Every previous wire we discussed in Chapter 31 reorganized economic value. The telegraph hollowed out independent local merchants; the telephone built AT&T; the internet eviscerated newspapers and built Amazon and Google; the smartphone rewrote retail and ride-hail. The fifth wire is doing its own version of this rerouting now, and the realistic time to look at it is while it is happening, not after. This chapter is the inventory: what is moving, who is losing margin, who is gaining it, and what plausibly happens next. Tone: realistic. No optimism-washing, no doom-washing.
The frame: where value moves, not whether
The mistake commonly made in AI economic discourse is treating the question as whether AI creates or destroys value. It does both, simultaneously, in different places, and the more useful question is where the displacement and accumulation occur. The pattern is consistent across general-purpose technologies: the inputs to the process get cheaper, the outputs get more available, the producers of complementary capabilities accumulate, and the producers of substituted labour lose. Looms displaced weavers and enriched mill owners; spreadsheets reduced accounting clerks and made financial analysts more productive; AI is doing the analogous thing.
What is different — and worth being honest about — is the breadth and the speed. Looms substituted weaving; AI substitutes a much wider set of cognitive tasks, on a faster timescale. Whether that produces a soft transition or a rough one depends on policy choices that have not been made, on geographic luck, and on individuals' ability to retool. Pretending the transition is gentle, or that everyone affected can simply learn to prompt their way to the next job, is not realism.
Labour: the routine cognitive layer
The category of work most directly affected is what economists call routine cognitive labour: tasks that follow patterns, can be specified in writing, and produce outputs that are themselves text or structured data. Customer-support L1 work, basic legal document review, basic accounting tasks, content moderation, routine report drafting, basic translation, low-end copywriting, formulaic financial analysis, much of what entry-level professionals in many fields actually do day to day.
The deployment data in 2026 is consistent with what models predicted. Customer support: vendor case studies and industry surveys put the share of tickets resolved without human escalation in a wide band — commonly 10–40% at large deployments, climbing past 50% on narrow product-support corpora — with the share rising in every cohort that publishes follow-ups. Legal document review: junior associate hours on contract review are down measurably at firms that have adopted AI assistance. Software engineering: AI-assisted coding has reduced the time-to-first-PR for new hires and pulled some entry-level work into the work AI does directly. BLS employment data and Indeed Hiring Lab tracking both show entry-level cognitive jobs being added at slower rates than non-cognitive ones, reversing a multi-decade trend.
What this means concretely: people in these roles are not all losing their jobs at once, but the on-ramp is narrowing. The ten new hires this year are five. The five-year-experienced person doing one task is now doing three. The ladder up exists but is harder to climb because the bottom rungs are partially occupied by software. This is not a forecast; it is happening.
Software and attention margins
Software-as-a-service margins are getting compressed in the middle and expanding at the edges. Compressed in the middle because AI commoditizes a lot of what used to be sold as software: text drafting tools, basic analytics, content management. The market is not vanishing, but pricing power is. A product that was sold for $50/seat/month in 2022 is being undercut by one with the same functionality at $5/seat/month or by an in-house implementation built in a weekend with an AI assistant.
Expanding at the edges because the products that integrate AI as their core capability — coding assistants, design tools, research platforms — are commanding premium prices because their value proposition relies on something the customer cannot easily build themselves. The new high-margin software is the AI-native product; the squeezed software is the boring SaaS.
Attention markets — search, advertising, content discovery — are in the middle of their largest reorganization since the launch of Google. AI-mediated retrieval (Perplexity, Google's AI Overviews, ChatGPT search) reduces the value of being a destination because users get answers directly. Publishers that depended on referral traffic from search are seeing their unit economics erode, and many are licensing content to model providers as the new revenue path. The realistic 2026 picture is a lot of long-tail publishers in serious distress and a handful of large brands negotiating direct licensing deals worth hundreds of millions a year.
Knowledge work, sliced finely
The temptation is to declare "knowledge work is automated." The reality is that most knowledge work is a bundle of subtasks, of which some are highly automatable and others are not. A doctor's job is partly information retrieval (highly automatable), partly diagnostic reasoning (partially automatable), partly procedural skill (not automatable in this generation), partly relationship management with patients (not automatable in this generation). The honest forecast is not that doctors disappear; it is that doctors do less of the tasks that were previously the apprenticeship for doctors, which changes both the job and the path into it.
The same fine slicing applies in law, finance, consulting, journalism, design, and dozens of other fields. The high-skill workers do not get replaced wholesale; the routine fraction of their work shrinks and they spend more of their time on the residual. Productivity per worker rises where the residual is genuinely valuable; it falls where the residual was actually padding all along, and those workers eventually face a different kind of question.
Who keeps what
Calling the spade a spade, here is the realistic accounting at end of 2025, with the usual caveat that forecasts more than three years out are not honest.
- Big winners: The chip and packaging supply chain (NVIDIA, TSMC, ASML, HBM vendors). The hyperscaler clouds. The frontier-model owners. Skilled workers in fields where AI complements rather than substitutes, particularly senior engineers, advanced scientists, and high-end specialists whose work requires judgment AI cannot yet imitate.
- Modest winners: AI-native product companies in well-defended niches. The routing and integration layer. Workers who have moved up the value chain by using AI to multiply their throughput. Some emerging-market workers whose costs were already low and whose access to AI tooling is now equal to wealthier counterparts.
- Squeezed: Mid-tier SaaS companies whose products AI commoditizes. Mid-career workers in routine cognitive jobs without a clear up-skill path. Long-tail content publishers reliant on search referral. Independent agencies whose value was in production capacity that AI now delivers cheaper.
- Material losers: Entry-level workers in highly automatable cognitive professions. Workers in jobs that were already at the margin of routine and creative, where AI tips the balance toward routine. Professions whose moat was the cost of training a human to do them, where AI now does them at marginal cost.
The fairness question — whether the gains are distributed in any sense proportional to the losses — is political, not technical. The technical layer makes the gains possible; it does not allocate them. That allocation is an open question, and the institutions designed to answer it (labour markets, education systems, social safety nets, antitrust) are responding at their characteristic pace, which is several years to a decade slower than the technology. The gap between the speeds is the actual policy challenge of this decade.
The book is almost over. There is one more chapter, and it is short, and it tries to say what all of this is, finally, when the parts are seen as one thing.