Associate Professor Jake Goldenfein
Melbourne Law School
ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S)
Associate Professor Jake Goldenfein makes the case for evaluation measures as part of AI procurement decisions, and that workers should be involved in this ‘evaluation design’.
The impact of AI on Australian workplaces is difficult to predict and measure. Australian tech firms have recently announced considerable job losses that they attribute to the adoption of AI. Professional and financial service firms have also cited the use of AI as the reason for retrenchments or reduced recruitment intakes. Recent figures show rising unemployment at junior levels, with the finger pointed at the automation of lower-level clerical and administrative work. Some analysts suggest this is just the start. Outspoken evangelists like Matt Schumer predict almost total job automation, arguing that if your work involves a computer screen, it will soon be performed by software.
There is, however, a competing narrative. An alternative view is that companies are ‘AI washing’ their job cuts. Even conservative media outlets like the Australian Financial Review have stated that rather than job losses arising from the adoption of AI they may be due to bad management, with large firms correcting for years of over hiring and managing broader sector rationalisation. Some argue that announcing job losses due to AI is strategic market signalling – an attempt to attract investment on the back of shifting organisational priorities from labour to capital.
Does AI increase productivity?
Establishing the truth is complex if not impossible. On one hand, large AI companies must make outlandish claims about the end of human work because their absurd capital investment in AI infrastructures can only be recouped by redirecting global money flows away from human workers and towards techno-capital. On the other, many organisations seem content to send money to Anthropic instead of human engineers. But do we really think there’s 20,000 less jobs in Australia in 2026 because of Claude and CoPilot? More importantly, how would we know?
Currently, it is impossible to validate organisational claims that productivity improvements from AI are leading to a reduction in the number of workers needed. This is not only because these companies are opaque, but also because nobody knows how AI is working. There are some experimental studies, based on simulations, suggesting productivity improvements for individuals performing information work tasks. Microsoft has attempted to extrapolate these into organisational figures. The Australian Government has even suggested it has achieved productivity uplift in public sector work, based on worker intuitions. However, there are also studies that suggest that worker intuitions about AI productivity benefits are typically wrong. And that even if individual tasks are performed more quickly, this isn’t translating into organisational productivity benefits. Very few of these studies looking at productivity gains take any consideration of software costs, and whether simply hiring more staff would achieve similar outcomes.
These contradictory results raise real questions about whether organisations should have to meaningfully demonstrate the effects of AI and automation on workplaces to justify reductions in workforces, and whether collective bargaining around AI should include clauses about evaluating AI productivity claims as part of worker participation in AI-related decision-making. The current problem, however, is that any claim about AI and productivity is difficult to validate and would take considerable institutional investment.
Victorian public sector: a case study
A group of researchers from University of Melbourne and University of Sydney, including myself, have been studying AI procurement and deployment in the Victorian Public Sector (publication forthcoming). The project has involved interviewing senior staff and bureaucrats in the Victorian Public Sector as well as software vendors about AI procurement, focusing on what it is used for, how it changes workflows within organisations, and how regulatory compliance is managed. During that research we learned a great deal about the impact of AI on workers and the way it is changing public sector work. Almost every discussion included comments about the gap between high-level productivity narratives that have been driving AI uptake, and the realities of trying to realise productivity benefits in practice.
Is the investment in AI worth it?
Figuring out whether you can use AI to achieve a 0.2FTE saving on a particular workflow requires massive organisational investment. Organisations have very little understanding, or capacity to outline what productivity improvements might look like from an AI perspective, or how to measure productivity and the impact of AI. This is particularly true if including weighing up productivity gains against upfront and ongoing software maintenance costs. We heard from software vendors that most organisations are unclear about what they hope to achieve through AI tooling, and they almost never have a vision for evaluating whether those goals are being met.
The way forward
What seemed to be missing from AI procurement generally was ‘evaluation design’. That is, going through a process to clarify the rationales for AI deployment, and developing mechanisms for auditing and assessing system performance against those goals, over time. This is not an exact or objective science. Organisations need to be thinking hard about their particular software needs rather than acceding to a generalised productivity story that translates increased software spend into vague and generic efficiency promises that inevitably lead to job losses for no good reason.
Involving workers in evaluation design
There is an important role for workers getting involved in designing evaluations of AI tooling in workplaces. Making evaluation design part of a consultation process may be a way to demand both managerial honesty at a high-level, but also provide a mechanism for ongoing engagement with technological workplace transformations. This is an opportunity for workers, through consultation, to influence how AI evaluations take place as a way to ensure those evaluations consider broader impacts on workers. More specifically, workers should be involved in designing the metrics used to evaluate whether AI is creating a productivity benefit, and the ways that AI software is changing the nature of work within their organisations. Tracking metrics like time savings versus costs, exposes the ways that AI results in the movement of money from workers to global technology firms without generating real cost savings or benefit. Defining what productivity means and ensuring it is assessed over time should be considered a critical mechanism for workers to exert meaningful control over AI-related workplace reorganisations.
This article is part of a suite of papers written following the symposium ‘Which Way Forward? AI and Decent Work’. Read more here.