AI changes the tasks inside a job, but it does not erase the job itself. Five skills keep you useful: defining the question, checking the numbers against a source your finance team has approved, talking to the person who must live with the answer, judging whether a draft is any good, and knowing when to stop and not paste.
Say it is Tuesday morning, and your big quarterly review with your bosses is on Thursday. You open the shared folder and find a file the new intern made with ChatGPT: six pages that define the key numbers the review will use. Page 1 is a pretty table. It says customers stay and spend 118% of what they spent last year, 4,260 accounts are active, and new customers get value in 2.1 days. (These numbers are made up for this example.) It looks finished, and that is the trouble.
Tasks move first, titles lag
The intern still has the intern title. The task that used to take a week of interviews, which was writing down what “active” means, took 11 minutes in a chat. That is how change shows up at work. Nobody’s job vanished on Tuesday. The definition work moved into a paste, and nobody owned the paste.
The International Labour Organization (ILO) has been useful here precisely because it is cautious. In 2023 it scored the tasks inside occupations, and in 2025 it refined that index. Few jobs are a pile of tasks you could fully hand to today’s generative AI, and nearly every occupation still has work that needs a person. The likely path is that jobs get transformed. Those scores describe potential exposure if the tools were fully adopted, so they are not a count of roles already gone. Cost, skills, and messy real-world operations all sit between the score and the headcount.
The Organisation for Economic Co-operation and Development (OECD), a research group funded by about 38 mostly wealthy countries, put the same caution in a chapter title of its 2023 Employment Outlook: no signs of slowing labour demand yet. Its reading of the evidence at the time was that AI had caused little sign of large negative effects on employment. Workplace case studies in manufacturing and finance pointed the same way, because tasks were reorganised more often than people were walked out. Your role can still change, so do not plan your week around a displacement percentage you saw on a thumbnail. You cannot take that thumbnail to the finance lead. You can take a definition, a source, and a conversation.
Rule of thumb: If Finance did not approve the definition, the model’s definition is only a draft, and a draft does not go in the deck.
Five skills that still matter
The National Institute of Standards and Technology (NIST) publishes an AI Risk Management Framework, which is voluntary and dates from 2023, with a generative-AI profile added in 2024. It spends a lot of space on human oversight, meaning who is responsible when a person and a model share a workflow. OpenAI’s own API (the way one program asks another for data) safety note says the same thing in one line you could tape above a monitor: wherever possible, have a person review outputs before they are used in practice. That person needs the original notes, the source table, and the memo in front of them. Reading the chat window does not count as a review.
On an analytics team, that oversight comes down to five habits you can name in a standup.
Define the question
The intern asked ChatGPT to write the KPI (key performance indicator, a number tracked to judge progress) definitions. That is a request to produce text, and it has no question inside it. The review had a real question: which agreed set of numbers does this company use this quarter for Net Retention, Active Accounts, and Time to First Value, and who has already approved it? A model will answer whatever prompt it is given. It will not walk down the hall and notice that you asked the wrong thing.
Write the question in one sentence before you open the chat. For example: which Net Retention does the finance lead want on slide 4, counted by customer logos or by annual recurring revenue (ARR), and with services in or out? If you cannot finish that sentence, you are not ready to prompt. The prompting series teaches how to turn a sticky note into a clear work request. This skill comes one step earlier, because you first have to pick the decision the meeting is actually making.
Check the numbers
The intern’s table looked arithmetically tidy, with three metrics, two decimal places, and a footnote that said “trailing twelve months.” None of that is a check. A real check means opening finance_metric_pack_q2.pdf, the three-page approved definitions file the finance lead sent last quarter, and recomputing one number against the warehouse. The gap between 118% and 97% is a fight over definitions. One figure includes professional services and uses ARR, while the other counts logos and strips out one-time services. The model cannot see that fight unless you put both memos in the prompt on purpose, and even then it will happily average them.
Pick one number and find its source. Then do the arithmetic once yourself, with a pencil or a SQL (the standard language for asking a database for data) query window. Do not ask the same model whether 118 looks right, because that question is how 118 survives until Thursday.
