Future of work

What AI-at-work evidence can tell a developer, a teacher and a factory team

AI is changing work through several mechanisms at once: people adopt assistants, organisations redesign tasks, some tools improve performance in particular settings, and other deployments add checking or coordination work. These mechanisms should be studied together without being treated as the same measurement. A conversation classified as a work task is evidence of use. It is not proof that the task was completed correctly, that a worker saved time or that an employer eliminated a job.

This distinction matters across industries. A software team can measure an accepted maintenance contribution; an educator can assess a lesson and learner understanding; a factory team can review an equipment-specific maintenance recommendation. Those outcomes have different units and consequences. The evidence below, checked on 15 October 2026, supports specific observations. The practical implications are editorial proposals for applying those observations carefully, not measured results for every sector or country.

Exposure describes a potential task relationship

ILO’s 20 May 2025 brief refines global occupational exposure using task data, expert input and AI predictions. It distinguishes levels of exposure across occupations and describes transformation as more likely than complete replacement overall. This is a model of potential task effects, not an observed count of redundancies. It does not say how a specific employer will respond, whether a particular tool is reliable or which workers will receive training. ILO

For an enterprise reader, exposure can help identify questions to investigate. A role combining document preparation, interpersonal work and physical inspection should be broken into those tasks before a pilot is designed. Even if document preparation is highly amenable to assistance, the job may remain constrained by inspection capacity, customer demand or a professional decision. The proposed action is to map the task bundle and its dependencies, rather than multiply an exposure share by headcount and call the result expected layoffs.

Platform traces reveal use, with a platform-shaped boundary

Anthropic’s 24 March 2026 Economic Index report analyses Claude usage sampled from 5–12 February 2026. It reports that the ten most common Claude.ai tasks represented 19% of conversations, down from 24% in its November sample, and describes coding’s movement toward first-party API workflows. The report also finds associations between longer user tenure and conversation success, while acknowledging that early-adopter characteristics can contribute. These are company analyses of its own platforms, using task classifications; they are not a representative census of employers, independent output audits or causal estimates of worker learning. Anthropic

OpenAI’s 30 June 2026 Signals account reports deeper and broader Individual-plan use over time. Its depth/breadth analysis uses a 0.1% sample of users whose accounts were created between 15 October 2025 and 1 May 2026, with activity through 31 May. Capabilities are classified into 53 categories; exclusions include under-18s, banned users and users inactive for the specified post-signup period. Regional growth is indexed to each region’s own July 2023 baseline. These are adoption and usage observations, not workplace productivity, wage or employment estimates. OpenAI

Together these sources show why adoption research belongs in an industry report, but needs its own column. They can reveal the kinds of activities people bring to a service and how use changes over time. They cannot establish that every recorded activity is paid work or that increased messaging creates proportionate economic value. A business application built on an API can also split one human task into many calls, making raw traffic counts a poor basis for comparing jobs across tools.

For a developer or founder, the proposed implication is to define a useful unit before collecting metrics. In a software repository, that might be a reviewed issue resolved. In retail, it might be an accurate product listing accepted by a merchandiser. In an agricultural advisory service, it might be a question correctly routed to a suitable source or person. Tool interactions can be retained as operational context, but should not silently replace the outcome the organisation actually wants.

Task performance can move in opposite directions

The final Quarterly Journal of Economics article “Generative AI at Work,” published on 4 February 2025, studies staggered introduction of assistance using data from 5,172 customer-support agents at one firm and its subcontractors. It reports a 15% average increase in issues resolved per hour, with heterogeneous effects: less experienced and lower-skilled workers gained more, while the highest-skilled workers saw small quality declines. The authors explicitly say their data do not establish aggregate employment or wage effects. The final 15% figure should not be confused with earlier working-paper versions reporting 14%. Published study

METR’s early-2025 randomised study offers a different setting: 16 experienced open-source developers completed 246 tasks in mature repositories they knew well. Allowing early-2025 AI tools increased completion time by 19% in that experiment, despite participants’ expectation that assistance would help. The authors emphasise that this does not establish uselessness in other settings or determine what future tools will do. It is a bounded result about particular participants, tasks and tool versions. METR

METR’s 24 February 2026 follow-up explicitly changes the measurement picture. It describes a later experiment with selection and time-measurement problems: some developers did not want to work without AI, some selected tasks accordingly, and concurrent agents complicated time accounting. METR says this gives an unreliable signal of the current productivity effect and only very weak evidence about the size of a likely improvement from early 2025. The historical slowdown should therefore not be presented as a current benchmark. METR’s 2026 update

The comparison does not establish which industry will “win.” The studies use different tasks, workers, tools and evaluation designs. It does establish a reason to resist a universal productivity multiplier. Assistance that disseminates useful support practices may have a different effect from assistance that must meet a mature repository’s demanding contextual requirements. The appropriate next question is what kind of knowledge the tool supplies and what new work the human must do to assess it.

For a teacher, this means that drafting lesson material and improving learning should be measured separately. For a clinician’s documentation team, typing time and the safety of the reviewed record are separate outcomes. For a factory technician, a clear recommendation is not useful if it refers to the wrong equipment revision. These are proposed applications of the measurement lesson, not findings that the cited support or software experiments establish in education, health or manufacturing.

Management changes can affect the experience of work

OECD’s 6 February 2025 employer survey covers more than 6,000 firms in France, Germany, Italy, Japan, Spain and the United States. It examines algorithmic management, which can include AI, and reports managers’ perceived decision benefits alongside concerns about accountability, understandable logic and worker health. The perspective is important: employer perceptions do not substitute for workers’ direct accounts, and the six-country sample should not be described as every workplace worldwide. OECD

An organisation can therefore ask two parallel questions during a pilot: did the output improve, and how did the work change? A dispatcher may receive a better organised incident summary but face tighter timing demands. A catalogue reviewer may correct more listings in an hour while losing time for difficult exceptions. A junior analyst may draft faster while getting less practice building an argument. These are illustrative possibilities, not allegations about surveyed firms or claims that AI necessarily causes these outcomes.

A proposed evaluation can give workers a safe way to record effort, interruption, discretion and the ability to challenge a result. It can include training time and the work of people who handle escalations. Consultation should occur while the workflow can still be changed. The proposal is a practical management choice, not an assertion of one worldwide statutory entitlement or a promise that consultation alone resolves every risk.

Worldwide reporting needs a local denominator

OpenAI’s regional series uses relative growth from each region’s own baseline; Anthropic’s sampled usage remains geographically uneven. Neither establishes equal adoption, infrastructure or workplace outcomes across countries. A fast-growing user population can begin from a small base. A widely used tool can still be inaccessible to a particular workplace because of language, connectivity, permissions, cost or unsuitable source material. Those factual conditions should be collected locally before a global headline becomes a staffing assumption.

The proposed industry profiles accompanying this synthesis keep these questions concrete. Each identifies a task boundary, affected responsibilities, skills, a bounded exercise, possible measures and local limits. Their diagrams describe proposed relationships, not observed productivity gains or job-loss flows. Their media briefs describe fictional editorial scenes, not documentary evidence of a deployment. The profiles are useful starting points for deeper sector reporting; they are not thirty measured industry studies or thirty accepted native articles.

A sound next step is small enough to inspect: state the task and baseline, involve the people doing the work, compare useful output and full effort, retain failures and explain which conclusions the evidence can support. A useful result may justify another carefully scoped trial. An unresolved result should remain unresolved. That approach gives readers a way to act on emerging evidence while keeping the distance between an interesting capability, a changed workflow and an actual employment outcome visible.