In 1995, I moved jobs from Gwent Tertiary College, a large vocational school in Wales, to take up a post as a researcher at the Institut Technik und Bildung (ITB) at the University of Bremen in Germany. I had never worked in a university before and was excited by the prospect of being a 'proper' researcher. More than that, I was deeply impressed by the central tenet of ITB's work: the social shaping of technology. At that period, there were growing worries in Germany about the impact of new technologies on skilled work, particularly in the engineering industry.

Some of my initial investigations were into Computer-Aided Design and Computer-Aided Manufacturing (CAD/CAM). There were widespread concerns that the leading German CAD/CAM machinery would be replaced by USA-produced automated systems. Comparative studies undertaken by ITB, notably under the leadership of Professor Felix Rauner, told a different story. Productivity remained higher in Germany because of the highly skilled workforce, as compared to the relative deskilling associated with fully automated CAD/CAM machines. It also became apparent that new technologies faced major problems due to the need for changing work organisation, and that simple automation was not the answer. Rather, there needed to be a human-driven approach, conceptualised in Rauner's influential work as human-centred Computer Integrated Manufacturing (CIM) systems (Rauner, F., Rasmussen, L., & Corbett, J. M. (1988)). The educational goal that emerged from this work was captured in the German concept of Technikgestaltung—the idea that vocational education should equip workers not merely to operate technology, but to actively shape and design it (Rauner, F. (1988)).

This discussion is not unlike the one taking place now about Artificial Intelligence (AI), with the European Union increasingly pursuing a human-first approach to implementation. I still believe that the social shaping of technology is a key analytical lens. Sadly, much of the research into the impact of AI has relied on task analysis - asking how many tasks AI could theoretically perform—and has failed to engage with wider social issues, including work organisation, accountability, and the nature of tacit knowledge.

I was therefore very impressed by a recent newsletter article, Why AI hasn't replaced software engineers, and won't: Coding agents as normal technology, published on 11 June 2026 by Arvind Narayanan and Sayash Kapoor. Their analysis challenges the dominant narrative of AI-driven job displacement and aligns, perhaps more than they realise, with the principles of work process knowledge and the social shaping of technology that I encountered at ITB Bremen three decades ago.

The Myth of AI-Driven Mass Layoffs

Narayanan and Kapoor examine the software engineering profession, where AI capabilities are currently the most advanced and adoption has been exceptionally rapid. They argue that there is sufficient evidence to reject the narrative that reaching a certain threshold of AI capability will automatically cause mass layoffs. Their examination of high-profile cases - from fintech firms to social media companies - reveals a consistent pattern they call "AI washing": the use of AI as a convenient public justification for job cuts that are in reality driven by financial pressures, investor demands, or corporate restructuring.

They cite empirical evidence to support their theory. A survey found that 59% of U.S. hiring managers admitted to emphasising AI when explaining hiring freezes or layoffs because it plays better with stakeholders than citing financial constraints (Mousa, D. (2026) A Forrester analyst noted that when companies preparing supposedly AI-driven layoffs were asked whether they had a mature, vetted AI application ready to fill the vacated roles, the answer was no in nine out of ten cases. Furthermore, data from New York State's WARN Act filings - the first U.S. jurisdiction to require companies to disclose AI as a reason for layoffs - revealed that in the first full year of reporting, only one company out of over 160 that filed notices checked the AI box (Narayanan and Kapoor, 2026).

This resonates with the findings from ITB in the 1990s. Just as the simple automation of CAD/CAM failed to deliver productivity without the deep, tacit Arbeitsprozesswissen (work process knowledge) of skilled German engineers, the effective use of AI in software development relies heavily on the contextual and organisational understanding held by human workers. Firing existing workers, as Narayanan and Kapoor observe, results in the loss of precisely the tacit knowledge and organisational capital that allows workers to operate AI effectively.

The Decide-Execute-Deliver Sandwich

To explain why coding agents have not led to labour displacement, Narayanan and Kapoor (2026) introduce the concept of the "decide-execute-deliver sandwich". They point out that writing code - the execution phase - has never been the primary bottleneck in software development. A 2019 Microsoft Research paper summarising existing studies found that developers spend surprisingly little time actually coding, ranging from 9% to 61% of their working time depending on the study ((Narayanan and Kapoor, 2026). When coding agents began to be widely adopted, developers discovered that having AI write most of the code had little impact on overall productivity, because code generation was never where the real work resided.

