Artificial Idea | AI careers · practical prompts · no hype Monday, December 1, 2025 · Issue #35 · Jobs

The year in review

Every prediction about 2025 was made in 2023. Here is what the year looked like from inside it, with the data that was not available when the predictions were made.

In January 2025, the predictions were already in circulation. AI would eliminate 85 million jobs. AI would create 97 million new ones. The net would be positive. The transition would be painful. The workers most affected would be those least able to absorb the disruption. The workers best positioned would be those who had already invested in AI fluency. The timeline was accelerating. The window for preparation was narrowing.

Most of those predictions were made in 2023, extrapolated from trends visible at that time and projected forward through modelling that assumed relatively linear adoption curves, relatively predictable technology development, and relatively rational organisational responses to both.

2025 was not linear, not entirely predictable, and not rational in the ways the models assumed. It was a year in which significant things happened that the optimistic scenarios did not anticipate and significant things failed to happen that the pessimistic ones predicted with confidence. Understanding what actually occurred, with the data available now that the year is closing, is more useful than the predictions were, because it is grounded in what happened rather than in what the models expected.

What got cut

The displacement that occurred in 2025 was real, substantial, and concentrated in ways the aggregate figures obscure.

The sectors with the largest absolute job losses attributable to AI adoption were, in order: financial services back office operations, content production and editorial, customer service and support, legal document review, and software development at the junior level. Collectively these sectors accounted for an estimated 2.3 million roles eliminated or significantly restructured in 2025 across the G20 economies, according to the McKinsey Global Institute's December 2025 workforce report.

Two things about that figure are worth holding simultaneously. It is a large number representing real disruption to real people. It is also approximately 0.7% of the total G20 workforce, not the civilisational-scale displacement the most alarming 2023 predictions described. The disruption is concentrated, which means it is severe for those in the affected roles and less visible than its total scale suggests to those outside them.

The concentration pattern within those sectors is also instructive. In financial services, the losses were almost entirely in roles whose primary function was processing structured data according to defined rules: trade confirmation, routine compliance checking, standard report generation. Roles requiring judgment about non-standard situations, client relationship management, and regulatory interpretation were largely untouched and in some cases grew as the organisations that automated their processing functions redirected investment toward the judgment-intensive work that automation revealed was the actual source of their competitive differentiation.

In content production, the losses were concentrated in templated, SEO-driven content at scale: the product description farms, the programmatic news summaries, the formulaic blog content that had already been in tension with quality standards before AI made the economics of human production untenable. Long-form analysis, original reporting, and content with a distinctive human voice contracted less and in some subsectors grew, as audiences increasingly distinguished between content produced at scale and content worth reading.

The pattern across all five affected sectors was the same one this newsletter has described since Issue #3: AI replacing tasks within roles more than replacing roles entirely, with the roles that survived being those whose remaining tasks after automation were the judgment-intensive, relationship-dependent, variable-context work that AI handles poorly.

What got created

The job creation side of 2025 was more diffuse and harder to measure than the displacement side, which is one of the reasons the displacement narrative dominated the coverage. Displacement happens at identifiable organisations and produces announcements. Creation happens across millions of workplaces in forms too granular and too varied to generate equivalent news coverage.

The most reliable estimate of net job creation attributable to AI in 2025 comes from the Brookings Institution's December analysis, which synthesised data from twelve separate labour market studies conducted across the year. Its central estimate is that AI-related activity created approximately 1.8 million net new roles in G20 economies in 2025, offset against the 2.3 million displaced, for a net displacement figure of approximately 500,000 roles.

That net figure sits at approximately 0.15% of the G20 workforce. It is not the catastrophic net displacement the pessimistic scenarios predicted. It is also not the net positive the optimistic scenarios promised. It is a transitional figure consistent with a technology adoption curve that is real, significant, and still early in its trajectory.

The 1.8 million created roles are worth examining by category. The largest single category, accounting for approximately 640,000 roles, is what the Brookings analysis describes as AI-augmented professional roles: existing professional functions in consulting, marketing, finance, law, and healthcare that have been sufficiently transformed by AI tool adoption to constitute new role definitions with different skill requirements and typically higher compensation than the roles they evolved from. These are not new job titles. They are existing job titles that now require AI fluency as a core competency, and the professionals filling them are earning the premium the Burning Glass data described in Issue #33.

The second largest category, approximately 420,000 roles, is AI operations and governance: the growing population of professionals whose function is to manage, evaluate, audit, and maintain the AI systems that organisations are deploying at scale. These are roles that did not exist in their current form three years ago and that the technology sector underinvested in creating, producing the governance gaps that regulators in multiple jurisdictions are now addressing through mandatory requirements that are themselves creating more roles in this category.

The third category, approximately 380,000 roles, is AI-adjacent creative and strategic work: the content strategists, brand professionals, creative directors, and strategic advisors whose value has increased as AI commoditised the execution layer of their functions and elevated the premium on the judgment and originality that distinguishes their work from what the tools produce. This category surprised the researchers who tracked it. The 2023 predictions were almost uniformly pessimistic about creative roles. The 2025 data shows a more nuanced picture in which the most skilled creative professionals are in higher demand than before, precisely because AI has made the gap between their work and commodity content more visible rather than less.

