Say you’re tired of sorting the same kinds of files. You could let a tool move them for you. That sounds useful. Then you notice which files it would touch. Some are spare copies. Others are the only originals you have.
The steps look the same every time. The cost of a mistake does not. Choosing to automate means looking at two things: the work the tool takes over, and what could happen after it runs.
This module is for everyday work and home tasks. Pick one where you can judge what could go wrong and how you’d put it right.
What you’ll get from this module
In this module
By the end of this module, you’ll know how to choose between three options: let a tool do the job, let it do the job with set safeguards, or keep the step that matters most in human hands for now. You’ll see why the choice depends on the whole task, not just the repeated part. The evidence comes from a design framework, workplace studies, a research review and real incidents. Then you’ll pick a situation from your own life and work through it. You’ll finish with a reasoned choice, or a clear note of what you still need to find out.
Ask Jing
Become a Jingius.
What makes a task worth automating?
Design framework: separate the parts of the task.
In 2000, Raja Parasuraman, Thomas B. Sheridan and Christopher D. Wickens proposed a framework for automation. It splits a task into four parts: gathering information, analysing it, choosing what to do, and carrying out the action.
Each part can be handed over to a different degree. The framework asks designers to weigh performance, reliability and the cost of a wrong decision or action. It was written for people who design systems.
Why it matters: The useful idea for you is that handing over a job is not all or nothing. A tool can prepare information without acting on it. And a mistake can cost something different at each step.
Read the framework for types and levels of automation.
Workplace study: assistance helped agents resolve more customer issues.
In 2025, Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond studied an AI assistant at one software company. The company rolled it out to 5,172 customer-support agents in stages. Agents could edit the suggested replies or ignore them.
On average, agents resolved an estimated 15% more issues per hour. The gains were not the same for everyone. Highly skilled agents showed small drops in quality. The researchers studied the staged rollout as it happened. They did not assign agents to it at random.
Why it matters: Solving more issues per hour and keeping quality up are two different parts of good work. The average gain makes help like this worth a look. The mixed results mean the person and the task matter.
Read the study of generative AI at work.
Randomized workplace experiment: assistance made these tasks take longer.
A different result came from Joel Becker and colleagues at METR in early 2025. Sixteen experienced developers worked on 246 real tasks in open-source projects they knew well. Each task was assigned at random to allow AI or not.
With AI allowed, tasks took 19% longer. That result is about this group, these tasks and the tools of early 2025.
In February 2026, METR reported selection and measurement problems in its follow-up study. Those problems made the follow-up’s productivity estimate unreliable. The update does not show a reliable, measured reversal of the earlier result.
Why it matters: Put the two workplace studies together and the lesson is this: the finished task is what counts. Producing suggestions is only part of getting useful work done. The effect depends on the actual set-up.
Read the early-2025 developer experiment. Read the February 2026 methodological update.
Research review: adding a person changes the comparison.
In 2024, Michelle Vaccaro, Abdullah Almaatouq and Thomas W. Malone pooled 106 experiments from 74 papers. The papers were published between January 2020 and June 2023.
On average, a person and an AI working together did better than a person alone. But they did worse than the better of the two working alone, whether that was the person or the AI. Those are two different comparisons, and they answer different questions.
Beating someone’s unaided work does not show that the pair is the best option on offer. The person has to add something useful. Being present is not enough.
Why it matters: Creation tasks did better than decision tasks in this review. But the apparent edge on creation tasks was too uncertain to show that the pair beat the better performer alone. So the review does not show a general advantage for pairing people with AI on creative work.
Read the review of human–AI combinations.
Incident postmortem: a stopped backup left a later gap.
GitLab’s own account of its outage on 31 January 2017 describes an accidental deletion from its main database. Recovery was harder because the automatic backups had been failing. A version mismatch had stopped them. The error emails had been rejected too.
GitLab recovered from an older snapshot and lost some data. The failed backups were part of a wider incident. They were not the cause of the deletion itself.
Why it matters: This first-hand account shows a cost that is easy to miss. A protective task can stop without breaking anything straight away. It leaves you without the protection you need later.
Read the database-outage postmortem.
Official deployment account: protection can require automatic action.
Keeping important actions in human hands is not always the answer. In 2021, Teresa Whiting of NASA reported that an automatic ground-collision avoidance system, Auto GCAS, had saved 11 F-16 pilots since it was introduced.
NASA describes a system that warns the pilot first. If the pilot does not act, it pulls the aircraft away itself. Control goes back to the pilot once the aircraft is clear of the terrain.
