task = one piece of work. A job is a bundle of tasks.
A job is not one thing. A nurse charts, takes vitals, explains a procedure, notices that something is off, and comforts a worried family. A bike-shop worker answers messages, builds bikes, handles refunds, and chats with regulars. When people ask “Will AI take this job?” a more useful question is smaller: which tasks, and what happens to the rest?
Two things a tool can do to a task
Substitute and complement
Substitute.The tool does a task a worker used to do. The worker's time isn't needed for that task anymore.
Complement.The tool helps the worker do the task faster or better. The worker's judgment is still part of it.
Notice that the question is about how the tool is used, not what the tool is. The same software can be either, depending on how people set it up.
ExampleOne screening tool, two designs
An invented company gets 2,000 applications for 10 jobs, so it buys a résumé screening tool.
Step 1 — Design 1: substitute
The tool rejects anything below a score, and no recruiter ever sees the rejected résumés. The screening task now belongs to the tool.
⚠️Watch out: “substitute” doesn't mean “the job disappears,” and “complement” doesn't mean “safe.” A substitute replaces a task. The rest of the job may stay, change, or shrink. A complement can still change which skills matter and how many people a team needs. Nobody can promise either outcome for every job.
New tasks show up too
When a tool takes over a task, other tasks usually appear: setting it up, checking its work, handling the cases it gets wrong, and answering the person who says “that's not right.” Whether those add up to as much work as the tool removed depends on the job. In 3.2.4 you'll look at real jobs task by task.
Try it
Which is it? · invented situations
For each situation, decide how the tool is being used. Ask: is a person's judgment still part of the task?
A small newspaper uses AI to write its daily weather blurb and posts it without anyone reading it first.
A company's tool rejects résumés that don't match its keyword list, and no recruiter ever sees the rejected ones.
A clinic's billing tool reads each insurance form and enters the numbers on its own, so no one types them anymore.
An electrician uses an AI app to look up code rules on site, then decides what to install.
A designer asks AI for ten rough logo ideas, picks one, and redraws it by hand.
A paralegal uses AI to find the contract lines that mention deadlines, then reads each one to decide what matters.
Correct so far: 0 / 6
Principles of AI · 3.2.1SkillConcept
Four Questions Before You Hand It Over
hand-off level = how much of a task the AI does and how much a person stays involved
“Can AI do it?” is the wrong first question. AI can make an attempt at almost any task. The questions that matter are: should it do this task alone, and how much should a person stay involved?Four questions help you decide.
The four questions
Routine or judgment?Are the steps the same every time, with a clear rule? Or does each case need someone to weigh things the rule can't list?
How costly is a mistake?If the AI is wrong, what does it cost, and can it be undone?
Does it depend on trust?Does the task depend on someone trusting a person, or on a relationship?
Who checks the AI's work?Would a person catch a wrong result before it hurts someone, and could they say no?
Score the answers
For each question, the facts of the task point one way or the other. Score 0 if it favors letting AI do the task, 1 if it's in between, and 2 if it favors a person. Add up the four scores, from 0 to 8. The total suggests a starting level.
Four hand-off levels
Level 1 · 0–1 points.AI does it. People sample the results and handle complaints. (closest to a substitute)
Level 2 · 2–4 points.AI drafts. A person reviews every result before it is used. (a complement)
Level 3 · 5–6 points.AI assists. A person makes the decision. (a complement)
Level 4 · 7–8 points.A person does it. AI stays out of the core decision.
Two stop signs
A total can hide one very bad answer, so two answers override it. They only ever move a task toward more human involvement.
Stop signs
Hard to undo.If a mistake can't be undone, nothing goes out unread. Level 2 at the least.
No check.If no one would catch a mistake, don't hand the task over until someone can. Level 3 at the least.
ExampleOne insurer, two tasks
Harborview Health Plan, an invented insurer, uses the same AI system for two tasks. Score each one.
Step 1 — Approve clear matches
Answers: 0 + 0 + 0 + 1 = 1 points. A wrong approval is a small payment that audits can catch. Level 1: AI does it. People sample the results and handle complaints.
🔑Question 4 is about the plan, not the task. Whether a task is routine, whether a mistake is costly, and whether trust matters are mostly facts about the work. Who checks the AI's work is a design choice. Building a real check is how a team earns the right to hand over more.
⚠️A tally is not a verdict. These points are a simplified practice model for organizing your thinking. People can reasonably weigh the questions differently, and real decisions involve laws, contracts and many more facts.
Try it
Pick a task from each field, answer the four questions, and see where it lands. Then try the change and see what moves.
Four-questions lab · invented tasks · nothing here is saved
Retail
Health care
Trades
Creative work
Answer "Where is my order?" messages
Pine & Pedal, an invented bike shop, gets hundreds of these a week.
1 · Routine or judgment?
“Every message asks the same thing, and the answer is in the tracking system.”
