Start here: you already know how to hand off work#
Think of the last time someone helped you with a job: a classmate on a group project, or a new hire. You said what you needed and looked over what they did. If something went wrong, it was still your name on it. So you already know how to delegate, which means handing a piece of work to someone else while you stay in charge of the whole.
Working with AI has the same shape, with one new part. This helper is fast, tireless and confident, and it cannot answer for anything. When someone asks about a wrong number, a person answers. This module is about that one difference.
Meet Elena. She is three weeks into a job as office coordinator at Greenline Landscaping, a made-up company with 24 staff in three crews. On Monday her manager, Grace, asks for two things by Friday:
- The monthly report: jobs done, fuel costs and overtime, from the company's job spreadsheet.
- A team announcement: paper timesheets are being replaced next month by an app on staff phones.
Elena has used a chat assistant a few times, mostly for recipes. We follow her through the week. Students: swap in a group project report, and every move carries over.
Elena's report will reach Grace with Elena's name on it. If the AI slips in a wrong number, who will Grace ask about it?
Show the answer
Elena. Grace will never talk to the AI, and “the AI wrote that” does not fix the number. The tool does the work; a person answers for it.
What to hand over and what to keep#
Job one: “turn this list of numbers into a table.” Job two: “tell me whether our overtime pay follows the law.” Which can Elena hand to an AI?
Show the answer
Job one, because she can check the table at a glance. Job two has real consequences and she cannot check the answer. It goes to the person whose job it is, such as Grace or whoever runs payroll.
Two questions sort almost every task. Can I check it? And who lives with the result? If you can check it and little rides on it, hand it over. If you cannot check it, or it lands on people, money or promises, keep it.
| Kind of task | Hand over or keep? | Elena's example |
|---|---|---|
| Drafts | Hand over, then rewrite | First draft of the announcement |
| Summaries of material you provide | Hand over, then check | Meeting notes she pastes in |
| First passes | Hand over | “What is missing from this outline?” |
| Formatting | Hand over | Turning her numbers into a table |
| Research leads | Hand over, then look up | “What do staff ask when timesheets change?” |
| Decisions with consequences | Keep | The date paper timesheets stop (Grace's call) |
| Anything you cannot check | Keep | Whether overtime rules are followed |
| Relationships | Keep | Replying to an unhappy crew lead |
| Final responsibility | Keep, always | Her name on the report |
Note the words “material you provide”. “Summarize these notes I pasted” is easy to check. “Summarize the overtime rules” asks the AI to remember, with no page to check it against.
What careful studies found
Erik Brynjolfsson, Danielle Li and Lindsey Raymond ran a field study, which is research in a real workplace rather than a lab. They followed 5,172 customer-support agents as an AI chat assistant reached different groups at different times. Issues resolved per hour rose 15% on average. Less experienced agents improved in speed and quality. The most experienced agents gained a little speed and saw small declines in quality.
Fabrizio Dell'Acqua and colleagues ran an experiment with 758 knowledge workers, together with the consulting firm Boston Consulting Group. Each person was randomly assigned to work with or without AI. On 18 realistic tasks the AI handled well, AI users completed 12.2% more tasks, finished them 25.1% more quickly and produced higher quality work. On one task chosen to sit just outside what the AI could do, AI users were 19% less likely to reach a correct answer. The authors call this uneven edge a jagged frontier. The AI is strong on some tasks and weak on others that look just as hard.
So how hard a task looks does not tell you whether the AI will get it right. As a newcomer, Elena may gain the most, but may also be the least able to spot a confident mistake. The check has to come from her.
When an AI names a rule, a statistic or a source, look it up in the real document. A wrong lead costs you five minutes. A wrong “fact” in a report costs you the report.
Ask for a recommendation, then choose#
Elena types: “Should I send the announcement by email?” What is weak about that question?
Show the answer
It invites a yes or no and hides the other options. It also tells the AI nothing about her team. A better question asks for options, a recommendation with reasons, and what would change it.
