Say you’re checking an invoice. The spreadsheet has worked out a total, and the sum looks right. You’re about to approve it when you realise you don’t know whether the quantities use the right units.
Doing the sum again wouldn’t settle that. You need to understand what belongs in it. But you don’t need to become an expert in every task you hand over, either.
That leaves a useful question. Which parts of this work need your understanding, and which can go to a tool or to someone qualified? The answer depends on the task, on what a mistake would cost and on who can put it right.
This module is for working adults deciding what to hand over in tasks they can judge themselves or take to someone who can.
What you’ll get from this module
In this module
By the end of this module, you’ll have a clearer way to choose what to hand over and what you still need to understand. We’ll look at a workplace study, a review of experiments, research on expertise and a medical study from the recent past. You’ll see when help works, what a useful check needs and when the expertise can sit somewhere else. Then you’ll pick an everyday task and work out who should do and check each part. You’ll finish with a plan for who does what, and the knowledge you need to keep—or a clear gap that needs qualified help.
Ask Jing
Become a Jingius.
What makes handing work over succeed?
Workplace study: help can improve real performance.
In 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied AI help for 5,172 customer-support agents at one business-software firm. The system reached different parts of the workforce at different times.
The researchers compared performance across that rollout, allowing for differences over time and between agents. They estimated that agents resolved 15% more customer issues per hour, on average, with the AI help.
The gains were not the same for everyone. Less experienced workers improved. The most experienced workers gained less speed, and some of their quality measures fell. The estimate comes from comparing agents during the rollout. Workers weren’t randomly assigned to work with or without the help.
Why it matters: This gives handing work over a positive starting point: help can get real work done. But you still need to ask what improved. More issues resolved per hour doesn’t mean every worker or every quality measure got better.
Read the workplace study of generative AI assistance.
Review of experiments: the comparison changes the answer.
In 2024, Michelle Vaccaro, Abdullah Almaatouq and Thomas W. Malone reviewed 106 experiments from 74 papers. Each compared people working with AI against people alone and against AI alone, across many kinds of task and many kinds of participant.
On average, the person and the AI together did better than people alone. But together they did worse than whichever of the two—the person or the AI—was better on its own at that task.
Those two findings answer different questions. AI might help you do better than you would alone, and still do better without you. Adding a person doesn’t automatically improve the result. You need to compare the ways the task could actually be done.
Why it matters: These averages pool many different experimental tasks. They support one thing: look at the real alternatives, rather than handing every kind of work to a person, a tool or a mix of the two by habit.
Read the review of human–AI performance comparisons.
Research summary: knowledge gives a check something to work with.
A 2000 chapter from the National Research Council describes what sets experts apart. Knowledgeable people spot meaningful patterns and organise what they know around the ideas that matter. They also know when their knowledge applies.
Knowing how facts fit together helps you see what a result means. You might be able to look up a formula without knowing whether it fits your problem. To judge the answer, you need to know when that formula applies.
Why it matters: For handing work over, the question becomes what you must understand to judge this particular result. That may include how things relate, units, exceptions and limits. A general intention to think carefully can’t supply knowledge of the subject that you don’t have.
Read the chapter on how experts differ from novices.
Reader experiment: experience doesn’t protect you from bad advice.
In 2023, Thomas Dratsch and colleagues asked 27 radiologists to read mammograms with suggestions presented as coming from an AI. After 10 training cases, each read 40 test cases. In 12 of them the suggestion was deliberately wrong.
Among the five most experienced readers, the mean share of mammograms rated correctly was 82.3% when the suggestion was right and 45.5% when it was wrong. The comparison was between right and wrong suggestions on different cases, not between reading with help and reading without it.
The researchers chose which suggestions to make wrong, so the size of the gap has to be read within that setup. Still, even experienced readers were swayed by a wrong suggestion. Their expertise gave them the knowledge they needed, but it didn’t guarantee that their final rating would reject bad advice.
Why it matters: Together, these accounts explain why keeping the right knowledge matters. They also show why naming an experienced reviewer isn’t enough. What that person actually checks still matters.
Read the experiment on AI suggestions and mammography judgments.
Human-factors analysis: a check has to be possible in the time you have.
Lisanne Bainbridge’s 1983 analysis of industrial automation looks at the work left to the human operators. Watching over a process means knowing what it should be doing. Yet a person may not be able to redo a computer’s calculations, or improve on its decisions, while the process is running.
