THE SIGNAL IN ONE SENTENCE
The loudest AI jobs debate has been running on a surprisingly small fuel tank. We know what models can do in a benchmark. We know what companies say they are buying. We know what workers fear and what executives promise. We know much less about which tools enter a workplace, which tasks people actually hand over, how jobs are reorganized, who gets more pay or less power, and whether the result lasts. A group of organizations used the London School of Economics' Global Forum on September 16 to announce pieces of a better measurement system. Cohere Labs released nearly 700,000 agent tools collected from public Model Context Protocol servers. The World Bank, IMF, OECD, and technology companies are developing a cross-country adoption index. The IMF plans a public hub with more than 100 indicators later this year. New projects will study UK labor-market effects, Latin American jobs, African economies, and workers across the AI supply chain. These are useful inputs, not a verdict. A catalog of tools does not prove use. Usage does not prove displacement. A national average can hide a brutal local change. The plain signal is that the jobs argument finally has a serious research backlog. The next job is making the evidence comparable, independent, worker-visible, and connected to decisions.
01
WHAT ACTUALLY CHANGED
LSE says 15 organizations used its September 16 and 17 Global Forum for AI and the Social Sciences to announce funding, datasets, collaborations, and research projects about AI and labor markets. The package includes USD 1.5 million from Google.org for LSE research on AI and the economy and USD 500,000 from the Collaborative for AI in the Public Interest for a 19-month project on workers across the AI supply chain. Funding announcements establish resources and intentions. They do not establish a method, completed analysis, independent finding, or improved worker outcome.
Cohere Labs released the Agentic Task Ecosystem, an open dataset that aggregates nearly 700,000 tools from public Model Context Protocol servers. Its analysis maps tools against the Occupational Information Network and says about one tool in 40 performs a recognized work task from end to end. Across 178 occupations, the team reports different patterns: tools reach specialized tasks in healthcare and computing while clustering around more routine edges in legal, production, and sales work. The dataset measures what developers have exposed as tools on public servers. It does not measure private tools, actual workplace deployment, usage frequency, reliability, worker consent, productivity, wages, layoffs, hiring, or bargaining power.
The Development Data Partnership is facilitating a Global AI Adoption Index with the World Bank Group, IMF, OECD, Anthropic, Google, Microsoft, OpenAI, OpenRouter, and other partners. The coalition plans a shared method using privacy-protected company data on consumer chat and API use, normalized against population, income, connectivity, and national baselines. Meta, LinkedIn, Ookla, and GitHub are named as contributors of complementary indicators. The first statistics are planned for joint release in mid-2027. No final methodology, coverage table, audit process, revision policy, public microdata, or first result was released on September 16.
The IMF announced an AI Economy Data Hub for later this year. It is intended to bring more than 100 open-source indicators into one public place, covering adoption and maturity, labor-market effects, infrastructure, and investment. The IMF says each indicator will carry its source and definition and will be viewable or downloadable. That can make scattered evidence easier to inspect. It cannot make unlike measures comparable by proximity, remove errors inherited from a source, or turn a descriptive chart into a causal finding.
Several projects widen the map. The Resolution Foundation and partners plan to develop and pilot UK labor-market measurement tools. LSE's International Inequalities Institute and the UN Economic Commission for Latin America and the Caribbean plan an initial project on the quantity and quality of jobs in Latin America. The four-year African Growth and AI Network will connect researchers in countries including South Africa, Kenya, Uganda, and Ghana, with attention to informal workers, women, and young people. Qatar Foundation opened a collaboration call for places where standard employment, occupation, and wage measures may not exist. These are project descriptions and geographic commitments, not published outcomes.
02
WHY THIS MATTERS
AI can touch work through several different doors. A developer can publish a tool. A company can buy it. A team can test it. A manager can require it. A worker can use it rarely, rely on it daily, or spend extra time correcting it. A task can shrink while the job remains, or a job can disappear while the occupation grows elsewhere. Treating any one door as the whole story produces very confident nonsense.