On a decision, do not ask the AI to decide. Ask it to lay out the choice, pick one with reasons, and say what would change its mind. Here is Elena's prompt:
I'm an office coordinator. I need to tell 24 landscaping staff in three
crews that paper timesheets end next month and a phone app replaces them.
Give me three ways to announce it, each with one upside and one risk.
Recommend one with your reasons. Then tell me what fact about my team
would change your recommendation.The last line makes the AI show its assumptions, and you know whether they are true. The reply recommends an email, unless staff rarely read email. Elena knows the crews read texts and are at job sites by 7 a.m. So she chooses a five-minute huddle at the yard, plus a text to the three crew leads. The AI did useful thinking. Elena made the call with knowledge it did not have.
Some decisions are not Elena's either: the date paper timesheets stop is Grace's. Keeping decisions includes knowing whose they are.
Kestrel Exams is a small study site built by one person working with an AI assistant. When a decision is genuinely the owner's, our AI assistant asks a short multiple-choice question. Examples: which payment service to use, what a button should say, how big a course should be. Each question comes with a recommendation and a reason. Small technical choices it makes itself. The rule: delegate the work, keep the decisions with consequences. Ask the AI for a recommendation and its reasoning, then choose.
Use a harmless made-up decision. Paste this:
I'm organizing a study group of 5 people for a test in two weeks. We
all have different schedules. Give me three ways to plan our sessions,
each with one upside and one risk. Recommend one with your reasons, then
tell me what fact about our group would change your recommendation.What to look for: Does every option have a real risk? Are the reasons ones you can test? Does the last part name a fact you actually know? Then finish this sentence: “I chose ___ because ___.” If the only reason is “the AI said so”, you have not chosen yet.
Three duties: disclose, verify, own#
Elena's first draft says “Fuel costs fell 8% from last month.” She only pasted this month's numbers. Where did the 8% come from?
Show the answer
From nowhere she can check. With no figures for last month, the AI wrote a comparison that sounded right. Elena deletes the line, or gets last month's real numbers and works out the change herself.
Three duties stop slips like the 8% from shipping, at school or at work.
| Duty | What it means | What Elena does |
|---|---|---|
| Disclose | Say plainly how AI was used where it matters: when rules ask, or when someone relies on the work. | Adds a line: “First draft written with an AI assistant. Every number checked against the October job sheet.” |
| Verify | Check the output against something real. Start with numbers, names, dates and promises. | Traces every number to the sheet. Cuts the 8%. |
| Own | The result is yours. If it is wrong, you fix it and say so. | Sends it under her name, ready for any question. |
Disclosure is not a confession. It tells the reader what was checked, which helps them trust the work.
These three duties come from a proposal about AI rules at a university. The person who runs Kestrel Exams wrote it. Instead of bans, it asked everyone who uses AI to do three things. Disclose how AI was used, verify what it produced, and own the result as their responsibility. We try to follow the same three while building this site with our AI assistant, including this course.
In the 2023 US case Mata v. Avianca, lawyers were fined for filing court decisions that a chat assistant had invented. The judge wrote that using a reliable AI tool is not improper in itself; checking the filing was the lawyers' job. Module 03 tells the full story.
Rules first: your policy and private data#
Greenline has no written AI policy. Does that mean Elena can use any AI tool for anything?
Show the answer
No. It means nobody has decided yet. If something goes wrong, Elena and Grace still answer for it. The fix is cheap: ask.
An AI use policy is a set of written rules from an employer or school. It says which AI tools you may use, for what, and with what information. Where one exists, follow it. At work, check the staff handbook or ask your manager. At school, read the school's rules on honest work (often called an academic integrity policy). Check each class's rules too, since they can differ from class to class.
Where there is no policy, ask a few short questions and keep the answers in writing:
- Which AI tools may I use, and on a work or a personal account?
- What information must never go into them?
- Do you want to know when my work was drafted with AI?
- Is there any work where I should not use AI at all?
- Who do I ask when I am not sure?
