Chapter 6 - Adjacent Professions: Software Development and Law

6.1 What the adjacent professions show

Software developers had the problem first, and lawyers had it next. In both cases, AI mediation has been operating at scale long enough to leave empirical traces: on displacement, on formation, on what institutional responses are available. This chapter reads those traces and develops four findings that Chapter 7’s stage-by-stage analysis then deploys.

The chapter works from published evidence on adjacent professions rather than from lived material in my own practice. I made that choice because the empirical and theoretical literatures on AI in software and law are, at the time of writing, richer than anything comparable on AI in consulting. The first-person voice is correspondingly sparing. 💡

The chapter is also where this work meets its principal analytical foil: the Susskinds’ The Future of the Professions 1 , a substitutionist account whose premises §6.5 answers directly.

It engages, too, the AI-washing critique that has emerged in the labour-market literature on the 2024-2026 tech sector layoffs, which §6.4 takes seriously and answers: the argument of this work is about formation, not headcount.

Software and law are adjacent to professional advisory broadly, not only to European infrastructure advisory. The formation mechanisms, the pipeline rupture risk, and the legitimacy narration requirement operate across knowledge-intensive professions where AI mediation has reached the formative base.

The comparative evidence from software and law contributes to the general register of this work: it supports the claim that the proletarianisation mechanism operates across advisory and advisory-adjacent professions, not only in the specific domain this work examines.

6.2 The software case

The software development profession has been the canary for AI mediation in knowledge work, and the evidence from this case is the richest comparative material available to this work. Three components of the evidence carry the chapter’s argument.

The first is empirical evidence of selective displacement at the level of individual practice. The published literature on AI mediation in software development shows that AI tools (particularly the family of code-generation systems that became widely deployed from 2022 onward) have absorbed certain components of the developer’s work. Routine code generation, boilerplate, basic test writing, code review at the level of style and convention: the historically most repetitive parts of the job now routinely pass through AI tools 2 3 4 . The displacement is selective because it takes some parts of the work and leaves others untouched 5 6 .

The second is empirical evidence of pipeline disruption at the level of formation. The formation pathway junior developers have historically travelled 💡 is being altered by AI mediation. Junior developers now produce “competent” code faster, and the cost is a compressed formative engagement with failure and correction. 💡 The evidence sits in three places: the literature on developer education 7 , senior software engineers writing about their concerns for the next generation 8 , and, by analogy, the cross-professional literature on AI-induced deskilling, which is most developed in medicine 9 .

The most direct evidence on this second component arrived in 2026, from inside the profession’s own career structure. Yu and Moon’s interview study of junior and senior software engineers in South Korea, a market with one of the highest rates of generative AI adoption, names the mechanism absorption: generative AI “redirects entry-level work into senior–AI workflows”, so that the tasks through which juniors once learned are performed by a senior working with a model rather than by a junior working under a senior 10 . Three consequences follow in their data. Juniors lose “the productive struggle through which expertise once developed”; the loss is reproduced upstream, because AI use is collectively normalised in university classrooms before anyone reaches a firm; and a “perceptual asymmetry between seniors and juniors” leaves neither side able to correct the dynamic alone (§7.2 takes it up for advisory).

Their conclusion is the one this work reaches, in another vocabulary: the technology “appears to be absorbing not just specific categories of tasks but also parts of the pathway through which the next generation of seniors is formed”, the pathway “was already unprotected; GenAI has made it consequential”, and its preservation “will require deliberate institutional design” rather than individual restraint 10 . The software case adds the structure of the mechanism: absorption is selective displacement observed at the level of who does the work, and its three consequences map onto the Foundation stage, the pre-entry formation the ChMC framework does not model, and the supervisory relation of Chapter 7.

The third component is labour-market data with the AI-washing complication. Analysts disagree about how much of the 2024-2026 tech sector layoffs is attributable to AI mediation and how much to broader cost-cutting framed in AI terms 11 12 13 . §6.4 takes it up.

