Who Is Exposed to the New Workbench?

As frontier models entered more office tasks, the ILO’s February evidence showed why AI exposure and job loss are different—and why gender matters.

Retrospective written 16 October 2026; focus month: February 2026.

In February, two kinds of announcement landed in the same week. A frontier model maker described a system that could hold a million tokens of context and perform longer work tasks. The International Labour Organization published a report asking who, in the Philippines, was most exposed to the tasks such systems might change. One story was about capability; the other about the shape of work and the people who do it.

On 5 February, Anthropic announced Claude Opus 4.6, with a one-million-token context window in beta and product examples spanning coding, financial analysis and document work. On the same date, the ILO published Generative AI and jobs in the Philippines, a research brief on occupational exposure and policy implications. Its estimates said more than one-quarter of Philippine employment—12.7 million jobs—was exposed to generative AI, but only 3.6% of jobs fell in its highest exposure category associated with elevated displacement risk. The brief stresses that exposure often means tasks can be transformed, not that an entire job disappears.

That distinction is the beginning of a more serious labor conversation. “AI will take jobs” compresses several separate questions: Which tasks can a system perform? Which tasks are actually adopted by employers? Does adoption substitute for labor or complement it? Do productivity gains create new demand? Who receives the savings? What happens to wages, hours, job quality and bargaining power? Exposure estimates answer only part of the first question.

The gender pattern inside the task map

The ILO brief found that women in the Philippines faced roughly twice the rate of exposure to generative AI as men, associated with women’s greater representation in occupations such as clerical and administrative support. It also reported that young women may face greater potential disruption than young men in some occupational groups. These are findings for the Philippine labor market using the report’s occupational method. They are not estimates for Kenya, and should not be copied into a Kenyan forecast.

Still, the mechanism deserves attention in East Africa. If a technology is strongest on text-heavy, routine, screen-based tasks, the impact will depend on which workers perform those tasks and on the way a local employer reorganizes work. Occupational segregation can make exposure unequal even when a model contains no explicit gender rule. Conversely, a tool that automates some repetitive clerical work may free time for more complex duties if workers receive training, better job design and a fair share of the productivity gains.

The ILO’s 5 March global research summary reported that, across countries with available data, female-dominated occupations were almost twice as likely to be exposed as male-dominated occupations: 29% versus 16%. It also found larger representation of women in the highest exposure categories. The report links this pattern to occupational segregation and women’s underrepresentation in AI and STEM roles. This is a structural risk signal, not a prediction that women will lose twice as many jobs.

What the evidence does not say

Exposure is not an outcome. A receptionist’s tasks may include scheduling, interpersonal support, judgment and confidential handling of sensitive situations. A language model might draft a routine message or summarize information, but it may not safely manage an upset patient, resolve a billing dispute or recognize when a person needs help beyond a standard script. An occupation is a bundle of tasks and relationships, not one line on an automation chart.

Nor does a model benchmark establish a productivity gain in a real workplace. Anthropic’s announcement described capabilities and vendor-run evaluations. A clinic or school would need to test whether the system reduces total workload after review, correction, integration and privacy safeguards. A system that produces drafts quickly but requires extensive checking may shift work rather than remove it. The relevant measure is the whole task, including errors and oversight.

The ILO’s later joint paper with the World Bank, published 27 March, examined exposure across 135 countries and argued that developing economies could experience disruption before capturing productivity gains. The authors point to digital divides and differences in task composition. Again, that is a broad comparative warning, not a Kenya-specific forecast of unemployment. Local labor-force data, adoption rates, internet access, employer practices and education pathways would be needed to estimate Kenya’s actual exposure.

Education is part of the transition, but not a magic answer

When tasks change, workers need ways to learn the new work. Digital fluency, critical reading, data literacy and the ability to check machine outputs can help. Training alone cannot secure workers if employers reduce headcount or new roles remain inaccessible. Education policy should connect classroom learning, technical and vocational education, employer demand and protections during transitions.

For young women entering the workforce, if AI absorbs entry-level writing, bookkeeping or scheduling tasks, the route by which workers learn an occupation may narrow. Employers and educators can preserve supervised apprenticeships and pathways into higher-skill work. This is a risk to investigate, not a claim that jobs have vanished in Kenya. The ILO recommends social dialogue and targeted policies: collect occupation-level data, consult workers, and monitor hiring, pay and promotion by gender and age. Women should help design and govern AI systems as developers, evaluators and decision-makers.

A Kenyan lens for employers and health services

Kenyan organizations should commission local task audits before making workforce decisions based on overseas exposure percentages. Start by mapping tasks, not job titles. Ask which tasks are repetitive, which require judgment or human trust, what data the tool can access, and who is accountable for errors. Test the system with examples from Kenyan work, languages and records. Examine results by role, gender, age, disability and location where lawful and feasible.

In health care, AI can draft a referral, summarize a chart or organize patient questions. Clinicians remain responsible for decisions, and patients need access to a human conversation. A tool may reduce paperwork or add privacy risks and review; its benefit should be measured in workflow, including staff workload.

February’s announcements made the capability frontier visible. The ILO’s evidence made the distribution question visible. If productivity gains are real, institutions should ask who gains time, whose work is monitored, whose roles become more precarious, and whether new opportunities reach people previously excluded from technical careers. The answers will determine whether AI changes work for the better or simply shifts risk onto less powerful workers.

From 16 October, a useful forward test is not a precise job-loss probability unsupported by Kenyan data. It is whether Kenyan employers publish task-level evaluations, workers participate in deployment decisions, and training pathways reach women and young people. Release cadence, tracked separately by F11’s 85% registered judgment, cannot answer that question. This is an uncalibrated probability about model releases, not a fitted output or labor forecast. The labor outcome will depend on adoption, institutions and distribution—not just the model’s context window.

Part 02b of The World Ledger: 2026, Read Backwards.

The dividend needs an owner

The founder-authored Kenya Vision 2060 proposes a shift toward broad productive ownership as cognitive work becomes cheaper. That is a political and institutional choice, not an automatic property of AI. Training without access to customers, capital or bargaining power may produce skilled people whose gains accrue elsewhere. Over the next year, assess who receives higher income, shorter hours or an ownership stake when a task becomes easier. Over several years, test whether women and young workers gain durable assets as well as access to software.

Read the World Ledger

All 30 articles. Previous: Four Banks, Four Clocks. Next: A Cheaper Rate Does Not Cheapen Supper.

Mathematical methods: 45 parameters, assumptions and limits. Download the research and executable model.

AI-assisted research and writing using parallel Luna agents; editorial synthesis under the APEX lens. Historical scenes are illustrative. Sources support the stated observations; forward scenarios and inherited probabilities remain conditional, uncalibrated judgments. Kenya Vision 2060 is a founder-authored thesis, not an adopted government plan.

Cover image: Woman with computer.jpg — Cumminsr at English Wikipedia · CC BY-SA 3.0. Photographed August 7, 2013; the Commons description identifies only a woman using a computer and gives no location. Use it as a generic illustration of computer-based work, not as evidence about a specific workforce or AI deployment. Resized for web display.