Why it matters: This case shows why protective automation may need to act without waiting for a person. It belongs inside specialist safety systems. It is not a reason to change critical controls on your own.
Read the NASA account of protective aircraft automation.
What this means for you
Work out exactly how the job would run. What does the tool do? What does a person do? What happens if the result is wrong? The questions below help you connect those parts.
Name the step and the acceptable result
Describe what you want to hand over. Is it gathering information, preparing an analysis, choosing an option or carrying out an action? Then say what a good result looks like, including the exceptions that matter.
Say a tool prepares a message to a client. Preparing the draft and sending it are two different steps. A draft can sit there until someone has checked it. Nobody receives it until then.
That split only helps if someone can actually check what matters before sending. If you can’t judge the content, an approve button doesn’t fix that. Make the task smaller, or find a reviewer who can judge it.
Compare the completed work
Compare your proposed set-up with a realistic alternative. Depending on the task, that might be doing it by hand, letting the tool run alone, or a mix where each side has a useful role.
Count the whole job: setting up, checking and fixing mistakes. Keep acceptable quality in the comparison. A quick draft that needs a lot of repair may not get the job done sooner. A good assistant may still save effort once the checking is counted.
Don’t add up the percentages from the workplace studies to predict your own gain. They measured different things in different places. Use them as a reminder that you need evidence about your own task.
If you can’t show acceptable results and a worthwhile benefit, keep the step with consequences in human hands for now. That only works if you can do it well yourself. You can still consider handing over a preparatory step that is useful on its own.
Describe the failure before choosing protection
Think through three ways it could go wrong. The task stops. The result looks right but is wrong. The tool acts on the wrong thing. For each one, work out who or what would be affected afterwards.
A failure is minor only if you can live with the task being missed or delayed, and nothing that depends on it comes to harm. A protective task that stops needs a closer look. Its absence can matter later.
Go back to the file-sorting example. Sorting spare test copies is easy to undo or redo, as long as the originals are untouched and nothing else is affected. Doing the same thing to your only originals is a different matter.
Never delete originals to test this comparison. Work it through on paper, or use spare test copies and leave the originals alone.
Choose protection to match the failure you found. An approval step helps only if the reviewer can check what matters before the action goes ahead. An undo helps only if you can still undo before the harm is done.
Make recovery and responsibility credible
If your set-up relies on recovery, name who would do it, what authority they need and how much time they’d have. That might be a qualified colleague or a service, not you.
A name on a plan is not enough if that person can’t act when things go wrong. If failures can’t be noticed, contained and fixed in time, keep that step in human hands or change how it works.
Look at the work around the task too. Would someone now have to check a result they used to produce? Does another task depend on it? Would it become unclear who is responsible when something fails?
Use those answers to judge the whole set-up again. If the protection cancels out the benefit, or can’t limit the harm, it doesn’t justify handing over the action.
Never override established safety systems to try this exercise. For critical controls, use the proper qualified process. If doing the step by hand is also unsafe or beyond your skills, pause or get qualified help.
Choose the arrangement that meets the conditions
Use your answers together. Don’t just count how many sound positive.
Automate when all of these hold. You’ve shown acceptable results and a worthwhile benefit. Any failure that could reasonably happen is either easy to notice or harmless in itself. Its effects stay limited, and you can undo it or try again before real harm is done.
Automate with safeguards when the benefit is credible but a likely failure would have real consequences. You need a credible way to limit the action, check it before it’s committed, reverse it, or get capable help in time. Choose protection that actually contains the consequence you identified.
Keep the step that matters in your hands for now when you haven’t shown acceptable results or a benefit, when you still lack the judgment this situation needs, or when you can’t credibly notice, contain or recover from failures. A useful preparatory step may still be handed over.
If neither option is within your skills or properly controlled, pause and find a qualified process. Keeping work manual doesn’t settle whether you can do it well.
Take the client message. A competent check before sending could support the safeguards route. Now take an unfamiliar decision you can’t judge and can’t reverse. An approve click would leave the real problem unsolved. That needs competent human judgment, or a pause.
How about your daily life?
Where do you prepare, sort or send the same thing again and again? Which part would you like to hand over? What would you want to keep control of?
You might look at sorting spare file copies, preparing a client draft, or another everyday task whose result you can judge. Choose something you can actually change. This is a choice of what to examine. It is not yet a decision to automate it.
Write down situations you’d like to make easier. One good idea is enough. Add more if it helps.
| A situation in my daily life | What I would like to improve |
|---|---|
| ____________________ | ____________________ |
| ____________________ | ____________________ |
Which one would be most useful to work through now?