6.3 The law case

Law is the second case where AI mediation has been operating at scale for substantially longer than in management consulting, and its evidence complements the software case. It is institutionally distinctive: a heavily regulated profession with chartered-status architecture (bar admission, seniority frameworks, ethical codes) and a long-standing preoccupation with how professional expertise is reproduced across generations.

The evidence again falls into three components 💡 .

The first is theoretical articulation of the formation problem. Lawyers have named the formation problem in terms that anticipate this work’s central concern. The legal-AI literature calls it mediated evolution 14 : the structural alteration of the pathway through which junior lawyers develop contributory expertise. The articulation is richer than software’s, partly because law has institutions (bar admission requirements, mandatory continuing legal education, formal apprenticeship) that make formation visible in a way less regulated professions cannot match.

The second component is empirical evidence of similar patterns to the software case. In law, AI mediation has taken the same kind of components it took in software (research, document drafting, contract review, due diligence) and left the rest: the same selective displacement. The pipeline disruption is similar too. Junior lawyers are now formed differently from their predecessors, and what they are being formed in has been altered by AI mediation.

The third component is institutional-response evidence. Because law is regulated, its institutional responses to AI mediation are visible. Bar associations, law schools, and regulatory bodies in multiple jurisdictions have begun to write explicit governance of AI in legal practice, from disclosure requirements to protections for the formation pathway. These responses show that a regulated profession can take the formation question into governance; Chapter 9’s recommendations build on them.

6.4 The AI-washing complication

The AI-washing complication is this: some firms have framed cost-cutting decisions as AI-driven displacement when the underlying decisions were made for other reasons (cost discipline, say, or interest-rate pressure on technology investment). So labour-market data alone cannot show that AI mediation is producing the displacement attributed to it. 💡

Our response is twofold.

First, the complication is real, and it limits what labour-market data can show. But the analytical chapters do not rest on that data. The claims this work makes about advisory transformation sit at the level of formation, sovereignty, extraction, and political economy, none of which depends on headcount in the consulting profession. Chapter 7 needs this to hold: its stage-by-stage analysis works at the formation level throughout.

Second, the formation evidence stands on its own. Whatever AI-washing does to the layoff figures, the evidence that the formation pathway is being disrupted is untouched: the content junior practitioners are formed in has been altered, whether or not the labour-market data is partly cost-cutting in AI dress. Part of what this work contributes is to hold the two questions apart: what is happening to professional formation, and how much displacement is attributable to AI specifically.

6.5 The Susskinds as foil

Richard and Daniel Susskind’s The Future of the Professions is the analytical foil against which this work develops its formationist account. The Susskinds read the transformation of the professions under AI mediation as substitution: they predict that professional services will increasingly be delivered through AI-mediated systems, with human practitioners displaced from the work and reduced to narrower roles where their contribution is institutionally recognised but quantitatively diminished. The book has been influential in the conversation about AI in professional work.

My disagreement with the Susskinds is about analytical apparatus rather than facts. It has three points.

The first is the unit of analysis. The Susskinds address professional services delivered to clients; I address professional formation across generations. The two units ask different questions about the same situation.

• The Susskinds’ question is: how will professional services be delivered as AI mediation matures?
• Mine is: what is happening to the formation pathway through which professional expertise is reproduced, as AI mediation operates on professional work? The two questions sit at different levels, and neither excludes the other. But the formation question cannot be reached from a service-delivery analysis, and the concerns of this work live at the level the Susskinds do not visit.

The second point is the content of expertise. The Susskinds treat professional expertise as the institutional architecture through which services have historically been delivered, so that AI mediation can replicate the expertise once that architecture is mature enough. Professional expertise is something else: contributory expertise in Collins’s sense 15 , the ability to contribute to the practice through embodied engagement, built through formative engagement with consequence over the arc of a career. That is a different thing from the delivery architecture, and AI mediation cannot replicate it even where it absorbs the architecture wholesale.

