Future-of-work researchers and AI-enabled knowledge work

Define the researcher's role and decision

Future-of-work researchers may produce labour-market studies, workforce scenarios, programme evaluations, occupational taxonomies, or evidence for policy and organizational planning. They may be an owner of research design, an influencer on strategy, a reviewer of external evidence, or a collaborator with workforce and functional leaders. They do not automatically own technology deployment, training budgets, or employment decisions.

Begin with the decision the research must inform. Examples include selecting occupations for deeper study, designing a worker survey, interpreting a productivity pilot, or deciding whether observed tool use warrants a workflow experiment. Name the audience, time horizon, population, unit of analysis, and consequence of error before collecting attractive statistics.

The International Labour Organization's Generative AI and Jobs: A Refined Global Index of Occupational Exposure uses task-level data, surveys, expert input, and model predictions to estimate exposure gradients. The report explicitly treats exposure as potential rather than observed impact. That distinction is central to responsible future-of-work analysis.

Sources: [1]

Build an evidence-layer map

Create one row per evidence item. Record whether it describes model capability, task exposure, access, use, work-process change, output quality, productivity, worker experience, employment, or distributional effects. Add the study design, sample, geography, occupation, tool, outcome definition, comparison group, date, and stated limitations.

Do not collapse levels. A benchmark showing that a model can complete a task is not evidence that an organization has deployed it. A survey of use is not evidence of output quality. A time-saving experiment in one workflow is not a forecast for every occupation. An exposure index is not a prediction that a task will be automated.

The OECD's Artificial Intelligence and the Future of Skills programme develops measures connecting AI capabilities with occupational requirements while noting that real-world effects also depend on adoption, regulation, organizational change, and social choices. Use such measures as structured inputs, not deterministic forecasts.

Sources: [2]

Worked artifact: a synthetic research briefing

Imagine a synthetic researcher asked whether generative AI will reduce analyst headcount over the next two years. The evidence pack contains an occupational exposure index, a company survey showing broad tool access, a small productivity pilot, and interviews describing review burden.

The evidence-layer map labels the exposure index as task-potential evidence, the survey as access and self-reported use, the pilot as workflow-specific outcome evidence, and the interviews as qualitative evidence about implementation. The researcher refuses to add the percentages together or translate them directly into jobs lost. Instead, the briefing identifies three scenarios: limited adoption, task redesign with stable staffing, and narrower automation of routine subtasks.

Each scenario lists observable indicators that would distinguish it: workflow coverage, task allocation, review time, output volume, quality defects, vacancies, redeployment, and hours worked. The final recommendation is to collect task-level baseline data and repeat measurement after a bounded deployment. This synthetic artifact is not a labour forecast or Two Prune research result.

Test claims against design and denominator

For every headline finding, ask who and what is in the denominator, what comparison creates the effect, whether the outcome is directly measured, and which alternative explanations remain. Check whether a study's unit is prompts, workers, tasks, firms, occupations, or jobs. Confirm that dates and technology versions match the question.

Preserve heterogeneity. Effects can differ by experience, task, organization, adoption quality, and time. Report null findings and adverse shifts alongside positive ones. Where evidence comes from a vendor-supported study, record the relationship and inspect the design rather than dismissing or accepting the result solely because of funding.

Verify and communicate uncertainty

Have another reviewer trace every material statement to the evidence-layer map. Mark whether language describes capability, association, experimental effect, projection, or scenario. Reproduce calculations where possible and retain data transformations. Reopen the analysis when the technology, occupation definition, deployment pattern, or outcome data changes.

This method improves traceability between evidence and future-of-work claims. It does not make a universal employment forecast, establish causal effects beyond a study, or prove that Two Prune changes workforce outcomes. Two Prune can provide task-bounded assessment evidence from realistic simulations for human interpretation; it does not automate employment decisions.