The Cohere dataset offers an early view of the supply side: what developers appear willing to let agents do. That is valuable because tools are closer to deployment than general model benchmarks. It also carries selection problems. Public Model Context Protocol servers are not the entire market, duplicates and abandoned tools can distort counts, one tool can cover several tasks, and a listed capability can fail under real constraints. Tool supply should be joined to observed use and outcomes, not promoted into a jobs forecast by itself.
Company logs can reveal adoption at a scale surveys may miss. They can also place the measurement system inside the firms being measured. A credible global index needs stable definitions, country and language coverage, disclosure thresholds, privacy review, independent governance, versioned methods, conflict rules, and a way to show when a provider changes its product or logging. Otherwise a harmonized index can make incomparable commercial traces look official simply because they share a chart.
Regional research matters because labor markets are not interchangeable. Informal work, household enterprises, public employment, agriculture, platform labor, language access, connectivity, worker protections, and statistical capacity vary sharply. A method built around salaried software users in wealthy countries can miss both displacement and opportunity elsewhere. Comparable does not have to mean identical, but every difference needs to be documented.
Good measurement should change a decision. Evidence should tell a government when to fund training or income support, tell a union where duties expanded without compensation, tell an employer when monitoring grew faster than productivity, and tell a worker whether a tool improved mobility or merely raised the pace. A dashboard that never triggers action is an attractive waiting room.
03
WHERE IT COULD HELP
- Publish a measurement ledger for every headline claim: unit of analysis, population, geography, occupation, task definition, period, source, missing groups, uncertainty, revisions, funding, conflicts, and the decision the measure is supposed to inform
- Follow a repeated worker panel before and after adoption, recording actual tool use, task mix, correction time, hours, pay, autonomy, monitoring, stress, training, promotion, exit, and subgroup effects instead of asking only whether an employer bought AI
- Make the Global AI Adoption Index publish provider coverage, country and language gaps, normalization choices, privacy thresholds, methodology versions, audit findings, suppressed cells, historical revisions, and a public route for researchers and civil society to challenge an indicator
- Join the agent-tool catalog to deployment receipts and outcome data: which version ran, who assigned it, what task it attempted, how often it succeeded, who reviewed it, what changed in staffing, and whether productivity gains reached workers or customers
- Pre-commit policy triggers so evidence has consequences, including consultation, bargaining, training funds, portable benefits, income support, inspection, procurement limits, or repeat evaluation when displacement, pay loss, unsafe monitoring, or unequal burden crosses a published threshold
KEEP A HAND ON THE WHEEL
The verified event is a collection of September 16 announcements, one openly released tool dataset, and several planned indexes, hubs, studies, funding programs, and collaborations. It is not a global estimate of AI job losses or gains. Nearly 700,000 public tools are not 700,000 deployed systems or automated jobs. The reported one-in-40 result comes from Cohere Labs' mapping method and does not establish workplace performance or economic impact. The Global AI Adoption Index is under development, with first statistics planned for mid-2027. Privacy-protected company data can improve coverage while limiting independent reproduction. The IMF hub is planned for later this year and has not yet demonstrated indicator quality or comparability. Regional projects have not published results. Announced funding does not prove research independence, worker governance, open publication, or policy uptake. Watch for dataset documentation and deduplication, code and licenses, coverage by country and language, worker and union roles, preregistered questions, access to company data, independent replication, causal designs, attrition, subgroup power, negative findings, revision logs, and proof that results changed a real decision.
04
TERMS WORTH KEEPING
OPEN GLOSSARY CARD
Task exposure
The degree to which a job task could be affected by a technology, whether or not that technology is actually used or changes employment.
OPEN GLOSSARY CARD
Adoption index
A composite measure intended to compare how widely a technology is used across places, sectors, organizations, or groups.
OPEN GLOSSARY CARD
Administrative data
Records created while an organization operates a service, program, workplace, or system rather than specifically for a research survey.
SOURCES AND VERIFICATION STATUS
This article was written from the materials below. Product claims and dates were checked against those sources on September 16, 2026.
PUBLICATION RECEIPT: Revision 1. Published September 16, 2026.
THE PUBLICATION ENGINE
WANT A SIGNAL OF YOUR OWN?
We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.
WORK WITH US