The third point is the normative content. The Susskinds advance a neutral account of the transformation they predict: AI-mediated service delivery will be more efficient than human-mediated delivery, and the one evaluative question left open is whether the professions’ institutions adapt to the gain. My account starts from what the work owes, and to whom: at stake in European infrastructure advisory are the answerability of the work to the third box (the publics and future generations its decisions affect), the European public-interest tradition, and the integrity of the formation pathway. These are at risk under contemporary AI mediation, and the question becomes whether the content survives the transformation; institutional adaptation to the efficiency gain does not settle it.

So the Susskinds are not refuted but answered, with an apparatus that works at a different level and reaches concerns the substitutionist account does not. 💡

6.6 Four findings carried forward

Four findings come out of the comparative evidence.

First finding: selective displacement. AI mediation operates as selective displacement of savoir-faire into technical systems. It absorbs routine analytical components, expressive layers of professional output, and formative tasks at the early stages of professional progression. Situated judgment, relational intelligence, engagement with consequence, and the contributory expertise exercised at the senior tiers remain beyond its reach. The selectivity is the proletarianisation mechanism we name: it leaves the practitioner in place and takes specific components of the work. 💡

Second finding: pipeline rupture regardless of headcount. Selective displacement disrupts the formation pathway. The disruption acts on formation, and stable headcount is therefore no protection: software and law both show the pathway being altered while employment levels hold, and what junior practitioners are being formed in has changed. It is also the most robust of the four, because the AI-washing dispute (§6.4) bears on headcount and this finding does not.

Third finding: uneven upward value flow. AI mediation redistributes value within the professions where it operates. The content absorbed at the formative tiers (the routine analytical work, the expressive output, the formation pathway) flows upward in value terms to the senior tiers, where situated judgment remains valuable, and to the owners of the systems 💡 . The redistribution has a direction: it runs upward through the professional hierarchy and outward to the platform owners.

Fourth finding: explicit narration of professional legitimacy. Put the first three together and the structures through which professional legitimacy has historically been recognised come under pressure. Legitimacy that was implicit in expertise, formation, and senior practice no longer suffices under AI mediation. It now has to be narrated explicitly: what is being done, who is doing it, what the AI mediation is doing, and what the practitioner is contributing. Law’s institutional response (bar associations developing AI governance, disclosure requirements, formation pathway protections) shows both that the pressure is real and that responses are available.

These four findings are the chapter’s contribution to the analytical apparatus, and Chapters 7–9 work with them from here on. 💡

Supporting evidence from adjacent fields

The medical literature adds depth to the picture from software and law. Natali et al.’s mixed-method review of AI-induced deskilling in medicine 9 synthesises the evidence across healthcare and identifies specific competency domains at risk from AI-driven decision support: physical examination, differential diagnosis, clinical judgment, and physician-patient communication. These are the formation-through-engagement capacities the tacit-as-formation account predicts will be most vulnerable. 💡

The authors draw a distinction that bears directly on the pipeline-rupture argument. Upskilling inhibition (the foreclosure of skill acquisition) is analytically distinct from deskilling (the loss of already-acquired skills). Upskilling inhibition is the primary concern of this work: the failure of junior practitioners to acquire the capabilities they would have acquired through formative engagement, not the loss of capacities already formed.

The paper’s cross-domain extension licenses its use here: the authors frame their findings as applicable wherever AI decision support displaces practitioner judgment. The PACES-MRCPUK clinical competency framework they deploy structures medical progression through stages, each dependent on formation-through-practice, and serves the same function the ChMC framework serves in consulting. The parallel extends the argument: the formation-pipeline-rupture concern appears wherever a structured competency framework governs professional progression.

A 2026 Perspective in Nature Medicine sharpens the conceptual vocabulary this evidence requires. Ke et al. (2026) distinguish never-skilling (the failure of foundational competency to develop during formative training) from deskilling (the erosion of established competencies in experienced practitioners) and mis-skilling (the internalisation of flawed AI outputs) 16 .

Never-skilling covers the ground Natali et al. call upskilling inhibition, and it carries the same weight for the pipeline-rupture argument: formation that does not occur, as distinct from skills acquired and later eroded. The cohort at risk is the Foundation-stage one, whose formation is disrupted; the experienced senior whose acquired skills are later reduced is a different case. The two risks need distinct responses.

