Key points
- A job is a bundle of tasks, and AI changes the tasks one at a time.
- Repetitive, rule-based work with text changes first, especially where a mistake is cheap to catch and fix.
- Work that needs physical presence, trust, accountability or judgement about people changes later.
- Junior roles are most exposed because their tasks are the easiest to hand over, which makes learning the trade harder.
- Checking AI output is a skill, and it depends on knowing the subject yourself.
1. Start with tasks, not jobs
Most arguments about AI and work ask which jobs will disappear. That question is too large to answer well. A more useful one is which tasks inside a job a machine can now do, and what happens to the person when it does.
This article reasons from the nature of tasks. It uses no statistics and makes no numbered forecasts, because nobody can honestly promise those. Where things are uncertain, it says so. The examples come from ordinary workplaces in Pakistan: offices, shops, schools, clinics and freelance work.
2. A job is a bundle of tasks
Consider an accounts clerk in a small trading firm. In one week he enters invoices, matches them against bank statements, chases a late payment by phone, prepares a summary for the owner, deals with the tax consultant and calms a supplier who was paid the wrong amount. That is one job title and at least six different tasks.
These tasks are not equally easy to hand to software. Entering invoices from a clear format is routine. Calming an angry supplier is not. When people say "AI will take the accountant's job", they are treating the bundle as one thing. In practice some tasks in the bundle shrink, some stay the same, and some become more important.
What happens to the job depends on how the bundle is put back together. If the routine tasks were most of the job, fewer people may be needed. If they were the dull part that kept the clerk from more valuable work, the job may improve. Both outcomes will occur in different workplaces, and it is hard to say in advance which will be more common.
3. What changes first: repetitive, rule-based work with text
Language models work with text. They are strongest when the input is text, the output is text, and the task follows a pattern seen many times before. So the first tasks to change are those where someone reads something and writes something according to a fairly fixed rule.
Offices are full of these tasks. Drafting a routine letter, summarising a long email chain, turning meeting notes into minutes, translating a notice between English and Urdu, filling a standard form from a document, and answering the same ten customer questions are all examples. A shop that sells online has its own versions: writing product descriptions, replying to "is this available" and "what is the delivery charge", and preparing order confirmations.
One more condition matters a great deal. A task moves early when a mistake is cheap to catch and cheap to fix. A clumsy sentence in a product description costs little, and anyone can spot it. This is why first drafts change sooner than final decisions.
- The input and output are mostly text or numbers in a known format.
- The task repeats often and follows a rule or template.
- A good result is easy to recognise.
- An error is cheap to notice and cheap to correct.
4. What changes later: presence, trust, accountability and judgement about people
Some tasks resist for reasons that have little to do with how clever the software is. The first is physical presence. An electrician rewiring a house in an old mohalla, a nurse giving an injection, a tailor taking measurements and a mechanic listening to an engine all work with their hands in an untidy physical world. Software that writes well does not do any of that.
The second is trust. A shopkeeper who gives a regular customer goods on credit until the end of the month is relying on years of knowing that family. A parent choosing a school wants to look the principal in the eye. People often want a person because a person can be known, and that preference changes slowly.
The third is accountability. Someone must sign the audit, the prescription, the legal notice and the structural drawing, and answer for it if it is wrong. A tool cannot be held responsible in that way. As long as law and custom require a responsible human, that human's role remains, even if a tool prepared most of the paperwork.
The fourth is judgement about people. A teacher noticing that a quiet child has stopped trying, a manager deciding which of two employees is ready for more responsibility, and a doctor sensing that a patient is not saying everything are all reading people, in person, with incomplete information. This kind of judgement is hard to write down as rules, so it is hard to hand over.
"Later" does not mean "never". It means the obstacle is not only technical, so better software alone does not remove it.
5. The same split, workplace by workplace
In a school, preparing a worksheet, drafting a test paper, writing report-card comments and building a timetable are text tasks that tools can speed up. Managing a class of forty children, earning their respect and speaking to a worried parent are not. A teacher who uses tools for the first group has more energy for the second.
In a clinic, the reception desk has routine text work: appointment messages, reminders, typing up notes and preparing bills. Examining the patient, deciding the treatment and taking responsibility for it remain with the doctor. Even a patient who has read a great deal online usually still wants a doctor to confirm it and to answer for it.
