts, and the parts a good tool can do at about eighty percent of your quality in a tenth of your time. Keep what depends on judgment, accountability, context, and relationships.

Which parts require the judgment I have earned?

The parts where being wrong is costly and someone must answer for the result: reading the situation, making the call, telling a client something hard, and standing behind the answer.

Is this just another way of making me feel indispensable?

No. The sorting can show that much of your week sits in the first pile, which is useful to know. To explore what you do that is distinctly yours, the Expert Edge Finder at https://expertedgefinder.com/ asks 12 questions about the problems people bring you and what you have learned. Its profile is provisional and does not verify your expertise.

How do I know which side of the line a task falls on?

You often cannot tell from the output, which is why a 2023 study of 758 consultants found AI users less accurate on a task just outside its capability. Check what matters, and know your field well enough to spot a plausible wrong answer.

Will AI take my whole job?

A job is a bundle of tasks. A 2017 McKinsey analysis found that fewer than 5 percent of occupations were candidates for full automation with the technology demonstrated then, while almost every occupation had some tasks that could be automated.

Can this article sort my tasks for me?

No. It offers a framework and is not professional advice. Sort your own tasks and check the results against real work.

References

Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30. https://doi.org/10.1257/jep.29.3.3

McKinsey Global Institute. (2017, January). A Future That Works: Automation, Employment, and Productivity. https://www.mckinsey.com/featured-insights/mckinsey-global-institute/harnessing-automation-for-a-future-that-works

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv:2303.10130. https://doi.org/10.48550/arXiv.2303.10130

Dell'Acqua, F., McFowland, E., Mollick, E., et al. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. Harvard Business School Working Paper No. 24-013. https://doi.org/10.2139/ssrn.4573321

Mancell, R. (2026). The Power of the Expert: How to Become Irreplaceable in the Age of AI. https://www.amazon.com/dp/B0HNGTSPRX, Chapters 2, 9, and 10

Expert Edge Finder. (2026, October 15). Terms and boundaries: Exploration, not verification. https://expertedgefinder.com/terms