Elena emails these to Grace. Grace replies: drafts and summaries are fine. No customer names or addresses, and no staff pay by name. Use the company account, not a personal one. She wants a note on any report drafted with AI. That dated reply is now Elena's policy.
Official guidance makes the same move. The UK government's AI Playbook (2025) gives government staff ten principles. One is “You have meaningful human control at the right stages.” In the US, the National Institute of Standards and Technology (NIST) publishes a voluntary AI Risk Management Framework (2023). It asks organizations to write down who is responsible for what. Same idea: a named person stays in charge.
Confidential data
Confidential data is information that your employer, school, customers or coworkers expect you to keep private. Greenline's job sheet lists customer addresses and overtime by employee name. So Elena pastes a copy with crew totals only: Crew A, Crew B, Crew C. The AI does not need to know who lives where. Module 04 shows how to strip private details step by step.
A missing rule is not permission. It only means nobody has written the rule yet. Five questions now cost far less than explaining a leak later.
Meetings, email and spreadsheets in practice#
Elena pastes her notes from the yard huddle and asks for a summary with action items (who will do what, by when). Which part must she not send out unchecked?
Show the answer
The action items: who will do what, by when. They become promises other people are held to, so each one goes back to the person named first.
Meetings
AI can draft an agenda and summarize notes you provide. Before anyone records a meeting or uses an AI note-taker, ask everyone and follow your workplace's rules. Then check the action items, because summaries firm things up. Elena's summary says the Crew B lead “will train his crew on the app”. He actually said he would “take a look at it”. So she sends the list to the three leads and asks: “Is this right?”
Email and messages
The AI drafts her follow-up text to the three crew leads. Elena rewrites it in her words, checks the start date against Grace's email, and deletes a promise of “full support, any time” nobody offered. When the Crew B lead replies, annoyed, Elena writes back herself: that is a relationship. The AI can still tell her if her reply sounds defensive.
Spreadsheets
AI can explain or write a formula, which is the rule a spreadsheet cell uses to calculate its value. Test it before you trust it by shrinking the problem. Two rows holding 10 and 5 should total 15. If the formula passes that and one real row you check by hand, keep it. Never paste AI-typed totals over cells that hold formulas.
This table is made up. Paste the whole box:
Here is a made-up monthly table for a small landscaping company.
Crew Jobs done Fuel cost Overtime hours
A 42 $1,840 12
B 38 $1,615 20
C 35 $1,502 6
Write a three-sentence summary for a manager. Use only numbers from this
table or totals you calculate from it. Then list every number you used
and show how you got each total.Predict first. Each crew did about 40 jobs, so total jobs should be close to 120. Write your estimate down before you run the prompt.
What to look for: The real totals are 115 jobs (42 + 38 + 35), $4,957 of fuel and 38 overtime hours. A reply far from your estimate, such as 151 jobs, is wrong before you add anything up. Are the reply's totals right? Did any number appear that the table cannot give, such as a change from last month? Did it claim a trend, like “fuel costs are rising”, from one month of data? Each is a line to cut.
Build your own workflow#
On Monday Elena cut the made-up 8%. On Thursday a fresh draft invents an overtime change the same way. What should she have done on Monday?
Show the answer
Written the mistake down and turned it into an instruction she gives every time: “Use only numbers from my material, and say where each came from.” A mistake fixed once is luck. A mistake written down becomes a rule.
A workflow is the steps you follow every time you do a kind of task. A good personal one has four small parts.
1. A mistakes log
A mistakes log is your own running list of how AI went wrong for you, and the rule each mistake taught you.
| Day | What went wrong | The rule |
|---|---|---|
| Mon | Invented “fuel fell 8%” | Only numbers from my material, each with its source |
| Tue | “Will train” for “will take a look” | Action items go back to the people named |
| Wed | Promised support nobody offered | No promises unless I make them |
Early on, the owner asked our AI assistant to keep a running log of its own errors, newest first, with the rule each one produced. It was also to actually apply those rules. Several later entries say a rule was “applied successfully”. Rule: keep your own list of how AI fails you. It is the most useful prompt you own.