That is a problem with the role the person was given. Saying that the person is still responsible doesn’t solve it. A review arrangement needs someone who can make the judgment while there is still time to act on it.
Why it matters: The analysis is about industrial systems. Its lesson for ordinary work is to look at what checking would take, including when it has to happen, before relying on a promise to review the result.
Read the analysis of human monitoring in automated systems.
Diagnostic-accuracy study: expertise can sit elsewhere in a system built for one task.
In 2018, Michael D. Abràmoff and colleagues studied a system that read retinal images—pictures of the back of the eye—in primary care. It enrolled 900 adults with diabetes and no history of diabetic retinopathy. 819 of them could be fully assessed.
The system looked for more-than-mild diabetic retinopathy, an eye disease linked to diabetes. The researchers compared its results with assessments from a specialist reading centre. The way it was meant to be used, no expert had to read every image. The system did miss some cases, and specialist expertise and referral routes stayed part of the wider arrangement.
This example from the recent past shows why every user needn’t repeat every specialist judgment. The system was tested for one task in a defined group of patients. The lesson is about where the expertise sits. It is not a recommendation to use a general tool to diagnose yourself.
Why it matters: The study was funded by the system’s developer, and the authors declared commercial ties. The evidence covers that defined clinical arrangement, with its limits and its specialist support.
Read the trial of autonomous retinal-image interpretation.
Online experiment: feeling able to explain is not the same as being able to.
In 2015, Matthew Fisher, Mariel K. Goddu and Frank C. Keil ran an experiment, Experiment 1a, with people online. Some searched the internet for explanations. Others didn’t. Then everyone rated how well they could explain subjects that had nothing to do with the search.
The search group rated themselves higher: 3.61 against 3.07 on a seven-point scale. These were self-ratings, not tests of what they knew. The two groups also spent different amounts of time on the first task.
The researchers hadn’t measured these ratings before the first task. So this experiment on its own couldn’t say whether searching raised people’s ratings, or whether going without search lowered them. What it does show is a difference in how able people felt after different ways of getting information.
Why it matters: When you decide what to hand over, don’t mistake a familiar explanation for proof that you can judge the work. How confident you feel and how well you can judge a result are different questions.
Read the experiments on searching and self-rated knowledge.
Research summary: gut feeling needs conditions it can learn from.
In 2009, Daniel Kahneman and Gary Klein asked when an expert’s gut feeling can be trusted. The situation needs patterns that are regular enough to learn. You also need chances to learn those patterns through experience, with feedback that tells you whether your judgments were right.
Feeling certain doesn’t show that those conditions were there. Years of experience matter partly because of what they gave you the chance to learn.
Why it matters: This sets a limit on handing work over, too. A weak gut feeling on your side doesn’t prove that a particular machine is reliable. Each way of doing the task needs its own evidence.
Read the account of conditions for intuitive expertise.
What this means for you
Start by looking at the work, rather than choosing a broad skill to keep or give up. Handing over the work and handing over the judgment about its quality are two separate choices. You might check a tool’s result yourself, or arrange for someone qualified to check it. Either way, the arrangement needs evidence that it can work, and someone who can act on an important mistake.
Split the task before deciding which skill to keep
Name the result you need. Then separate the information going in, the working out, the decision and the action that follows. A different person or tool can handle each part.
In the invoice example, working out the total is one part. Choosing the quantities and units is another. Deciding whether to approve the invoice is a further judgment.
If you are checking quantities and units, you need to understand how they fit together. Handing over the sum doesn’t answer that question. If an unfamiliar term needs knowledge you don’t have, someone qualified needs to look at that part.
This is why “keep maths” or “hand over admin” is too broad. Breaking the task down shows which judgments stay yours. It also shows what you need to know about a tool, or about another person’s ability, before handing over the rest.
Name an important mistake and who can act on it
Describe a mistake that would matter, when someone would have to notice it and who could stop or correct it. Say who resolves a problem the first reviewer can’t. Include losing access to the tool or the information, if that would stop the work or stop you putting things right.
For the invoice, the wrong units could make the total wrong for its purpose even when the sum is correct. Checking those units only after approval would miss the point. The decision depends on getting them right first.
