AI Guides › Workbench
By Nigel Guy · 6 min read
Most people ask "will AI take my job?" as if a job were one object. It is not. It is a bundle of tasks, and the useful question is which tasks in your bundle a chatbot can already do, which it merely speeds up, and which still need you. The headline question feels sensible and teaches you nothing, which is why it generates so much worry and so few decisions.
The rule: audit tasks, never the job title, and sort each task into one of three buckets (Hand over, Speed up, Keep) before you touch any tool.
Three free or cheap things do the work. You do not need to buy anything to start.
| Tool | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| O*NET OnLine | US government database listing task statements for 900+ occupations | Free | A neutral starting list of tasks for your role | US-built, so UK job titles and UK-specific duties may not map; it will miss your own quirks |
| A general chatbot (Claude, ChatGPT or similar) running the audit prompt below | Drafts your task list and a first-pass sort | Claude Free is $0; Claude Pro is $20 a month (about £16 at time of writing, check the £ price at checkout) | Turning a rough job description into a sorted table | Its sort is a hypothesis, not a finding |
| Your own one-week task log | Records what you actually did | Free | Catching the work you forgot to list | Takes a week; most people skip it |
| Bucket | Meaning | Test |
|---|---|---|
| Hand over | AI produces a result you would send with a quick check | A competent colleague could verify it in minutes |
| Speed up | AI does a first pass; you steer, correct and finish | You would not sign it off unedited |
| Keep | The value is your judgement, relationships, accountability or physical presence | A wrong answer costs trust, money or safety, or the task needs you in the room |
Anthropic's research on how people use Claude, which mapped around a million conversations to O*NET tasks, found a slight lean towards augmentation (57%) over automation (43%). Treat that as context for the middle bucket being large, not as a forecast for your job: it describes how Claude was being used, not what any employer will do.
Paste this into a chatbot. Fill in the square-bracket placeholders. Do not paste client names, personal data or anything your employer treats as confidential.
You are a careful workplace analyst helping me understand which parts of my job current AI tools can handle. You are not predicting the future of my profession.
Context
- My job title: [JOB_TITLE]
- Sector and country: [SECTOR_AND_COUNTRY]
- Seniority and team size: [SENIORITY_AND_TEAM]
- My task list with rough weekly hours: [TASK_LIST_WITH_HOURS]
- Tools and software I already use: [CURRENT_TOOLS]
- Things that go wrong most often in this role: [COMMON_FAILURE_POINTS]
Goal
Sort every task into exactly one bucket and explain why, so I can choose what to try first.
- HAND OVER: an AI tool could produce something I would send after a short check.
- SPEED UP: AI can draft or prepare, but I must steer and finish.
- KEEP: the value is my judgement, relationships, accountability or physical presence.
Steps
1. If my task list is empty, thin or vague, ask me up to five questions before doing anything else. Do not invent tasks.
2. Put each task in a table with columns: Task | Weekly hours | Bucket | Reason in one sentence | What the AI would actually do | What I would still check.
3. Add any tasks that a person in this role usually does but I left out, marked "SUGGESTED, CONFIRM", and do not count them in totals.
4. Give hours per bucket and the three tasks where a trial would save me the most time.
5. For each of those three, name the main way the AI output could be wrong or risky.
Constraints
- Judge what tools can do now, not what they might do later.
- Do not give salary, job-loss or career predictions.
- If a task involves personal, legal, medical, financial or safety-critical decisions, default to KEEP or SPEED UP and say why.
- Say "unsure" where you are unsure. Do not guess.
Self-check before answering: every task has one bucket, the hours add up to my totals, and no reason is generic enough to fit any job.
The table is a first draft. Take each Hand over task and ask three questions:
Then run one real trial per task, using your own material, and compare the result with what you would have done. Only after the trial does a task earn the Hand over label. Use this second prompt for the trial:
You are my working assistant for one task I do regularly: [TASK_NAME].
Context: [BACKGROUND_AND_AUDIENCE]. Here is the input I normally start from: [INPUT_MATERIAL]. Here is an example of a result I was happy with: [GOOD_EXAMPLE].
Produce the result in this format: [DESIRED_FORMAT]. Match the tone and level of detail of my example.
Rules: ask me for anything missing instead of filling gaps with assumptions; mark any fact you could not take from my input as "UNVERIFIED"; do not invent figures, names or sources.
Before replying, check that every claim traces back to my input and the format matches.
Make a small table with the task, its bucket, the date you trialled it, what you had to fix, and the minutes saved. Revisit it every few months; tools change, and a task that failed in the spring may pass now.
Pick tasks that are frequent, low-stakes and easy to check: first drafts, meeting summaries, reformatting, tidying data. Leave anything where an error reaches a customer or regulator for later, and run it with a human review step.