The most expensive mistake a firm can make with generative AI is to treat it as a way to do the same work with fewer people. Erik Brynjolfsson (2022) named this the Turing trap. When we benchmark machines against human performance, there are excess incentives to automate rather than augment. If a task is completely handed to a model the worker ends up with less bargaining power and a complete loss of capability follows. The distinction is not rhetorical, where AI substitutes for a worker it captures the value created; where it complements one, the worker keeps a claim on it. Which of the two actually pays is now the key empirical question.
Brynjolfsson, Li & Raymond (2025), studying 5,172 customer-support agents for the Quarterly Journal of Economics, found a 15% average rise in issues resolved per hour, concentrated among novice and lower-skilled workers; the most experienced gained little and lost slightly on quality. The system worked by allowing tacit knowledge to spread. AI here became a complement to human skill rather than a replacement for it. The gain, though, may be conditional on the context. Dell'Acqua et al. (2023), in a pre-registered trial with 758 Boston Consulting Group consultants, found AI cut performance by around 19 percentage points on tasks just outside its capability frontier, where workers over-trusted output that looked correct but was in reality flawed.
That miscalibration acquires a name once it reaches the inbox. A survey of 1,150 US desk workers by Niederhoffer et al. (2025) for Harvard Business Review found 40% had received "Workslop" in a single month. Under this definition we find AI output that looks finished but lacks substance. Unfinished results that will be pushed onto a downstream colleague for further correction. Each incident is estimated to cost two hours, an invisible tax the authors estimate costing roughly 9 million USD a year for a 10,000-person organisation. The sample is self-reported and US-centric, but the mechanism is sound. Speed here is gained at the front of a workflow, but is compensated with a slower development further down the line.
The papers converge on an awkward point for any cost-cutting case. The productivity is real, but it mostly concentrates around complementarity. Improvements seem significant when AI extends what a worker can already do, while thinner gains are present when substitution becomes the rule of the game. Acemoglu, Autor & Johnson (2023), writing for CEPR, argue that what we are about to face has much more to do with a choice than a pre-determined destiny. Firms currently lean towards automation and surveillance, but a human-complementary path stays open if tax and innovation incentives stop favouring the machine over the worker. For Europe, where the regulatory instruments already exist, that path is something policy can set.