What Happens to the World's 800 Million Clerical Workers When AI Arrives

The 9.0/10 score explained

When WorldJobsData assigns a 9.0/10 AI exposure score to general and keyboard clerks (ISCO 41), it is not a forecast about what AI will be able to do in the future. It is a description of what AI can already do right now. The tasks that define clerical work - data entry, document classification, appointment scheduling, routing correspondence, form processing, and generating standard written communications - are exactly the task categories where current large language models perform at or above human level.

This is not an abstract capability claim. GPT-4 class models can process and classify documents at speeds no human can match, at a marginal cost close to zero. They can draft and respond to standard business correspondence. They can extract structured data from unstructured text. They can route customer queries to the right department faster and more consistently than a human operator reading the same query. These are not edge capabilities. They are the core outputs of most clerical roles.

The 9.0/10 score reflects the degree to which the task content of the role is directly replicable by current AI systems. A score of 10.0 would imply complete replicability with no human judgment required anywhere in the workflow. The 1.0-point gap from a perfect score reflects the irreducibly human elements that still exist in some clerical contexts - the edge case that requires escalation judgment, the client relationship that requires a named human contact, the internal process that requires navigating undocumented institutional knowledge. Those elements exist. They do not dominate the role.

Customer services clerks (ISCO 42) score 8.5/10 rather than 9.0/10 because their work involves more direct human interaction and more variable conversation contexts than pure keyboard and data processing roles. The human-facing element provides a modest protection. As the Anthropic data shows, however, even that protection is already being eroded at scale.

The Anthropic data: disruption is already happening

Massenkoff and McCrory (2026), publishing through Anthropic Research on March 5, 2026, produced what is arguably the most direct measurement of AI labour market impact available. Rather than modelling theoretical exposure, they measured actual Claude usage patterns against occupational task structures - producing an "observed" exposure rate that can be compared directly against theoretical maximums.

The results for clerical workers are unambiguous. Customer services clerks (ISCO 42) have a theoretical AI exposure of 78% - meaning 78% of their work tasks are in principle automatable by current AI. The observed rate, measured from actual AI usage in real workflows, is 70%. The gap between theoretical and observed is just 8 percentage points. In practical terms, this means the displacement of customer service clerks by AI is not a projection - it is a current reality that has already reached 90% of its theoretical maximum.

General and keyboard clerks (ISCO 41) tell a slightly different story. Theoretical exposure: 82%. Observed: 55%. Gap: 27 percentage points. This wider gap means the displacement is real and significant but has not yet reached its theoretical limit. There is both a current impact and additional runway for further displacement - not a reassurance, but a statement about where in the disruption curve this occupation group currently sits.

ICT professionals (ISCO 25) provide the clearest contrast. Theoretical AI exposure: 94% - the highest of any major occupational group. Observed exposure: 33%. Gap: 61 percentage points. This enormous gap explains why tech workers have not experienced the mass displacement their theoretical scores would suggest. The gap is not caused by the AI being incapable - it is caused by the friction of deployment in complex technical environments. Code review, enterprise compliance, the political and social complexity of replacing senior engineers, and the genuine difficulty of integrating AI outputs into production systems without human oversight all contribute to the gap. This gap will narrow - but it provides meaningful runway that clerical workers simply do not have.

Where displacement hits first: the velocity map

The Anthropic data measures what has already happened globally. The velocity data from WorldJobsData tells you where it will accelerate next. Risk Velocity is a composite score reflecting a country's digital infrastructure, AI adoption rates, regulatory environment, and enterprise technology investment. A score of 10.0 means the country has the infrastructure to deploy AI at scale in clerical roles within 1 to 3 years. A score of 0.1 means the infrastructure barriers are so significant that large-scale clerical AI deployment is a decade or more away.

Countries at 10.0 velocity: UK, US, Sweden, Switzerland, Singapore, Poland, South Africa, Vietnam. These are economies with high enterprise cloud adoption, established SaaS customer service infrastructure, and active AI procurement by major employers. A UK company replacing its customer service tier with an AI agent is facing no significant technical or infrastructure barrier. The decision is purely economic - and the economics are increasingly compelling.