In a shop, replying to common messages, writing listings and keeping a simple stock record can be assisted. Choosing what to stock for the neighbourhood, bargaining with the wholesaler and deciding who gets credit stay with the owner.
Freelancing shows the change most sharply, because much freelance work is delivered entirely as text or files. Basic article writing, simple translation, data entry and plain transcription are exactly the tasks described in section 3, and clients can now do much of this themselves. Freelancers who understand the client's business, take responsibility for the result and handle the difficult cases are in a different position from those who sold only the typing.
6. What happens to junior roles
This is the hardest part of the picture. In many professions, the tasks that are easiest to automate are the ones traditionally given to beginners. The junior at a law office drafts routine notices. The trainee at an accounting firm does the entries. The new developer writes the simple functions and fixes the small bugs.
Those tasks were never only about output. They were how a beginner learned the trade. Someone becomes a senior who can spot a wrong figure at a glance by first doing thousands of entries. If a tool now does the entries, a firm has less reason to hire a trainee, and the trainee who is hired gets less of the practice that used to build judgement.
How this resolves is uncertain. One possibility is that firms hire fewer juniors and later find they have a shortage of experienced people. Another is that training changes shape, with juniors starting on review and harder problems under closer supervision. The practical lesson for a beginner is clear either way: do not let the tool do all your practice. Do the work by hand often enough to understand it, even when a tool could do it faster.
7. Why checking AI output becomes a skill
A language model produces fluent text whether or not the content is right. It can give a wrong date, an invented reference or a calculation that looks tidy and is incorrect, all in the same confident tone. Fluency used to be a rough sign of competence in a human writer. With these tools it is not.
So when a tool produces the draft, the human's work moves from producing to checking. Checking is harder than it sounds. To find an error in a contract clause you must know contract law. To notice that a summary left out the one point that mattered you must have understood the original. This is why subject knowledge becomes more valuable, not less, for the person who remains responsible.
There is also a human weakness to guard against. After a tool has been right twenty times, people stop reading carefully, and the twenty-first answer goes through unchecked. Good checking is a habit as well as a skill.
- Verify every name, date, figure and reference against the original source.
- Redo a sample of the calculations yourself.
- Ask what is missing, not only what is wrong.
- Read as the person who will receive it: would it mislead them?
- Remember that your name is on the result, not the tool's.
8. What is honestly uncertain
Several things cannot be known from reasoning alone. The pace is one. A task being possible for a machine does not mean workplaces adopt it quickly. A small office in Pakistan faces real frictions: the cost of tools, unreliable power and internet, records kept on paper, documents in mixed Urdu and English, and owners who reasonably distrust what they cannot check. These slow adoption, but they do not prevent it.
The second unknown is what new work appears. When a task becomes cheap, people often want more of it, or want things that were not worth doing before. A shop that could never afford a catalogue may now have one, and someone must manage it. Whether such new tasks make up for the ones that shrink is not something anyone can state with confidence.
The third is how far the tools will improve at the harder tasks. The boundary described in this article is where things appear to stand now. It will move, and probably unevenly.
9. What a student or worker can do this year
None of this calls for panic or for ignoring the subject. It calls for a few steady habits. The first is to look at your own work as a bundle. Write down the tasks you do in a week and mark which are routine text work and which need presence, trust, accountability or judgement. That list tells you where to use tools and where to invest in yourself.
The second is to become the person in your workplace who uses these tools well and checks them properly. That person is useful in almost any office. For those who want to go further and build automations and agents themselves, structured courses exist, including the 4-month bootcamp that TIDES runs in Lahore, but the habits below need no course.
- List your weekly tasks and sort them using the tests in sections 3 and 4.
- Pick one routine task and learn to do it with an AI tool, checking every result.
- Deepen your knowledge of one subject, because checking depends on it.
- Practise the human parts on purpose: explaining clearly, listening, negotiating, taking responsibility.
- If you are a beginner, do the basic work by hand until you understand it, then use the tool.
- Keep a record of small real things you have built or improved. It speaks louder than a claim on a CV.
The safest position is not to avoid the tools or to depend on them, but to be the one who can tell when they are wrong.
