AI, Jobs and Developing Economies

Tasks, not jobs

The most useful correction to the public conversation: AI automates tasks, not occupations.

Almost every job is a bundle of tasks. A radiologist reads images, but also consults with clinicians, handles ambiguous cases, explains findings to patients, supervises trainees and takes responsibility for decisions. A model that reads images extremely well automates one item on that list.

The effect is usually a change in the mix — the person spends less time on the automated task and more on the rest. Sometimes that makes the role better, sometimes worse, and sometimes fewer people are needed for the same volume.

The right analysis is therefore not "will AI take this job?" but: which tasks in this role are repetitive, fast judgements with recorded outcomes — and what proportion of the week are they? A role that is 70% such tasks will change profoundly. One that is 10% will barely notice.

What actually determines exposure

More exposed: high-volume repetitive judgement; work already conducted through a screen and logged; tasks with a clear right answer; roles where the output is a document or a classification.

Less exposed: physical work in unstructured environments — a technician diagnosing a fault in a building nobody documented; work whose core is relationship, negotiation or persuasion; work requiring accountability, because someone must be answerable and a model cannot be; genuinely novel problems with no precedent to learn from.

Note that this cuts across the usual status hierarchy. Some highly-paid analytical work is more exposed than skilled manual work, which is a reversal of the pattern of previous automation waves and part of why this one feels different.

An honest position

Anyone who tells you confidently what AI will do to employment is overreaching. What can be said responsibly:

Historically, technology has displaced specific occupations while total employment grew. This has held over long periods.
That transition is genuinely painful for the people displaced, and the aggregate statistic is no comfort to them. Displaced workers frequently do not obtain the new jobs; those go disproportionately to a younger cohort in different places.
The speed matters more than the magnitude. Economies absorb change across decades comfortably and across a few years badly.
Nobody knows whether this wave behaves like previous ones. Arguments that it is different — the breadth of tasks affected, the pace — are serious and not obviously wrong.

For an organisation, the actionable part is narrower and clearer. If you automate work people currently do, you will make decisions about those people. Making them thoughtfully and early is both decent and practical: the staff who understand a process best are the ones whose knowledge you need to automate it well, and they will not help you if they believe the outcome is their redundancy and nobody has said otherwise.

Developing economies

AI adoption does not follow the same path everywhere, and if your organisation operates in or sells to developing markets this matters commercially.

Leapfrogging is real. Just as many countries moved to mobile phones without building landline networks, some are adopting AI-enabled services without an intermediate layer of legacy enterprise software. There is nothing to migrate from, which can make adoption faster rather than slower.

The constraints are different. Connectivity, device capability, electricity reliability and the cost of cloud compute shape what is feasible. A system assuming an always-connected high-end device excludes most of the market.

Data is thinner, and biased toward elsewhere. Models trained predominantly on data from wealthy countries perform worse elsewhere — in languages, in accents, in what "normal" looks like for a transaction, in agricultural conditions, in disease presentation. This is the Gender Shades problem at the scale of countries.

The opportunity concentrates where formal infrastructure is absent. Some of the most consequential deployments address gaps rather than optimising existing systems: agricultural advice where extension services are thin, medical triage where specialists are scarce, credit assessment where formal credit histories do not exist.

The risk of dependency is real. If the models, the infrastructure and the expertise all sit elsewhere, value accrues elsewhere. This is why the question "what data do we have that others do not" from module 3 is a national and regional question as much as a corporate one.

Suppose you are automating a task that occupies three people for much of their week. Here is what actually works, and it is mostly about sequence and honesty.

Say what you know, when you know it. The worst option is silence, because the rumour will be worse than the truth and will arrive first. If you do not yet know the headcount implications, say exactly that — "we do not yet know, and here is when we will".

Involve them in building it. They know the exception cases, the informal rules, the reasons the obvious approach fails. Excluding them produces a worse system and guarantees resistance to it. This is not manipulation; their knowledge is genuinely the scarce input.

Be honest about direction. If the intention is that fewer people do this work, say so. Reassurance that later proves false destroys credibility for every future change. If the intention is to redeploy rather than reduce, say what to and when.

Name what the role becomes. "You will handle the exceptions" sounds like a demotion unless you make clear that the exceptions are the difficult, judgement-heavy cases — which they are, since the routine ones are what got automated. Adjust titles, grading and expectations to match, or nobody will believe it.

Watch the skill trap. If people only ever see the hard cases, they lose the routine practice that built their judgement, and new joiners never acquire it at all. This is a real, documented pattern in automated environments. Plan for how expertise gets maintained when the easy cases stop reaching humans.

Give them a route to challenge the system. The people supervising a model are the first to see it failing. If reporting that is inconvenient or unwelcome, you lose your best monitoring — and you were relying on it, whether or not you wrote that down.

❓ Knowledge Check

What is the most accurate framing of AI's effect on employment within an organisation?

📚 Flashcards1 / 6

Term

Tasks, not occupations

Click to flip

Definition

AI automates tasks within jobs rather than whole jobs. The useful question is what proportion of a role consists of repetitive, fast judgements with recorded outcomes.

Click to flip back

💡Key Takeaway

AI automates tasks, not occupations, so assess roles by what share of the week is repetitive, fast, recorded judgement — and note that exposure cuts across the usual status hierarchy. Be honest with affected staff early, because the people who understand a process best are exactly the ones whose knowledge you need to automate it well. In developing markets, expect leapfrogging, different infrastructure constraints, and models trained mostly on data from elsewhere performing worse locally.