Now look at the proposed reviewer. Do they understand the quantities and the rules the sum has to follow? Can they look at them in time? Who handles a problem they can’t resolve?
“I’ll review it” becomes useful when those answers describe a check that someone can actually do. If the knowledge or the timely response is missing, get qualified help or narrow the task before relying on the arrangement.
Use evidence for your own ability and for the delegate’s
Look for past work that someone else checked, fair comparisons between the options you have, or review by someone qualified. The evidence needs to be about this task and the mistakes that matter. If you don’t have an answer, write “unknown”.
An answer can look familiar without showing that you could spot an important mistake. The self-rating experiment explains why feeling confident isn’t enough to settle who should judge the work.
If you rely on gut feeling, ask whether the situation has patterns you could have learned. Has feedback told you when your judgments were right or wrong? If you don’t know, keep that doubt in view.
Ask for evidence about the alternative too. Doubting your own ability doesn’t show that a tool is reliable. You may need a competent specialist when neither your own check nor the automated answer has enough support.
Choose where the expertise will sit
Use the evidence to decide who will do the work, who will judge it and who will deal with problems. These jobs don’t all have to stay with you.
| Who does what | When it fits | What stays with you |
|---|---|---|
| Hand over storing or doing | Evidence about performance, the cost of mistakes, the checks available and access supports handing it over. | Know the purpose and the inputs that matter. Do any checks or responses to problems that are assigned to you. |
| Hand over the doing and keep the judging | You need to judge important mistakes and can do so competently, in time for someone to act. | Understand the ideas, relationships, rules and limits that judgment needs. |
| Use a competent specialist, or a system tested for one specific task | You can’t or needn’t do the expert checking yourself, and the qualified alternative covers the work. | Explain what you need, spot what the arrangement doesn’t cover and know where to take an unresolved problem. |
| Leave the decision open | You lack the evidence you need about someone’s competence, the tool’s reliability or how problems will be handled. | Get qualified judgment or narrow the task before treating the arrangement as sound. |
For the invoice, the evidence may support letting the spreadsheet do the sum while you stay responsible for quantities and units. That doesn’t qualify you to interpret every unfamiliar term. Put a specialist judgment with someone qualified when the task needs it.
The eye-disease example shows how expertise can be part of a defined system without each user repeating the expert judgment. You still need to know what that system covers and where unresolved cases go.
Name the knowledge your remaining role needs
Once the split is clear, write down what you still need to understand. Name the ideas, patterns, quantities, exceptions and limits your checks need, and what you need for your response when something goes wrong.
“Understand the invoice” leaves the job unclear. “Know which quantities belong in the total, their units and whether the result looks plausible” gives you something specific to compare with your own ability.
If you can’t show that you have that ability, keep the gap in view and find qualified help. Look at the split again when your responsibilities change. A new responsibility may need knowledge or checks that weren’t part of your old role.
How about your daily life?
Where do you accept a tool’s result while still making a decision yourself? You might check spreadsheet sums, pull up stored information or review work before someone acts on it.
Which part do you want help with: doing the work, deciding whether it’s right or dealing with a problem? Write down situations where a clearer split would help. One suitable idea is enough.
| Situation I want to look at | What I need to understand or improve |
|---|---|
| ____________________ | ____________________ |
| ____________________ | ____________________ |
Which situation can you judge yourself, or take to someone who can?
My chosen situation: ____________________
Carry it into the exercise below. Choosing a situation doesn’t yet show that the way you hand it over now is sound.
Choose who does each part
Start with your chosen situation. If you came straight here, pick one everyday task you can judge or take to someone who can. Work through it on paper before relying on a new split of responsibility.
- Describe the task and its parts. Write the result you need, the information going in, the working out, the decision and the action that follows. Mark the part you want to hand over.
Task, parts and proposed delegation: ____________________ - Name an important mistake and the response. Describe what could go wrong, how soon someone would have to notice and who could stop or correct it. Name who would handle a problem the first reviewer can’t resolve. Include loss of access if it would stop the work or the response.
Error, time to notice and person responsible for the response: ____________________ - Record evidence for each proposed role. Look for past work that someone else checked, fair comparisons between options or review by someone qualified. Mark unknown ability or reliability as “unknown”. Confidence and familiar wording don’t answer these questions.