2. An instructions note
An instructions note is a short set of rules you paste at the start of a chat, so the AI follows your standards. Some tools let you save one; check yours. Each line of Elena's note came from her log:
About me: office coordinator at a small landscaping company. Readers are
crew members and my manager. Write plainly, briefly and kindly.
1. Use only facts and numbers from material I give you. Label estimates.
2. For every number, say where it came from.
3. Never invent names, dates or promises. Write [CHECK] instead.
4. For decisions, give options, a recommendation with reasons, and what
would change it. I decide.
5. At the end, list anything you are unsure about.3. A checklist
- Every number traced to its source.
- Names, dates and promises match the original.
- No private details in what I pasted or what I send.
- Read once as the reader will.
- Disclosure line added where it matters.
- I opened the actual file or message that will be sent.
4. A review step
A review step is a deliberate pause to check the final version, as the reader will see it. The last checklist line is the one people skip. On Thursday night Elena fixed the overtime line in her document. On Friday she attached the report and nearly clicked send. Then she opened the attachment itself. It was Wednesday's PDF, made before the fix.
All day, our AI assistant told the owner that the day's work was “ready to deploy”, meaning ready to upload to the live website. But the newest upload file was four days old. Uploading it would have published none of that day's work. The owner noticed the old dates himself. Rule: “done” means the thing that actually ships is current. Check the file that goes out, not the intention.
Start a new chat. Paste your own note, or Elena's, then the made-up crew table above, and ask for the same summary.
What to look for: Does every number come with its source? Did [CHECK] markers appear where it could not know something? Any rule it broke is your first log entry. Rewrite that rule more plainly and test again.
Our free kit sets up memory files your AI reads each session. Get the starter prompt.
Final project: one real task, start to finish#
Use the whole routine once, on something real, while it is fresh. Pick one small, low-risk task from your week that your rules allow: a summary of your own notes, an email draft, a study plan. Then:
- Split it. Mark each part “hand over” or “keep”. Can I check it? Who lives with the result?
- Check the rules. Find your policy, or ask. Strip private details before you paste.
- Run it. Paste your instructions note, then ask. For decisions, ask for a recommendation, then choose.
- Verify. Run your checklist, and open the actual thing you will send.
- Disclose and own. Add a disclosure line where it matters, and send it as yours.
- Write down what you checked, using this record:
Task:
What I handed over:
What I kept:
What I checked, against what:
What the AI got wrong (log it):
What I will change in my note:Elena's record is short. Her note's first rule now adds “no comparisons unless I give you both months”. On Friday, Grace reads the report and asks one question. Elena knows the answer, because the work is hers.
Where you go from here
That is the course. You began by asking an AI a question and wondering whether to believe it. You finish with a way of working: ask clearly, check what matters, keep private things private, decide, and stand behind what you send. None of that goes out of date when the tools change.
Thank you for staying to the end. Every module is on the course overview, ready to review or share. This course is free and supported by readers; if it helped, you can support it here.
What was new in this module#
- Delegating to AI is like delegating to a person, with one new part: the AI cannot answer for the work, so you do.
- Two questions sort tasks: can I check it, and who lives with the result?
- Studies found real gains where AI is strong and worse results just past its edge, a “jagged frontier” you cannot see in advance.
- For decisions, ask for options, a recommendation with reasons, and what would change it. Then choose.
- The three duties: disclose, verify, own.
- Follow your AI policy; if there is none, ask the five questions and keep the answers in writing. Keep confidential data out.
- Action items go back to the people named; relationships stay in your words; formulas get tested on a tiny case.
- A personal workflow: a mistakes log, an instructions note, a checklist and a review step.
- “Done” means the thing that actually ships has been checked.
Delegate the work, keep the decisions, and own what ships. AI can draft, summarize and tidy; you decide, you check, and your name answers for the result.
