Inside a European financial institution, a cluster of software agents now spends its working hours simulating the institution’s entire decision structure. Each agent competes with the others, filtering noise and arguing for the strongest signal before a trade is approved. A year ago the same system fielded questions. Today it executes work.
That transition, from conversational assistant to operator of whole stretches of a job, is the subject of a 68-page report assembled by researchers at Fudan University together with colleagues at Stanford University and the University of California, Berkeley, drawing on input from twelve institutions. The paper, which circulated through academic and industry channels this month, argues that AI has moved out of the chat window and into operating roles across finance, law, manufacturing, agriculture, trade and medicine.
The authors document the progression field by field. In law, the technology has traveled from retrieval assistance to making its reasoning explicit at each step, and in some pilot settings, to modeling an entire trial proceeding. On factory floors, it has moved from inspecting finished goods for defects to scheduling production and weighing operations decisions against live cost data. On farms, systems orchestrate vision models and knowledge graphs to diagnose crop disease, combining image recognition with accumulated agronomic knowledge.
In cross-border trade, the technology handles the understanding and orchestration layer, reading contracts and invoices, while handing tariff calculations and sanctions screening to deterministic software modules that regulators can audit. In hospitals, the report traces a completed path from knowledge injection, through capability alignment, to agentic execution at the point of care. The case studies share a pattern, the authors argue: each deployment starts with a narrow task, gains confidence, then absorbs adjacent responsibilities until the human role shifts from doing the work to supervising it.
The paper calls this process task absorption, and it says this is the mechanism by which AI converts into measured productivity rather than a headline metric. It arrives as the productivity question has become the central debate in AI economics. Spending on data centers, chips and models is running at hundreds of billions of dollars a year, and executives are under pressure to show the outlay is producing more than clever demonstrations. Economists have been skeptical: productivity statistics in most major economies have shown little acceleration since the technology became broadly available, a lag that matches historical patterns for previous general-purpose technologies such as electricity and the personal computer.
The Fudan-led group is careful about its claims. The paper cautions that early case studies overstate results, that the savings documented in pilots do not always survive scale-up, and that the hardest part of adoption is not the model but the reorganization of work around it. It notes that the firms reporting the largest gains are those that redesigned processes rather than simply adding an assistant to existing ones. It also flags what it calls the measurement gap: much of the work AI now performs, simulating a decision structure, modeling a trial, orchestrating a supply chain, does not show up in the categories statisticians use to track output. If the gains are real but invisible to current instruments, the report argues, policy decisions based on those instruments will systematically underestimate the technology’s effect.
Analysts who follow the sector read the paper as a tempering document. The evidence is real but uneven, one technology economist said. The deployments that work are narrow, well-instrumented and supervised, and that is not the same as the economy-wide transformation the marketing suggests. The report’s own numbers support the caution: across the case studies it examined, the median productivity gain was meaningful but concentrated in a minority of deployments, with the rest showing gains that were small or not yet measurable.
For the institutions in the case studies, the questions are practical rather than philosophical. Banks are weighing how much of a trader’s workflow to hand to agents when regulators require a human to explain every decision. Hospitals are deciding how far a diagnostic model can go before a physician signs off. The report offers no single answer, only a framework: start narrow, measure relentlessly, and keep the human accountable for the outcome.
The report’s scope is deliberately international. Its case studies span companies and institutions in China, the United States and Europe, and the authors argue that the adoption pattern looks similar across borders: the same task absorption curve appears in a Shanghai trading desk, a German machinery plant and a California hospital network. What differs, the paper says, is the speed of regulatory accommodation, which in finance and medicine lags the technology by the widest margin.
Back at the financial institution where the story began, the agents have stopped competing for a few minutes each day. Their human supervisor reviews the log of what they simulated, what they filtered and what they recommended, then signs off. The work is faster; the job is different. That, in miniature, is the transition the report says is now underway across the economy, less dramatic than the product launches, harder to measure, and, if the authors are right, where the real productivity gains will be found.
