In 2023, Shakked Noy and Whitney Zhang ran a controlled experiment that produced one of the clearest findings in the AI productivity literature. Professionals given access to generative AI completed writing tasks – including analyses, reports, memos, client communications – roughly 40 percent faster, with independent judges rating the AI-assisted outputs as higher quality. This was not a survey of perceived usefulness, but a randomised controlled trial with direct output measurement, published in Science. The effect was sufficiently large to be unambiguous, and the underlying mechanism it revealed is worth careful examination.
Knowledge work in finance, law, and public administration shares a structural feature that most productivity discussions overlook: the bottleneck is rarely judgment. Experienced professionals in these fields generally know what a contract review should conclude, what a credit memo should recommend, or what a policy brief should argue. What consumes most of their working time is the preceding step, assembling relevant information, structuring it coherently, and producing a first draft that can then be refined. This is precisely the stage at which large language models perform well. They are, at their core, first-draft machines: fast, coherent, and capable of synthesising large volumes of structured text into readable output.
AI does not improve professional judgment. It removes the bottleneck that comes before it. Rather, it is that AI removes the first-draft bottleneck and allows professionals to spend more time on the tasks they were trained to perform. A lawyer who previously spent four hours drafting a contract review followed by two hours of refinement may now spend forty minutes on the draft and two hours on refinement. Quality improves because more cognitive bandwidth is allocated to the stage that actually requires expertise. This is the channel documented by Noy and Zhang, and it holds across professional writing tasks with striking consistency. The next question is where this mechanism matters most in practice.
The implication for European firms is concrete. ECB data on AI uptake show that financial and professional services sectors lead adoption across the euro area, precisely where this productivity channel is most accessible. A bank’s credit analysis team, a law firm’s contract review practice, or a public administration unit responsible for regulatory reporting all spend most of their working hours on exactly the type of task in which a roghly 40 percent reduction in completion time has repeatedly been observed. As a Corvinák analysis on the digital skills gap has documented, the critical variable is not access to tools but organisational capacity to integrate them into existing workflows.
The prescriptive implication is more specific than general AI adoption advice tends to be. The gains in this channel are concentrated in tasks with three properties: a clear output format, a large volume of structured input, and a revision stage in which professional judgment is the primary value added. Contract review, regulatory filing, internal research synthesis, client communication drafting, and compliance documentation all fit this profile. Tasks involving unstructured negotiation, novel legal reasoning, or discretionary client judgment do not. As examined in a Corvinák article on digital ecosystems, the productive question is not whether a tool is transformative in the abstract but whether it removes a specific friction in a specific process.
Measurement matters here. Firms tracking only output volume will capture only part of the gain. Those tracking quality, accuracy of analysis, reduction in revision cycles, and lower escalation rates will capture more. As argued in a 2021 Corvinák article, each major technological transition has required institutions to rebuild their measurement frameworks from the ground up, and those that failed to do so converted the technology’s potential into visible economic gains far later than those that did. The first-draft bottleneck is one of the rare places where AI already delivers measurable productivity gains, not because it replaces expertise, but because it allows expertise to begin sooner.
