ManageTime

Article

An AI can rank your to-do list: which part of that job to keep

Keep the ranking decision and hand over the clerical half. An AI tool can gather your tasks, strip the duplicates and group what belongs together; deciding which one you do first is a judgment about work only you can see. The strongest evidence on human-AI teams finds the combination loses ground on exactly that kind of decision. Every ranking feature on the market is sold by the company that built it, so the claim arrives without a control group. This page covers what the research on human-AI teams actually measured, which half of prioritizing is safe to delegate, and how to check an order that comes back.

Should you let an AI tool decide what to work on first?

A button at the top of the task list, offering to sort it for you. Most task tools have grown one over the past two years, and the pitch is the same everywhere: paste in the backlog, get back a ranked plan. The offer is real, and the ranking does arrive in seconds.

What it hides is that prioritizing is two jobs rather than one. The first is clerical. Somebody has to collect what you are actually holding, pull the tasks out of threads and meeting notes, strip the duplicates, and group what belongs together. The second job is the decision: this one first, that one Thursday, that one never. A tool is better at the first job than most people expect, and the second job is the one the button is offering to take.

Those two jobs run on different inputs.

Collecting runs on what is written down, which is precisely what a tool can read. Ranking runs on what is not. Your manager mentioned the board pack in passing. The colleague waiting on your review is blocked until Thursday. Friday's deadline is the kind that moves and next month's is not. Feed a tool the text of your tasks and it ranks the text of your tasks, which is half the picture, chosen by whatever you happened to type.

A ranked list from a tool is also harder to argue with than one you wrote yourself. You know which of your own lines are vague; the generated order comes back as a clean numbered column and the vagueness has gone from the surface. The prioritization pillar treats ranking as a routine you repeat rather than a problem you solve once, and a routine that hands out its judgment stops teaching you anything about your own week.

So the first move is a split rather than a verdict. Give the tool the gathering: merge the four places your tasks live, name the duplicates, group what shares a deadline or a stakeholder. Then close the tab and rank the top three yourself.

Is an AI tool any good at ranking your to-do list?

Nobody has measured it. Not on a to-do list, and not on anything near enough to count, so what you have to reason from is evidence about human-AI teams in general. It is unflattering in a way that lands directly on your half of the question.

Vaccaro, Almaatouq and Malone published the review in Nature Human Behaviour in 2024. They pooled 106 experimental studies reporting 370 effect sizes, published between January 2020 and June 2023, and every included study had to measure three things: how people performed alone, how the AI performed alone, and how the two performed together. On average the combination came in below the better of the two alone, at a Hedges' g of -0.23, with a confidence interval running from -0.39 to -0.07.

The average is not the useful part.

Underneath it sits a split by task type. Where the task involved making a decision, the combinations lost ground, and that loss held up statistically. Where it involved creating content they gained instead, though on far fewer studies and by a margin too slim to reach significance on its own. What the review could show was that the two task types pull in opposite directions. A second split runs alongside that one. When the people were better than the AI alone, pairing them helped, and when the AI was better than the people alone, pairing them hurt. So the combination tended to pull the stronger party toward the weaker one rather than landing above both, which is worth knowing before you assume your own ranking plus a model beats either of you.

Hold that against your own two jobs from the previous section and the mapping is uncomfortably clean. Assembling your list is content work. Ranking it is a decision. The review's worst result sits on exactly the half of prioritizing that a vendor demo hands to the model.

The limits here are large, though, and they cut both ways. Not one of those 106 experiments involved anybody's own task list. The underlying studies cover work like sorting images and spotting fake reviews, none of it shaped like your Tuesday, and the authors name publication bias and the variety of study designs among the limitations of their own review. So the task-type split is a mechanism you can reason with, not a measured finding about backlogs. Anyone who tells you the research shows AI ranks work badly is overreading it. Anyone selling you a ranking feature has no measured result either.

How do you check whether an AI ranking of your work is any good?

Start by assuming the warning label does nothing, because that part has been tested.

Kiyak, Coskun and Budakoglu ran a randomized controlled trial with 186 fourth-year medical students, published in Medical Education in 2026. Each student worked three deliberately ambiguous clinical cases, chose a diagnosis, then read feedback attributed to ChatGPT arguing for the other option. Half of them saw a prominent warning above it: 'ChatGPT can make mistakes. Check important info'. The warning changed nothing. Students revised a diagnosis 15.3 percent of the time without it and 15.9 percent with it, an odds ratio of 1.09 whose confidence interval ran from 0.46 to 2.59.