2 · How costly is a mistake?
“A wrong update gets corrected by the next message. Nobody is hurt.”
3 · Does it depend on trust?
“Shoppers want a fast, correct update, not a conversation.”
4 · Who checks the AI's work?
“A manager reads a sample of chats each week, and a shopper can ask for a person.”
Principles of AI · 3.2.1ConceptSkill
A Review That Works
human in the loop = a person who is part of the process and can check the result and change it
“A person looks at it” sounds like a safety net. It is one only if four things are true:
Time. The reviewer can actually look at the work.
Evidence. The reviewer can see what the AI used, or enough to judge it.
Power. The reviewer can say no, and saying no changes the result.
Attention. The reviewer hasn't slid into just agreeing. People tend to stop looking when a tool is usually right, so a good process checks for that too.
Two reported cases
These are real, as described in news coverage. Details are limited, and the companies' accounts differ, so treat them as examples to think with.
Reported case · realA review in seconds
In 2023, ProPublica reported that doctors at Cigna, a large U.S. health insurer, signed off on denials in batches. It said they denied more than 300,000 claims over two months in 2022, spending about 1.2 seconds on each on average. Cigna said it disagrees with that description, and that the system automatically pays claims with correct codes and lets doctors focus on complex reviews. We can't settle that here. The question to take away is the test: how long would a real review take?
Reported case · realA tool that learned the past
In 2018, Reuters reported that Amazon had built an experimental tool to rate résumés, trained on résumés the company received over ten years, most of them from men. It reported that the tool learned to mark down résumés containing the word “women's” and graduates of two all-women's colleges, that recruiters looked at its recommendations but never relied only on its rankings, and that the effort was dropped. A learned tool copies patterns in its training data. A reviewer who only sees the tool's ranking may never see who was pushed down.
Two kinds of tools, two kinds of mistakes
Rules and learned models
A rule-based filtergives the same answer for the same input. That's called deterministic. Its mistakes are consistent: a gap in its word list misses the same kind of person every time. Testing can find the gap, and changing the list can fix it.
A learned model or chatbotworks from patterns in its training data. Its answers can vary from run to run, which is called probabilistic. Its mistakes can be inconsistent and harder to explain, so one test isn't enough. It needs checking on an ongoing basis.
⚠️A chatbot is not a résumé scanner. Pasting a résumé into a chatbot tells you what that chatbot says today. It doesn't tell you how any company's screening system works. Many real systems mix rules and learned models, and companies don't always say which.
Here is a rule-based filter in action. The posting, the résumés and the word list are all invented.
Keyword scanner · invented posting and résumés
A library needs an event helper who can organize, design, work with volunteers and communicate. The scanner counts how many words from its list appear in a résumé, and passes anyone with 3 or more. Click words to add them to or remove them from the list.
Résumé B describes the same kind of work as A, and it was rejected. Résumé C only lists words, and it passed. The filter gave exactly this answer every time.
Run it again and you get the same result. That's what makes a rule-based filter's mistakes consistent, and why a person can find them by testing. Real tools differ, and many use more than a word list.
Mistakes aren't equal
In the invented insurer from the last section, a wrong approval costs money that audits can recover. A wrong denial can delay care. The AI could make mistakes at the same rate in both cases and the cost would be very different. So when a reviewer can't look at everything, the question is: where would a mistake hurt most, and be hardest to undo?
💡Confidence is not correctness. A model can report that it's 96% sure and still be wrong, the same way a chatbot can state a false fact in a confident voice. The queue below is invented, and the confidence numbers are made up, but the habit of mind is real: don't use how sure the AI sounds as your only guide.
Review queue · Harborview Health Plan (invented) · nothing here is saved
An AI decided all 8 of these claims this morning. A reviewer has time to look at only 3. Choose which 3 to send for review. You can see what's at stake if the AI is wrong, but not whether it is wrong.
Flu shotAI says APPROVE · 97% sure · stake 1/4
If wrong: A small payment the plan didn't need to make. Audits can recover it.
If wrong: A covered procedure is refused until an appeal.
Reasons and a way back
Review inside a company isn't the only check. A person affected by an automated decision should be able to learn that a system was involved, get the main reasons, and ask a person to look again. That's about fairness, and it's also a check: it gives mistakes a way to reach someone who can fix them. Remember question 4. “Only if the person appeals” is a weaker check than “a trained reviewer reads every one,” and the people with the least time and money to appeal are the ones it protects least.
💡Your side of it: every shop, insurer, claim and résumé on this page is invented. Don't paste a real résumé, medical information, or anyone else's details into a chatbot to “test” it. If your teacher asks you to try a chatbot, use made-up examples only.
🔑Next you'll look at what jobs are made of. In 3.2.2 you'll build a skill stack, and in 3.2.4 you'll score a real job task by task. The four questions and the review test are the tools you'll use.