Choosing Which Tasks to Hand to AI and Which to Keep

By Trevor L Taylor

Hand AI the work where a wrong answer is cheap and easy to catch, and keep the work where a mistake would be expensive or hard to see. The common advice says to hand over whatever takes the most hours or whatever you like least. That rule sorts tasks by how they feel to you and ignores what a mistake would cost.

What should you hand to AI first?

Start with work where a wrong answer costs little and takes seconds to spot, such as a first draft of a status note, a reformatted schedule or a list of questions for a call. Keep anything where an error could travel a long way before anyone notices it.

Picture a project coordinator on a Tuesday morning with six jobs open in six windows. One is a status note for the team, one is a supplier price comparison, one is an apology to a client for a late delivery, one is a reformatted schedule, one is a summary of a long thread, and one is a choice between two vendors for the next order.

The usual rule says to give AI the three that will take longest. That sends the price comparison and the thread summary to the tool, which may be right, and it also sends the apology, which probably is not. Hours spent describe the clock, and they say nothing about which of the six can be wrong without somebody paying for it.

The advice has a reason behind it. A task that eats hours feels like the obvious thing to give away, and the relief is real when the tool finishes it. The trouble is that the relief arrives before the checking does, and the checking is where the cost of a mistake shows up.

The cost of a wrong answer sets the first sort

Two questions do most of the sorting. The first is what a wrong answer would cost if it got through. The second is how quickly you can check an answer before it leaves your hands.

A task that is cheap when wrong and quick to check goes to AI, and the status note and the schedule belong there. A task that is costly when wrong but quick to check can still be drafted by AI, as long as you verify every figure against the source. A task that is cheap but slow to check needs spot checks, because small errors pile up where nobody looks.

Take the summary of a long thread. If it drops the line where a colleague moved a deadline, the error looks small on the page and large three days later. A price comparison that feeds a purchase order is costly when wrong, yet each figure can be set against a quote, so AI can build the table while you verify the totals.

A wrong answer you can spot in ten seconds costs you ten seconds.

A task that is costly and hard to check stays with you, and the vendor choice is that kind of task. The tool can list considerations, but you carry the consequence, and you cannot verify a judgment by glancing at it.

In a Harvard Business School field experiment with 758 knowledge workers, people using AI on a task outside its capabilities were 19% less likely to produce a correct solution than people working without it. On tasks inside its capabilities, the same group finished 12.2% more tasks and finished them 25.1% more quickly. The researchers call the uneven edge a jagged frontier, which is why you cannot judge a task by how hard it feels.

How do you know a task is easy to check?

A task is easy to check when you can confirm the result against something you already hold, such as a source document, a total or a calendar, in less time than the work itself would take. When checking means redoing the work, the task was never cheap to hand over.

Three properties make a result quicker to check. A short answer takes less time than a long one, an answer with a visible source behind it takes less time than one built from the general knowledge of the tool, and a result with a known shape, such as a table whose rows must add up, takes less time than free text. When a task has none of these properties, assume that checking will take as long as the work.

The price comparison passes this test, because the coordinator can set each figure beside the supplier quotes in a couple of minutes. The thread summary passes only if the coordinator has read the thread, and it fails for anyone who has not. Checking depends on what you know, so the same task can be easy for one person and hard for another.

A framework from the MIT Center for Information Systems Research, built on interviews with 30 executives, treats routine decisions with low ambiguity and low risk as strong candidates for automation. That is the same instinct as sorting by cost and checkability, applied to a whole company instead of one Tuesday morning.

Here the usual advice earns some credit. Long, repetitive tasks are often good first candidates, but only because they tend to be checkable, and the hours they take are a side effect of that. The test that decides is still the cost of an error.

Where the sorting stops working

The line moves. A task that was costly and hard to check last spring can become cheap and quick to check as tools improve, and a task that looked safe can turn risky when the tool changes or the stakes rise. Reread your list every few months and move tasks across it on purpose.

The sort also leans on your own judgment, so it fails when you stop checking. Fluent answers invite relaxed checking, and a checkable task stops being checkable the day nobody checks it.

The sort also says nothing about whether a task is worth doing at all. A cheap and checkable task that nobody needs finished is still waste, and handing it to AI only makes the waste arrive faster.

A team can apply the same sort together, and the piece on teaching a team to use AI without losing its judgment covers how a group agrees which decisions stay with a person. The personal version starts smaller, with one written list and two questions.

Take the tasks of one ordinary week and give each one a mark for cost and a mark for checking time. Hand over the cheap and checkable, draft the costly and checkable with care, spot check the cheap and slow, and keep the rest. A task belongs with the tool only when a wrong answer is cheap and checking it is faster than doing it.

Trevor L Taylor is an investor, entrepreneur and lifelong athlete committed to supporting education.

Photos by Jo Szczepanska and Jakub Żerdzicki on Unsplash

Sources

Navigating the Jagged Technological Frontier Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality
A framework for determining when AI can make decisions