The authors' explanation is worth more here than the null result. Those students were already discounting the AI heavily, with a weight-of-advice score of 0.15 against the 0.30 average from earlier advice-taking research, so there was no excess trust left for a disclaimer to remove. Whatever the small print under the box achieves, it does not shift how much weight you give the answer.

The check has to be structural instead.

The cheapest one takes a minute. Write your own top three before you ask, then compare the two orders. Agreement tells you nothing much. A disagreement is the entire point, because now you have to say why you put the review first, and either you have a reason the tool could not have known or you find out you were ranking by what felt easiest to start.

One more reason to rank first and ask second comes out of the advice-taking literature. Across six experiments reported in Organizational Behavior and Human Decision Processes in 2019, Logg, Minson and Moore found that lay participants leaned on advice more heavily when they thought an algorithm was behind it than when they thought a person was. Two conditions weakened that lean. The first was a comparison against the participant's own estimate rather than an outside adviser's. The second was expertise: people who forecast for a living discounted the algorithm more heavily than novices, and paid for that discount in accuracy.

Your own task list is where both exceptions arrive together. You are the resident expert on your work, and the thing the ranking competes with is your own judgment rather than a colleague's. The prediction that follows is not that you will over-trust the output. It is that you will skim it, keep your original order, and have spent four minutes producing a document you ignored. Those experiments ran on estimates and forecasts rather than backlogs, so read them as a description of the pull, not a measurement of your Tuesday.

How does your working style change the answer?

Which half of this you get wrong is fairly predictable from the way you already work. The four names below are this site's own, from the working-style self-check on the home page.

Architects are most at risk from a good-looking output. A ranked, numbered, reasoned list is the artifact an Architect wanted, it arrives already formatted, and interrogating it feels like vandalism. Sprinters fail the other way round. No Sprinter waits on a tool mid-burst, so the ranking gets consulted in the flat hour after lunch, and whatever sits on top of it goes next.

Visionaries lose the goal work. A tool ranks what it can read, and the long-horizon project that lives in your head rather than in a ticket cannot compete with ten written tasks, so it sinks a little further every time the list is regenerated. Improvisers have the strongest case for the tool and the least use for its ranking: their problem is that the list does not exist in one place at all, so for them the gathering half is the whole value.

Two cases sit outside what this page covers. A shared team backlog scored against agreed criteria is a different object, because the inputs get written down for several people to see, and a tool applying a stated rule to stated numbers is doing arithmetic rather than judgment. A sort by due date is the same thing: useful, mechanical, and nothing you need a model for.

If you want the split without the subscription, the Ivy Lee Method already runs it. Six tasks chosen the evening before, in order, worked top down. A tool can produce the candidate six. The order is what deserves your own five minutes, and the Eisenhower Matrix gives you two questions to produce it with. Ask tonight for the list, then rank it yourself before you shut the laptop.

Frequently asked questions

Should you let an AI tool decide what to work on first?

Hand over the gathering and grouping, and keep the ranking call yourself. Collecting tasks runs on what is written down, which a tool can read. Ranking runs on what is not written down: who is blocked, which deadline moves under pressure, what your manager said in passing. A tool fed only the text of your tasks ranks only the text of your tasks.

Is AI any good at prioritizing your to-do list?

Nobody has tested it on a to-do list, so the honest answer is that it is unmeasured. The nearest evidence is a 2024 Nature Human Behaviour review of 106 experiments. It found human-AI combinations scoring below the better of human or AI alone (Hedges' g = -0.23), with the losses in decision tasks and a content-creation gain too slim to be sure of. No experiment in the set used a personal task list, so treat the task-type split as a mechanism rather than a measured result.

What should you use AI for on your task list, and what should you keep?

Use it to collect, de-duplicate and group; keep every judgment about what matters. Pulling tasks out of four tools and a week of meeting notes is clerical work you are slow at and it is slow at nothing. Deciding that the half-finished review beats the report nobody has asked for yet needs context that never got typed anywhere. The first job is worth paying for and the second is the job.

How do you check whether an AI ranking of your work is any good?

Write your own top three before you look at its order, then compare the two. The disagreements are where the information is: each one forces you to name a reason, and the reason is either something the tool could not have known or a preference you had not noticed. Do not expect the warning under the box to help. A 2026 randomized trial with 186 medical students found that a 'ChatGPT can make mistakes' notice changed advice-taking not at all (15.3 percent versus 15.9 percent).

Sources