Countries at low velocity: Nigeria (0.1), Pakistan (0.3), Philippines (0.3), Indonesia (2.4), Mexico (2.2). In these economies, the theoretical AI exposure of clerical workers is just as high - a Nigerian customer service clerk's tasks are no less automatable than a UK counterpart's - but the infrastructure to deploy AI at scale does not yet exist. Unreliable power, limited cloud infrastructure, low enterprise software penetration, and large informal sectors create deployment barriers that will take years to resolve.

This matters for timing, not for direction. Clerical displacement in low-velocity countries is not "if" - it is "when." The delay provides time for workforce transition that high-velocity countries do not have.

The 61-point gap: why tech workers have time

The ICT professional data deserves specific attention because it is frequently misread in both directions. Some commentators see the 94% theoretical exposure and conclude tech workers are the most at-risk group. Others see the 33% observed rate and conclude tech workers are safe. Both readings are wrong.

The 61-point gap between theoretical and observed exposure for ICT professionals (ISCO 25) reflects deployment friction, not permanent safety. The reasons for the gap are specific and time-limited. First, code review and system integration require human judgment that current AI produces but cannot yet verify reliably - an AI-generated function may compile correctly and still introduce subtle security vulnerabilities that only an experienced engineer will catch. Second, enterprise compliance requirements in regulated industries (finance, healthcare, defence) create institutional constraints on replacing human engineers with AI outputs. Third, the social and political cost of eliminating senior technical roles is high in ways that eliminating a data entry function is not - enterprises face different calculus when the displaced workers are well-paid, unionised, or difficult to replace.

None of these factors are permanent. AI code review tools are improving rapidly. Regulatory frameworks are being updated. The economics of senior engineering replacement become more compelling as AI capabilities advance. The 61-point gap is a measure of today's friction, not a long-term protection. For ICT professionals, the honest reading of the Anthropic data is: you have time that clerical workers do not have, and you should use it.

AI exposure vs observed deployment: the gap table

Source: Massenkoff, M. and McCrory, E. (2026). Labor Market Impacts. Anthropic Research, March 5, 2026. Theoretical % reflects share of tasks in principle automatable by current AI. Observed % reflects actual AI usage measured in real workflows.

What clerical workers should actually do

Most career advice for people in AI-exposed roles defaults to platitudes: "upskill," "embrace AI," "be adaptable." That advice is not wrong, but it is too vague to act on. The data suggests three specific and concrete directions that are grounded in the actual structure of AI displacement.

Direction 1: move toward high-judgement administration. AI is replacing the structured, routine core of clerical work. It is not yet replacing the human judgment involved in policy coordination, case management, complex client relationship management, and roles where the output requires accountability that an institution cannot legally assign to a machine. A data entry clerk and a policy coordinator may both carry the title "administrative professional" - but their AI exposure is radically different. The move is not away from administrative work; it is toward the part of administrative work where human judgment is load-bearing.

Direction 2: move toward care and human services. Personal care workers score 2.0/10 on AI exposure - the lowest in the WorldJobsData dataset - and face structural labour shortages across every aging economy. The UK had 131,000 care vacancies at any given time in 2024 (Skills for Care). The transition from clerical to care work is not trivial - it requires retraining and involves different physical and emotional demands. But the structural protection is genuine and the demand is structural. This is not advice to accept lower pay and harder conditions as a consolation prize. It is advice that the labour market is signalling real demand in care that does not exist in clerical.

Direction 3: build AI operation skills. The displacement of clerical workers by AI does not mean there are no roles left in AI-adjacent clerical contexts. GitHub Copilot usage data (2024) showed developer productivity up 55% - but firms still need humans to review, integrate, quality-check, and manage AI outputs. The same pattern applies in clerical contexts. The person who can manage an AI agent's outputs, catch its errors, handle its escalations, and integrate its work into institutional workflows has a role that did not exist five years ago and is genuinely in demand. Building AI operation competence - not AI development, but AI operation - is a practical and accessible move for people currently in clerical roles.