A familiar mantra in radiology claims that AI will not replace radiologists but those who fail to adopt it. While reassuring and adoption-friendly, this framing avoids a harder economic question: with massive investment in imaging AI and mounting healthcare costs, can sustainability be achieved without replacing human labor?

The current positioning of AI in radiology is internally inconsistent. AI is promoted as augmenting clinicians and keeping radiologists and technologists fully in the loop, while its economic value proposition, as in other industries, is labor substitution, thus delivering more output with fewer staff [1]. It is time we confront the possibility that, for AI to be economically viable, it may indeed need to substitute human roles.

To understand the economic trajectory of AI in radiology, we must look outside our reading rooms. In sectors such as retail, manufacturing, and logistics, the integration of AI and automation has rarely been purely about human “augmentation” for its own sake. Rather, this process has been focused on efficiency, i.e., a reduction in the unit cost of production. In customer service, AI chatbots have not merely “augmented” call center agents; they effectively replaced the first tier of support, handling routine queries so that a smaller workforce can manage complex issues [2]. In manufacturing, automation did not just make workers faster; it reduced the number of workers required to maintain the same output. The economic model relies on “labor polarization,” where technology complements high-skill tasks while substituting low-skill or routine labor [3].

Radiology is not immune to these market forces [4]. While we view medical imaging as a high-stakes, high-expertise field, it is composed of numerous process-driven tasks, such as patient positioning, protocol selection, image reconstruction, and initial triage. If AI and, more recently, robotics follow the same trajectory in healthcare as in the wider economy, their persistence will depend on the same mechanism: substituting human labor in routine tasks to reduce costs [3]. Current AI applications have also, rightfully, targeted high-volume, standardized diagnostic tasks, including cancer detection. However, we must recognize a fundamental limitation in these tools: they are largely products of supervised learning. Algorithms trained on expert labels can approach, but rarely exceed, expert consensus [5]. They replicate existing detection and characterization tasks rather than uncovering new diagnostic capabilities. This performance ceiling, where AI matches but does not surpass competent specialists, strengthens the case for substitution rather than augmentation. If an AI system offers performance matching a human reader, clearly, the primary mechanism to extract value from it cannot be represented by improved clinical performance. Rather, operational efficiency becomes the main added value provided by AI.

We are already seeing this shift in some screening environments. In breast cancer screening, recent pivotal studies suggest that the most viable economic path is not adding AI as a “third reader” to catch missed cancers but rather using AI to replace the second human reader in double-reading protocols [6]. In this scenario, technology is not being used to achieve super-human results, but to achieve standard-of-care results while requiring half the radiologist workforce.

The hesitation to acknowledge this reality appears to stem, at least in part, from the current lack of robust economic evidence. A recent systematic review highlighted that while studies on AI’s technical performance abound, thorough economic evaluations (such as cost-effectiveness analyses) are still lacking in quality and quantity [7]. Most current AI tools add a cost layer (e.g., software licensing, ongoing operational costs) on top of existing overheads. To justify this expense, AI must generate a return on investment or add quality-adjusted life years (QALYs). In a fee-for-service model, ROI comes from increased volume (working faster); in a value-based model, from cost reduction. In both cases, the efficiency gains required to cover the high cost of AI point toward a reduction in human hours per exam. Historically, productivity gains in healthcare rarely translate into reduced workload. Instead, they expand service capacity and diagnostic volume. AI is therefore more likely to shift radiologists toward supervising larger exam volumes and AI outputs rather than creating additional patient-facing time. If AI merely makes a radiologist 10% more accurate but does not increase their speed or reduce the need for downstream testing, it is a cost-adder, not a cost-saver, unless this leads to quantifiable health gains such as QALYs. For AI to be sustainable, meaning, for hospitals to keep paying for it once the hype cycle fades, it must prove it can do the job cheaper. History suggests that “cheaper” usually involves reducing the most expensive variable in the equation: the human workforce. However, the continuous increase in imaging volumes also implies that simply expecting radiologists to work faster is unsustainable. A more financially viable alternative lies in deliberate task splitting and role separation [8]. By deconstructing the diagnostic workflow into discrete components, AI can be deployed to autonomously handle lower-complexity tasks or even clear “normal” exams. This effectively redistributes the workload, allowing healthcare systems to manage surging imaging volumes without a proportional increase in the human workforce, thereby optimizing the overall cost per case. We are at a crossroads: marketing presents a future of “human-AI partnership” without job losses, yet the economic fundamentals of automation demand efficiency that is typically achieved through labor substitution. In other words, we must stop pretending that “efficiency” is a victimless metric.

Does AI need to substitute radiologists or radiation technologists to achieve economic sustainability? The uncomfortable answer is likely yes, in the form of redistributing expertise and reducing human involvement in routine tasks. Potentially, we will see a future with fewer technologists per scanner and fewer radiologists per study volume, with AI bridging the gap. Recognizing this trajectory should not be interpreted as professional decline, but as a call to realistically reassess how radiologists create value. The profession must prioritize developing expertise in supervising AI systems, integrating imaging into clinical decision pathways, and generating robust evidence of clinical and economic benefit. Ignoring the economic drivers of automation will not prevent workforce transformation; it will only reduce our ability to shape it. The future role of radiologists, and the thriving of the profession, will likely depend less on image interpretation alone and more on how effectively we manage, validate, and clinically contextualize automated diagnostic systems [9, 10].

This article was first published in European Radiology.

Source: The augmentation myth: AI, economics, and workforce substitution in radiology | European Radiology | Springer Nature Link

References