How should AI systems evolve alongside human work practices?

Parker, Ballard, Billinghurst, Collins, Dollard, Griffin, Johal, Jorritsma and colleagues (Australian Journal of Management, 2025) argue that sociotechnical systems (STS) theory — which holds that "the social aspects of work (e.g. leadership, culture, task allocations) and technical aspects of work (e.g. AI, robots) need to be jointly optimised to achieve quality work" — must be expanded into a "co-evolving sociotechnical systems" (CeSTS) approach to keep pace with AI-driven "radical change in work." They name the failure mode they are responding to a "technocentric fallacy": the assumption "that technology itself is the primary driver of successful digital transformation," which "fails to give sufficient attention to human and work issues and that threatens to undermine the quality of human work." As a cost of the fallacy, they cite that "IBM Watson recorded an 80% failure rate in healthcare applications, with little or no positive impact on care and billions of dollars wasted, largely due to an inability to integrate the technology into the existing work systems and culture."

CeSTS extends the "human in the loop" idea along two axes instead of treating technology as a fixed, already-deployed thing. First, a time axis: STS thinking should enter at the design and commissioning phase, and "consider technology design, development, and use as co-evolving and dynamically affecting each other," because even stable deployments see "workers engaging in multiple work arounds, adaptions, and improvements" long after rollout. Second, a levels-of-analysis axis: equal consideration of technology design itself, individual workers (cognition, physiology, personality), work teams (knowledge sharing, trust, algorithmic management of team processes), organizations (psychosocial safety climate, finance, size), and societal systems (regulations, industry, union density, labour markets). The authors argue prior sociotechnical approaches "have not achieved hoped-for outcomes at scale" partly because they failed "to embed the ideas into wider power structures and knowledge systems," while corporate stakeholders "tend to align with a technocentric agenda in which the emphasis is on replacing human labour and efficiency."

CeSTS's multilevel, co-evolving lens sits alongside other organization-design framings of human-AI relationships already in the library. How do AI novelty and scope shape employee relationships? is itself part of the sociotechnical tradition the paper cites and extends, but where Makarius's typology classifies a given AI system once, by novelty and scope, CeSTS argues the system and the classification both keep changing and need tracking at every level simultaneously. CeSTS's concern with human autonomy and accountability over opaque, self-learning systems runs alongside Can we separate reliability, safety, and trust into distinct mechanisms?'s separation of technical reliability from independent oversight, and alongside What UX principles do workplace users want in AI agents?'s finding that control ranks first among workplace agent design principles — CeSTS supplies an organizational-level argument for why that ranking matters. Its societal-level worry that accountability dissolves as "the line between AI development and use increasingly blurs" (citing Grote et al., 2024) points in the same direction as Does incremental AI replacement erode human influence over society?, though CeSTS's proposed remedy is decentralized stakeholder governance and legislation (citing Australian psychosocial-risk law reforms) rather than alignment mechanics.

The excerpt is a conceptual position paper, closing with the hope that its perspective "helps scholars to move forward on this crucial agenda," not a study that tests CeSTS against any organization or measures the five levels empirically; its supporting examples (IBM Watson, algorithmic healthcare cuts) are drawn from prior literature the authors cite, not new data they collected. The paper does not specify how researchers should weight or sequence the five levels when they conflict, so CeSTS reads as a research orientation rather than an operational method — applying it to resolve a specific AI deployment decision would be an inference beyond what the excerpt establishes.