If you rely on gut feeling, name the patterns you could have learned and the feedback you’ve had on your judgments. Check the proposed delegate’s ability separately. Doubt about your own ability doesn’t prove theirs.
Evidence for me, the tool or the reviewer; unknowns: ____________________ - Choose the split the evidence supports. Hand over storing or doing when the evidence supports it, weighing performance, the cost of mistakes, the checks available and access. Keep the judging with you when your role needs it and you can judge important mistakes competently and in time.
Use a competent specialist, or a system tested for one specific task, when you can’t or needn’t do the expert checking yourself. That alternative has to cover the actual work. Say what it covers and where unclear or failed cases will go.
Leave the decision open if you lack the evidence you need about competence, reliability or how problems will be handled. Get qualified judgment or narrow the task before treating the split as sound.
Division of work, or unresolved decision: ____________________ - Say what stays yours. Name the purpose, the inputs, the checks and the response to problems that are assigned to you. Record the ideas, relationships, units, rules and limits you need to understand for those jobs.
If the expertise sits elsewhere, say how you’ll spot what the arrangement doesn’t cover and where you’ll take an unresolved problem. Keep any remaining gaps in view, and name the next piece of information or qualified help you need.
Knowledge and responsibilities I retain: ____________________
Unresolved gap and next step: ____________________
Ask Jing
Become a Jingius.
Your next step
You have finished when you can explain the split you’ve proposed and the evidence behind it, or say exactly why it is still open. Keep the record and look at it again when your responsibilities change.
A note on the evidence
The workplace study and the review of experiments show why performance depends on the task, the person and the comparison. The reader experiment shows how wrong advice can sway people, within the limits of its design. The research on expertise and on industrial monitoring helps explain what a meaningful check needs.
The eye-disease example is about one defined clinical use with specialist support. It doesn’t make general-purpose systems interchangeable with that arrangement. Your decision still needs evidence about the actual task, and a route for important mistakes.
How we prepared this module
This module brings together research on assisted performance, expertise, self-assessment and human oversight, with a diagnostic study from the recent past. A made-up invoice example shows how working something out and judging it can be split. The written exercise helps you look at the knowledge and responsibility attached to one task.
Sources
- Erik Brynjolfsson, Danielle Li, and Lindsey Raymond. (May 2025; advance publication 4 February 2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. Final published article; abstract p.889, introduction pp.891–892 and §IV.C p.904.
- Michelle Vaccaro, Abdullah Almaatouq, and Thomas W. Malone. (28 October 2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. Published article, Europe PMC XML; Results opening and “Overall levels of human–AI synergy”; Methods, “Literature review.”
- National Research Council. (2000). How People Learn: Brain, Mind, Experience, and School: Expanded Edition; Chapter 2, How Experts Differ from Novices. National Academies Press. Published chapter; opening principles p.31, “Organization of Knowledge,” and “Context and Access to Knowledge.”
- Thomas Dratsch, Xue Chen, Mohammad Rezazade Mehrizi, Roman Kloeckner, Aline Mähringer-Kunz, Michael Püsken, Bettina Baeßler, Stephanie Sauer, David Maintz, and Daniel Pinto dos Santos. (2023). Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology, 307(4), e222176. University-hosted advance article with provisional headers; printed pp.1–3, Table 3 p.7 and Discussion limitations pp.8–9.
- Lisanne Bainbridge. (November 1983). Ironies of Automation. Automatica, 19(6), 775–779. Published article scan; “Monitoring,” p.776.
- Michael D. Abràmoff, Philip T. Lavin, Michele Birch, Nilay Shah, and James C. Folk. (28 August 2018). Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices. npj Digital Medicine, 1, article 39. Published article; Results, “Primary care participants” and “AI system characteristics”; Discussion, paragraph containing “Primary care providers”; Methods and Competing interests.
- Matthew Fisher, Mariel K. Goddu, and Frank C. Keil. (2015). Searching for Explanations: How the Internet Inflates Estimates of Internal Knowledge. Journal of Experimental Psychology: General, 144(3), 674–687. Published article; Experiment 1a, Method and Results, pp.675–676.
- Daniel Kahneman and Gary Klein. (September 2009). Conditions for Intuitive Expertise: A Failure to Disagree. American Psychologist, 64(6), 515–526. Published article; “Conditions for Skilled Intuition,” pp.520–522, and Conclusions, pp.524–525.