Talk to the human who has to live with it
The finance lead has to sit in the review and defend Net Retention if a board member pokes at it. Product has to defend Time to First Value if a customer success lead says onboarding takes “two days now.” Those people are not in the ChatGPT sidebar. Twelve minutes at the finance lead’s desk, or a six-line Slack message asking her to approve three definitions, beats an 11-slide surprise.
This is also where agents fool people. A coding agent or a work-mode agent (AI tools that do multi-step work on their own) can file a code change, fill a sheet, or draft 11 slides while you get coffee. The watch, approve, walk away ladder still ends with a person. If nobody talks to the owner of the number, you have handed off the embarrassment along with the work.
Taste, and the courage to stop
Taste here means noticing that Time to First Value came back as 2.100 days, that the prose sounds like a glossary page, and that “active” was defined three different ways in one table. Fake precision is a warning sign. So is a Wikipedia tone, and so is a metric that nobody in the building has ever said out loud.
Stopping is the harder half. The deck already had 11 slides built on 118%, and pulling them on Tuesday evening feels late, but shipping them on Thursday is later. The closing post of the prompting series covers closing the chat and doing the task yourself. This is the same muscle applied to a different object: close the folder, and do not paste.
What the model can draft and what you still own
Use this as a wall chart. The left column is the skill. The middle column is fair work for ChatGPT, Claude, Gemini, or Grok. The right column stays yours even when the draft looks done.
| Skill | What the model can draft | What you still own |
|---|---|---|
| Define the question | A rewrite of your prompt, a list of possible questions, three ways to frame the QBR | Which question this meeting is answering, and who the answer is for |
| Check the numbers | A formula, a sample table, a “spot the inconsistency” pass on text you pasted | The source system, the approved definition, the arithmetic you would defend |
| Talk to the human | A six-line Slack, a meeting agenda, a summary of last quarter’s thread | The actual conversation with Finance, Product, Legal, or the intern |
| Taste | Three shorter versions, a less glossary-like tone, a cut of the extra decimals | Whether this artifact should exist, for this audience, in this week |
| Stop | A list of risks if you ask it to list risks | The decision to pull the 11 slides, or to leave the chat unpasted |
Drafting is cheap, while ownership is a person with a calendar and a name. If your weekly AI habit stops at the middle column, you will only get faster at producing files like the intern’s six pages of definitions.
Worked example: three KPIs, two days out
Here is the mismatch laid out as a closed case. The numbers are this imaginary team’s toy figures, not industry research.
| Metric | ChatGPT paste (the intern) | Approved definitions (Finance / Product) |
|---|---|---|
| Net Retention | 118%, ARR, services included, trailing twelve months | 97%, logo count, one-time services stripped |
| Active Accounts | 4,260, any login in 90 days | 2,941, paid seat with a billable event in the last 30 days |
| Time to First Value | 2.1 days, first login after invite | 11 days, first successful export |
The intern’s chat title was still in the sidebar: “kpi defs please be thorough.” Thorough is how you end up with six pages. The model mentioned the product name in three sentences, invented a “North Star” paragraph, and formatted a table that would have looked fine in a vendor blog. It never opened the warehouse and never read finance_metric_pack_q2.pdf. It also did not know that Product counts first value as an export, because nobody types that unless they already know it.
Now picture the version where you catch it. You open Finance’s Slack thread before you open the intern’s file, and you find a 14-line message that is ugly but correct. You walk to the finance lead’s desk, and 18 minutes later you both agree to pull the 11 slides that evening. Thursday still happens, and the deck uses 97%, 2,941, and 11 days. The intern’s chat stays in history as a practice artifact and not a verdict on anyone’s career.
If you are the intern, send the draft to the owner labeled as a draft. Do not drop it in the review folder under a filename that pretends to be the official book. If you are the manager, opening the intern’s file first is already late. Put the approved PDF in the folder on Friday, with the three definitions at the top, before anyone writes a prompt.