Two further concepts apply directly.

• False proficiency names why the mechanism is invisible to productivity metrics: AI-assisted performance appears adequate during training, masking absent independent competency, until the support is withdrawn and the deficit becomes visible, by which point the formation window has closed.
• The calibration paradox maps onto the Collins interactional/contributory distinction deployed here 15 : effective AI oversight requires an independent cognitive architecture to verify against. A practitioner who has not built that architecture cannot verify; they can only read and accept. Interactional fluency without contributory foundation cannot perform the oversight it appears to perform.

Ke et al. ground these distinctions in learning theory: desirable difficulties, deliberate practice, cognitive load, and the expertise reversal effect 💡 (the finding that AI assistance beneficial to experts may harm novices by substituting for schema formation rather than augmenting it). This provides educational-science grounding for the stage-specific pattern the ChMC framework produces, and for the claim that never-skilling is a structural consequence of formative-stage AI mediation rather than an occasional risk. The authors are explicit about the claim’s status: “Direct evidence for never-skilling in clinical trainees remains absent” 16 . It is a hypothesis grounded in learning theory, not an observed result, and it is in that status that I carry it forward.

Ke et al. also take never-skilling past competence, to how trainees see themselves. Where AI routinely supplies the diagnosis and the management plan, trainees “may come to view themselves as intermediaries who interpret and relay AI outputs”, instead of reasoners in their own right. The authors leave this open too: “Whether this shift in professional identity occurs, and whether it has measurable consequences for clinical performance, is an empirical question” 16 . I give it the same status as never-skilling itself: a hypothesis about who the trainee comes to be, which Chapter 7 takes up for the Foundation-stage consultant.

Audit gives the fourth finding a profession-wide case. Goto’s 2021 study analyses how Japanese auditors described their role as AI arrived, in the national association’s journal and reports, where AI first appears in 2016, and in 42 interviews with the association and the “Big Four” firms’ AI task forces. It finds that “a new collective professional role identity was constructed”, in six themes that recast the auditor as, among others, a “Data-empowered Advisor” and a “Guardian of Capitalism as before”, and that together “framed the audit profession’s future positively” 17 . This is one profession in one country, and the study ends in 2018, before generative AI. It still shows who was narrating: the association and the people leading adoption, describing their own future. The optimism has a formation detail inside it. Long hours of manual checking had been a sign of auditors’ dedication, but a partner observed that “junior staff [were] becoming increasingly doubtful about the value of time-consuming manual work”, and the new identity offered to remove it 17 . That work was also how auditors had been formed, and the new identity says nothing about what replaces it.

The four findings also have theoretical support from a different direction: the formal model of Acemoglu, Kong and Ozdaglar set out in Chapter 5 (§5.6) 18 .

Their model formally derives what the cross-professional evidence here records descriptively. Displacement is selective, targeting the production of expertise rather than its use. Pipeline rupture operates on knowledge reproduction independently of headcount. And the value flows named in the third finding are robust features of how AI mediation interacts with knowledge-intensive work.

6.7 Four findings, one picture

Read together, the four findings are one mechanism seen from four sides. Selective displacement is the mechanism itself: AI takes the routine, the expressive, and the formative, and leaves judgment. Pipeline rupture is what that displacement does to formation, and it happens whether or not anyone is laid off. Uneven value flow is what it does to distribution, upward through the hierarchy and outward to whoever owns the systems. Explicit narration is what it does to legitimacy, which can no longer be assumed from expertise, formation, and seniority. Software showed the mechanism first; law has articulated it most fully and begun to govern it; medicine, audit and formal economics now show it from outside the two professions this chapter set out to compare.

Neither of the two objections engaged above reaches this level. The AI-washing dispute is about the layoff figures, and the picture does not rest on them. The Susskinds’ account is about service delivery, and from there pipeline rupture, sovereignty risk, and third-box answerability are simply not visible.

Chapter 7 carries the four findings into the stage-by-stage core of the analytical work: how each manifests within the three ChMC stages of European infrastructure advisory specifically.