Don’t Predict the Vanishing Job—Map the Task Stack

Slug: map-task-stack-ai-work
Tags: Future of Work, AI Jobs, Future Skills
Meta description: Stop treating occupations as single units. Map the task stack to decide what AI may assist, what humans should retain and which skills require investment.

Featured image: AI-generated conceptual illustration.

Headlines often ask which jobs artificial intelligence will replace. Employers and workers need a more useful question: which tasks inside a job are changing, and what happens to the work that remains?

A job title compresses many activities into two or three words. A project manager may plan schedules, interpret risk, negotiate priorities, reassure clients, write reports and notice when a team is quietly struggling. Treating that bundle as one unit makes both optimism and panic too easy.

Exposure is not the same as disappearance

The International Labour Organization’s May 2025 working paper Generative AI and Jobs: A Refined Global Index of Occupational Exposure uses task-level data, worker input, expert discussion and model-assisted scoring to estimate occupational exposure. The authors report that one in four workers globally is in an occupation with some exposure to generative AI, while 3.3% of global employment falls into the highest exposure category.

Those figures describe potential exposure under the study’s framework, not measured job losses. The paper explicitly argues that job transformation is the likeliest effect because most occupations include tasks requiring human input. That distinction should change workforce planning.

A role can be highly exposed because several tasks are technically susceptible to automation or assistance, yet still require people for judgement, responsibility, physical work, relationships or handling exceptions. Conversely, even a modestly exposed role may change sharply if one task controls the economics of the whole job.

Build the task stack

Choose one role and list what a competent person actually does during a typical week. Do not rely only on the formal job description. Include invisible work: checking assumptions, repairing misunderstandings, obtaining consent, mentoring colleagues and deciding when the standard process does not fit.

Group the tasks into five practical layers:

• Information tasks: finding, sorting, summarising and formatting material.
• Production tasks: drafting, calculating, coding, designing or preparing outputs.
• Judgement tasks: interpreting ambiguity, weighing trade-offs and choosing a course of action.
• Relational tasks: persuading, supporting, negotiating, teaching and building trust.
• Accountability tasks: approving consequential decisions, explaining them and repairing harm.

These layers overlap. The point is not to produce a perfect taxonomy; it is to stop discussing the occupation as though every activity will change at the same speed.

Assess assistance before automation

For each task, ask whether AI could support preparation, produce a first draft, check consistency or surface options. Then separately ask whether it should complete the task without review.

This separation matters. A system might draft an interview question but should not silently redefine the hiring criteria. It may summarise customer feedback but miss minority experiences hidden by aggregation. It may prepare a clinical note while remaining unsuitable to determine the treatment plan.

Rate each proposed use against evidence quality, error cost, privacy, reversibility and the availability of a qualified reviewer. A technically possible automation can still be a poor organisational decision.

Redesign the workflow, not only the task

A faster draft does not automatically create a better job. If management responds by doubling workload, removing reflection time and keeping the same accountability, employees may inherit more risk rather than more capability.

Measure the whole workflow. Did the tool reduce rework? Did exception rates rise? Did junior staff lose opportunities to learn? Did customers receive clearer explanations? Did experienced workers spend more time on meaningful judgement—or merely review a larger volume of machine output?

Pilot changes with the people doing the work and those affected by its outcomes. Record where they override the system and why. Overrides are not necessarily evidence that employees resist innovation; they may reveal missing context or a badly defined objective.

Protect the learning ladder

Many professions develop judgement through apparently routine tasks. A junior employee who researches, drafts and receives feedback is not merely producing text; they are learning how the field reasons.

If AI performs every entry-level component, organisations may weaken the pathway by which future experts are formed. Deliberately preserve supervised practice, require workers to explain important outputs and rotate people through tasks that build domain understanding.

Training should follow the redesigned work. Generic prompt instruction is insufficient if the role requires source evaluation, data protection, stakeholder communication or responsibility for decisions. Teach the human capabilities that become more important when production becomes easier.

Watch unequal effects

The ILO study reports meaningful differences by occupation, income level and gender. It found higher shares of women than men in the highest exposure category globally, with the gap wider in high-income countries. The authors connect this pattern partly to the concentration of women in clerical occupations.

Exposure does not predetermine harm, but it identifies where employers, worker representatives and policymakers should examine transition risks. Ask who receives training, who gains higher-value responsibilities, whose work becomes more monitored and who bears the consequences of system errors.

A one-page task-stack plan

For every priority role, document the task, current owner, proposed AI assistance, evidence required, reviewer, escalation route and success measure. Add a learning requirement: what must the worker continue practising even if the tool can generate a plausible result?

Revisit the map after a defined trial. If the workflow changes, the control structure, performance targets and job description may need to change too. Do not let responsibilities drift informally while official accountability remains frozen.

The better future-of-work question

The future of work will not be decided only by what models can do. It will also be decided by how institutions divide tasks, distribute gains, protect learning and preserve human authority where consequences matter.

So replace “Will AI take this job?” with four questions: Which tasks are changing? Which capabilities become more valuable? Who benefits from the redesign? Who remains answerable for the result?

That task-level view is less dramatic than a list of doomed occupations. It is also far more useful for building work people can actually trust.