A 45-minute weekly practice
Do not spend the 45 minutes on a talk about skills of the future. Pick one live file instead: a deck, a metric, an email, or an agent output. If your week is quiet, reuse last week’s leftover, the way you might reuse the intern’s chat the following Monday.
Spend the first five minutes writing the question in one sentence and naming who has to live with the answer. If the sentence will not finish, go find that person and come back. Spend the next fifteen opening one model you already use, such as ChatGPT, Claude, Gemini, or Grok. Paste the question first, then ask for a tight draft of six lines, a three-row table, or three options, and do not ask it to confirm what Finance thinks or to “be thorough” unless you want six pages. Spend the next fifteen opening the source, recomputing one number or checking one claim against a file a human already approved, and then messaging or walking to the owner. Use the last ten to write down stop, edit, or ship, where ship means you would put your name on it in the review. Checking agent work is the same last block, used when the artifact came from an agent’s loop and not from a single chat.
Paste this into a note and fill it in every week. The time stays at 45 minutes and the file changes.
Weekly 45-min AI practice (one artifact)
Date:
Artifact (deck / metric / email / agent output):
Owner who has to live with this:
Question in one sentence:
Model used (ChatGPT / Claude / Gemini / Grok):
What the model drafted (six lines max in this box):
Source I opened (file, query, or memo):
Number or claim I recomputed:
Person I talked to, and their yes/no:
Stop / edit / ship:
What I will not let the model own next time:
Chat title to keep (or to ignore):The checklist forces the right-hand column of the table before you close the laptop. A filled row for the quarterly review (QBR, short for quarterly business review) story would read like this: artifact is the QBR deck, owner is the finance lead, question is which Net Retention is the approved one, model is ChatGPT, and source is finance_metric_pack_q2.pdf. The recomputed number is 97%, not 118%, the person said yes, and the decision is stop, which means pulling 11 slides. The last line, “what I will not let the model own,” would be “the definition.”
Common mistakes
- Asking a model to write the definitions before anyone wrote the question.
- Pasting the reply into the review folder with a filename that looks official.
- Checking 118% by asking the same chat whether it looks right.
- Skipping the finance lead because the table was pretty and the intern was thorough.
- Spending the week on future-of-work threads instead of one live artifact.
- Letting an agent fill 11 slides while you only watched, then calling that an approval.
- Keeping 2.100 days because extra decimals felt precise.
How to practice this week
Run the 45-minute plan once on a file that already exists, and fill every line of the checklist, including stop, edit, or ship. If you manage an intern, put the approved PDF in the shared folder on Friday and require a named owner on any metric that will be spoken aloud in a meeting. If the weak step is the prompt, read Prompting for everyone next. If the weak step is an agent you walked away from too early, read AI agents for everyone. This series ends here, and the rest of the catalog is on Learn.
Quick recap
- Plan the week around tasks and definitions, not around a viral jobs-gone percentage.
- Keep five skills: question, number check, conversation, taste, and stopping.
- Models draft, and you own the approved definitions and the decision to pull a deck.
- 118% against 97% two days before a review is the failure to watch for, so pull the slides.
- Run one 45-minute practice on one artifact, with stop, edit, or ship written down. so each practice ends with a decision you can learn from
Series notes
This is Part 5 of AI safety for everyday life. Previous: Kids, family, and shared devices. Related: Learn, Prompting for everyone, AI agents for everyone.
Sources
- ILO: How might generative AI impact different occupations? (task scores inside jobs; transformation more likely than a clean wipe; exposure is not actual impact)
- OECD Employment Outlook 2023: Artificial intelligence and jobs (little evidence, at the time of that edition, of large negative employment effects; task reorganisation showed up first)
- NIST: AI Risk Management Framework (voluntary; human oversight and roles when people and models share a workflow)
- OpenAI: Safety best practices, Human in the loop (review outputs before they are used in practice; give the reviewer the original notes)
- Analytics Made Simple: Learn (paths after Phase P)
- Analytics Made Simple: Prompting for everyone and Analytics Made Simple: AI agents for everyone (the typing and the watch-approve ladder this post sits on)
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