The real bottlenecks are found at the two ends of the sandwich: first, deciding and specifying what to build, which requires a deep understanding of user needs, market signals, organisational priorities, and regulatory constraints; and second, verifying and delivering what has been built, which involves accountability for the final product and ensuring that mission-critical software is fully tested and understood. AI has compressed the middle "execute" layer considerably - one study found an eight-fold increase in lines of code generated by AI agents - but this translated into only a 30% increase in software releases, strongly suggesting that the human bottlenecks at the "decide" and "deliver" ends remain firmly in place (Narayanan and Kapoor, 2026). This model is a precise contemporary illustration of the concept of Technikgestaltung advocated by Rauner and his colleagues. The goal is not to automate the human out of the loop, but to enable people to shape technology and their work processes. As Narayanan and Kapoor put it, a central insight of viewing AI as normal technology is that "we can collectively choose to keep humans accountable through shared norms, law, and policy". The software engineer's role evolves - perhaps becoming analogous to that of a crane operator, where AI does much of the cognitive heavy lifting, but the human supervises and remains in control - but it does not disappear.

Implications for Vocational Education and Other Professions

The insights from software engineering have direct implications for other professions, including teaching and research. If the profession most exposed to AI is not facing mass displacement, due to the necessity of the "decide" and "deliver" layers, we can expect similar resilience in roles that require high levels of human interaction, contextual judgement, and accountability. A teacher deciding what and how to teach, and then verifying that learning has actually occurred, occupies precisely the same structural position in the educational process as a software engineer deciding what to build and delivering a working product. The middle layer - the execution of instruction - may be increasingly supported by AI tools, but the ends of the sandwich remain irreducibly human.

As Deena Mousa has argued, broad, economy-wide analyses based on theoretical "AI exposure" metrics are superficial and there is a critical need for careful, occupation-specific analysis (Mousa, D., 2026). The history of the CAD/CAM debate in Germany is an illustration of why this matters. At the time, aggregate analyses suggested that automation would displace skilled manufacturing workers. The reality, as ITB's comparative research demonstrated, was far more nuanced: productivity and quality depended on the preservation and development of skilled work, not its elimination.

Whether we are discussing the introduction of CAD/CAM in the 1990s or generative AI in the 2020s, the fundamental lesson remains the same. Technology is not an autonomous force that dictates the future of work. It is, in Narayanan and Kapoor's phrase, a "normal technology" that we can and should collectively choose to shape through shared norms, policies, and a commitment to keeping humans in control of the processes that matter most. The social shaping of technology is not a relic of 1990s German industrial sociology. It is, if anything, more urgent now than it was then.

References

Mousa, D. (2026, May 19). The leading indicator graveyard: Nobody knows what 'AI exposure' means. Under Development. https://newsletter.deenamousa.com/p/the-leading-indicator-graveyard

Narayanan, A., & Kapoor, S. (2026, June 11 ). Why AI hasn't replaced software engineers, and won't: Coding agents as normal technology. AI as Normal Technology. https://www.normaltech.ai/p/why-ai-hasnt-replaced-software-engineers

Rauner, F. (1988 ). Die Befähigung zur (Mit)Gestaltung von Arbeit und Technik als Leitidee beruflicher Bildung. In Gestaltung von Arbeit und Technik: Ein Ziel beruflicher Bildung. Gesellschaft zur Förderung arbeitsorientierter Forschung und Bildung.

Rauner, F. (2024). On the constitution of a new educational idea: "Enabling people to shape technology." In F. Rauner (Ed.), Handbook of fundamentals of modern vocational education (pp. 199–208). Springer. https://doi.org/10.1007/978-981-97-0987-8

Rauner, F., Rasmussen, L., & Corbett, J. M. (1988 ). The social shaping of technology and work: Human centred CIM systems. AI & Society, 2(1), 47–61. https://doi.org/10.1007/BF01891434

---

 

 

 

 

 

[9]Mousa, D. (2026, May 19). The leading indicator graveyard: Nobody knows what 'AI exposure' means. Under Development. https://newsletter.deenamousa.com/p/the-leading-indicator-graveyard