What surprised everyone

Three findings in the 2025 data were sufficiently inconsistent with the dominant predictions to warrant specific attention.

The first is the resilience of middle management. The 2023 and 2024 predictions about middle management were almost uniformly pessimistic, driven by the logic described in Issue #17: AI tools automating the coordination and reporting functions that constituted a large portion of many middle management roles. The 2025 data shows that middle management headcount in large organisations declined by an average of 6%, less than half the 15% reduction that the 2024 Gartner survey of HR leaders had projected. The explanation offered by researchers across multiple studies is consistent: organisations that attempted to eliminate management layers discovered that the judgment, development, and cultural functions that good middle managers perform were not as easily automated or redistributed as the coordination functions, and that the organisations that moved fastest on management reduction experienced the talent development and cultural cohesion problems that follow when those functions are inadequately performed.

The lesson the data draws from this is not that middle management is safe. The 6% reduction is real and the pressure continues. It is that the organisations that managed the transition best were those that redesigned management roles rather than eliminating management layers, keeping human judgment in the functions where its value is highest and automating the functions where the value was always lower than the time cost suggested.

The second surprise is the pace of AI skill diffusion. The 2023 predictions assumed a relatively long lag between AI tool availability and broad professional adoption, based on historical patterns of technology diffusion in professional contexts. The 2025 data shows that diffusion has been significantly faster than those models predicted, with AI tool usage becoming standard rather than exceptional across most large organisation professional functions within approximately eighteen months of meaningful tool availability in each sector.

The practical implication of faster diffusion is that the window during which AI fluency represents a genuine differentiator is shorter than the 2023 predictions suggested. The professionals who invested early are ahead. The window during which they remain differentiated by that investment alone, without continued development, is narrowing as the tools become ubiquitous and the baseline expectation rises. The premium is shifting from early adopters toward those who have gone deepest, not just first.

The third surprise is geographic. The predictions about AI's impact on the Indian technology services sector were predominantly pessimistic, focused on the automation of the process-oriented work that constitutes a significant portion of India's business process outsourcing revenue. The 2025 data shows a more complex picture. The lower end of the BPO sector, high-volume, low-judgment process work, has experienced the displacement the predictions anticipated. The higher end, complex analytics, specialised professional services, and domain-specific consulting, has grown at a rate that has partially offset the lower-end contraction and that shows a trajectory the pessimistic models did not capture.

The Indian technology sector's aggregate employment declined by an estimated 3.2% in 2025, against predictions ranging from 8 to 15%. The decline is concentrated at the lower end. The growth is concentrated at the higher end. The transition is producing a bifurcation within the sector that mirrors the broader labour market bifurcation this newsletter has been describing all year: those positioned at the intersection of domain expertise and AI fluency are growing. Those positioned in the task categories AI handles most effectively are contracting.

What 2026 looks like from here

The trajectory the 2025 data describes suggests three things about 2026 that are worth holding as context for the second half of this newsletter.

First, the pace of displacement will likely accelerate modestly rather than dramatically. The technology is improving and adoption is broadening, but the organisational change management requirements that limit how fast AI integration can proceed without creating the operational and cultural problems that fast movers discovered in 2025 are a genuine constraint on the pace of change.

Second, the salary premium for AI-fluent professionals will continue to grow in the specific functions where domain expertise and AI fluency interact most productively, and will compress in the functions where AI tool usage is becoming baseline rather than differentiated. The professionals who invest in depth will maintain and grow their premium. Those who invested early but stopped developing will find their early advantage eroding.

Third, the governance and oversight requirements for AI deployment are going to create a significant and underappreciated source of professional demand in 2026, as regulatory frameworks in the European Union, India, and increasingly the United States impose requirements on AI-deploying organisations that require human professionals with specific combinations of technical understanding, domain expertise, and compliance knowledge to fulfil. This is the job creation category this newsletter has covered least and that deserves more attention in the issues ahead.

The action

Review the five career decisions you made in 2025 that were most influenced by the AI transition narrative. For each one, assess whether the decision was made on the basis of the prediction that was circulating when you made it or on the basis of the data that was available.

The gap between those two bases for decision-making is the gap this newsletter exists to close. The predictions are always made with less information than the data that follows. The professionals who wait for the data before acting are always later than those who acted on the predictions. The professionals who act on accurate predictions are ahead of both.

Understanding the difference between an accurate prediction and a confident one is the analytical skill that determines which category you end up in.

Thursday is the last issue before the end of year period. It is also, not coincidentally, one of the most practically useful issues we have published: the prompt framework for building the 2026 AI toolkit, the specific set of capabilities and practices that the 2025 data suggests will produce the highest return in the year ahead.

The year told us something. Thursday explains what to do with it.

— The Artificial Idea team