My chosen situation: ____________________
You can change your choice. Carry this situation into the exercise below.
Make a decision about one task
Start with your chosen situation. If you came straight here, pick an everyday work or home task where you can judge what could go wrong and how you’d recover. Begin on paper. You don’t need to switch anything on.
Never override established safety systems or change critical controls for this exercise. Never delete originals to test a file task. If you need an example, use spare test copies and leave the originals alone.
If doing the step by hand or by tool is unsafe or beyond your skills, pause and use a qualified process.
- Define the handover. Name the part you want to hand over: preparing, analysing, choosing or acting. Write down what counts as a good result, including the rules and exceptions that matter.
Task, boundary and acceptable result: ____________________ - Compare the completed task. Describe a realistic alternative. Write down the evidence that the quality is acceptable and the benefit is real. Count setting up, checking and fixing. If evidence is missing, mark it as a gap.
Alternative, benefit and missing evidence: ____________________ - Describe plausible failures. Think about the task stopping, giving a wrong answer that looks right, or acting on the wrong thing. Name the people, belongings or information it could affect, and anything else that depends on the result. Say whether a missed or late task would cause real harm.
Failures and consequences: ____________________ - Identify the protection and recovery. Say how a serious failure would be noticed and contained before it does harm. Name a capable person or service who could recover, with the time and authority they’d need. If a step is unverified, mark it unresolved.
Protection, recovery and unresolved conditions: ____________________ - Check the surrounding work. Note what changes about approval, dependencies and responsibility. Then look at the benefit again with that work included. An approval step is not enough if nobody can judge the result before it matters.
Changes around the task: ____________________ - Choose an arrangement, or pause. Use the conditions below together.
Choose automate only when you’ve shown acceptable results and a worthwhile benefit. Any failure that could reasonably happen must be easy to notice or harmless in itself. Its effects must stay limited, and you must be able to undo it or try again before real harm is done.
Choose automate with safeguards when the benefit is credible and a specific protection can contain the consequences. Say how the action is limited, checked before it’s committed, reversed, or recovered by a capable person in time.
Choose keep the step that matters in your hands for now when you haven’t shown acceptable results or a benefit. Choose it too when the judgment you need is unresolved, or when you can’t credibly notice, contain or recover from failures. A useful preparatory step may still be handed over.
If doing it by hand is also unsafe or beyond your skills, pause or seek qualified help.
My choice and reason: ____________________
Missing information or next useful step: ____________________
Ask Jing
Become a Jingius.
Your next step
You’ve finished when you can explain the task boundary, the benefit, the consequences and the reason for your choice. An unresolved condition with a clear next step is a useful result too.
A note on the evidence
The workplace studies are about particular people, tasks and tools. Their results don’t give a current, universal figure for productivity. The incident and deployment accounts describe specific events and systems. Use the distinctions they draw to look at your own situation, and keep the real consequences and the support you have in view.
How we prepared this module
This module draws on a design framework, workplace studies, a human–AI research review, an incident postmortem and an official deployment account. Made-up file-sorting and client-message situations explain the choices. The written exercise helps you apply the questions to one everyday task.
Sources
- Raja Parasuraman, Thomas B. Sheridan, and Christopher D. Wickens. (May 2000). A Model for Types and Levels of Human Interaction with Automation. IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, 30(3), 286–297. Author-uploaded published article; abstract, p.286; §§II–III; §IV.B.2, p.292.
- Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond. (May 2025; online 4 February 2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942. Final journal article; abstract, p.889; introduction, pp.890–893; §III, pp.898–901; §IV.A, pp.901–902.
- Joel Becker, Nate Rush, Beth Barnes, and David Rein. (10 July 2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. METR research report/preprint, arXiv:2507.09089. Institutional account; Methodology, Core Result and limitations table.
- Joel Becker, Nate Rush, Tom Cunningham, David Rein, and Khalid Mahamud. (24 February 2026). We Are Changing Our Developer Productivity Experiment Design. METR. Methodological update; opening assessment and “Wider adoption of AI has made it more difficult to measure task-level productivity”.
- 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; Results, “Overall levels of human–AI synergy” and task-type moderator analysis.
- GitLab. (10 February 2017). Postmortem of Database Outage of January 31. GitLab Engineering. First-party postmortem; Timeline; “Database backups using pg_dump”; “Recovering GitLab.com”.
- Teresa Whiting. (4 October 2021; updated 26 July 2023). Resilient Autonomy Ends, Autonomy Research Ongoing. National Aeronautics and Space Administration (NASA). Official deployment account; paragraphs beginning “Mark Skoog” and “The Auto GCAS system”.
