The Task-Level Reality of AI and Work
Consider a bank clerk whose invoice workflow has changed. She still processes invoices, but the rhythm is different: the machine extracts most of the data, and she spends her time catching the edge cases—the handwritten supplier numbers, the mismatched tax codes, the invoices that arrived in a language the model was not trained on. Her job title is the same. Her day is not.
This is the pattern the workforce debate keeps missing. We talk about jobs vanishing or staying, about mass displacement or no impact, as though the unit of analysis is the entire role. It isn’t. The useful evidence points in a quieter, more granular direction: skills shift before jobs disappear, and retraining starts with the task, not the title.
The job-replacement narrative is a crutch
The headline is easier than the analysis. “AI will replace jobs” fits in a tweet and drives clicks, but it flattens a distributed, uneven process into a single claim. O*NET’s occupation model is useful precisely because it describes work in terms of tasks, work activities, knowledge, and skills rather than treating a job title as one indivisible block. A radiologist does not become unemployed because a model reads scans. The radiologist’s work shifts toward case review, patient communication, and the ambiguous images the model flags as uncertain. The job changes, then it stays.
The ILO’s global analysis of generative AI and jobs reinforces this. Effects are not uniform across sectors, regions, or even roles within the same company. The question is not “will this job exist in five years” but “which tasks in this job will a model do, and which ones will the person still need to own.”
What the task-level evidence actually shows
O*NET and ILO evidence both point toward a pattern that is neither utopian nor catastrophic. Occupations are bundles of tasks, and those tasks differ in how exposed they are to automation or augmentation. The distribution matters more than the average. A clerical role with heavy data entry looks different from a management role with heavy judgment and negotiation, even though both are “office jobs.”
The evidence is thin on exact percentages across industries—that varies by firm, by data quality, by regulatory environment. But the direction is consistent: exposure is partial, task-specific, and shaped by how organizations choose to deploy the technology. The same model that automates invoice extraction in one company becomes a decision-support tool in another.
A concrete classroom example
Consider a curriculum for accounting clerks. A traditional program spends weeks on manual data entry, ledger matching, and transaction coding. If models now handle most extraction and coding reliably, the curriculum needs to change—not by removing the accounting clerk role, but by shrinking the data-entry module and expanding the exception-handling one.
The new skills are not about the model. They are about what happens when the model is uncertain: verifying a flagged entry, resolving a discrepancy between two systems, communicating with a supplier whose invoice format broke the pipeline. The job title stays. The task composition shifts. The training must follow.
Why skills frameworks fail without granularity
Most workforce skills frameworks describe broad competencies: “data analysis,” “communication,” “critical thinking.” These are useful for résumé filters but useless for curriculum design. A task-level analysis breaks “data analysis” into concrete operations: extract, validate, reconcile, flag, escalate. Each has a different exposure to automation. Validation and escalation are harder to automate than extraction. If the framework treats them as one category, the training misses the shift.
The O*NET framing suggests that policy and education need to move toward this granularity. Not “retrain workers for AI” but “identify which specific tasks in a role are changing, and build targeted modules for the tasks that remain human-owned.”
What policy and training should do differently
First, stop asking whether AI will replace jobs. The question is not answerable at that level. Instead, ask: which tasks in which roles are most exposed, and which tasks are most resilient? That is an empirical question, and task-level occupational data is the evidence base that makes it answerable.
Second, redesign training around task composition, not job titles. A module on “exception handling for automated workflows” serves a data entry clerk, a claims processor, and a logistics coordinator, even though their job titles are different. The skill is the same. The context varies.
Third, fund longitudinal studies of task change, not one-off displacement estimates. The ILO’s work shows that effects vary by country, by sector, by firm size. A global number hides the local reality. Track the same roles over three to five years and watch how the task mix evolves.
The practical worry is not only whether the bank-clerk role survives. It is whether training catches up with the work people now actually do. That is the gap the evidence points to, and it is the one worth closing.