Key terms#
- Delegate
- To hand a piece of work to someone, or something, else while you stay responsible for the whole.
- Jagged frontier
- Researchers' name for the uneven edge of what AI does well: strong on some tasks, weak on others that look just as hard.
- Disclose, verify, own
- The three duties: say how AI was used where it matters, check its output against something real, and take responsibility for the result.
- AI use policy
- Written rules from an employer or school about which AI tools may be used, for what, and with what information.
- Confidential data
- Information that your employer, school, customers or coworkers expect you to keep private.
- Mistakes log
- Your own running list of how AI went wrong for you and the rule each mistake produced.
- Instructions note
- A short, reusable set of rules you paste at the start of a chat so the AI follows your standards.
- Review step
- A deliberate pause to check the final version, as the reader will see it, before it goes out.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. A coworker says: "I used AI for that report, so a wrong number is on the AI." What is wrong with that?
Readers ask the person who sent it, not the tool. Using AI does not move responsibility: the coworker had to verify before sending, and must fix and own the error now.
2. Option one: paste a ten-page handbook in and ask for a summary. Option two: ask "what does a typical handbook say about this?" Which is safer to hand over?
Option one. It summarizes material you provided, so you can check each line against the page. Option two asks the AI to remember, with nothing to check it against, so treat it as a lead at most.
3. The AI recommends option B with sensible reasons. You pick A. Is that a mistake?
Not necessarily: you may know something it does not, and the choice is yours. The test is whether you can finish "I chose A because ___" with a real reason. If the only reason is habit, look at B again.
4. Your class allows AI for brainstorming but not drafting. You brainstormed with AI, then wrote the essay yourself. Should you mention it?
Check the exact rule first; some classes want any use disclosed. If it is unclear, a one-line note or a quick question to your teacher costs little and protects you.
5. You checked your report carefully, saved it as a PDF and sent it. A colleague finds an old number in it. Which step was missing?
Opening the actual file that was sent. The check was done on the document, but an older PDF shipped. "Done" means the thing that goes out has been checked.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can split a task into parts to hand to AI and parts to keep, using "can I check it?" and "who lives with the result?"
- I can ask AI for options and a recommendation with reasons, then make and explain my own choice.
- I can apply disclose, verify and own to my work, including a one-line disclosure.
- I can find my school's or employer's AI rules, or ask the right questions when there are none.
- I can keep a mistakes log and turn it into an instructions note and a checklist.
- I can check the actual file or message that will go out before I call a task done.
Frequently asked questions#
Is it okay to use AI at work?
Often yes, for the right tasks, if your employer allows it. Check your employer's AI policy, or ask your manager which tools are allowed and what information must stay out. Start with drafts, summaries of material you provide and formatting, and check everything before it goes out.
Do I have to tell my boss or teacher that I used AI?
Check your school's or employer's rules first; some require disclosing any use. Where there is no rule, disclose when someone relies on the work. One plain sentence is enough, such as: first draft written with an AI assistant, every number checked against the source.
Who is responsible if AI makes a mistake in my work?
The person who sends the work answers for it, not the tool. In the 2023 US case Mata v. Avianca, lawyers who filed made-up court decisions from a chat assistant were fined. This is general information, not legal advice.
What should I never put into an AI tool at work?
Anything your employer's policy forbids. Common examples are customer details, personal information about coworkers, pay, passwords and anything covered by a contract. When in doubt, leave it out or use made-up placeholders.
What if my workplace has no AI policy?
Do not treat that as permission. Ask your manager which tools you may use and what information must stay out. Ask whether they want to know when work was drafted with AI. Keep the answers in writing.
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942. doi.org/10.1093/qje/qjae044
Dell'Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. R. (2023, revised 2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013; forthcoming in Organization Science. hbs.edu
US District Court, Southern District of New York (2023). Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, June 22, 2023. Court record via Justia
Government Digital Service, UK Government (2025). AI Playbook for the UK Government, 10 February 2025. gov.uk
National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 (see GOVERN 2.1). nvlpubs.nist.gov