What is different about AI

The routinization hypothesis and the job-polarization literature built on the claim that tacit knowledge cannot be codified, therefore work requiring tacit judgment is automation-resistant. Foundation models trained on large corpora substitute for tacit pattern recognition without explicit codification. The substitution occurs precisely on the categories of work the literature classified as non-routine cognitive - drafting, summarisation, diagnosis, customer-facing communication.

The empirical predictions of the polarization framework reverse. The framework predicted that automation would continue to displace middle-skill routine work while protecting non-routine cognitive work at both ends of the skill distribution. Foundation models do not respect that partition. The categories of work most exposed are precisely the categories the literature classified as safe. This is not a marginal correction; it inverts the maintained hypothesis. The implication for the wage-distribution dynamics that drove the polarization literature is that the high end stops being protected and the distribution compresses, rather than continuing to hollow at the middle.

• Routinization (ALM) Core

  The routinization hypothesis held that work needing tacit, judgment-rich pattern recognition cannot be codified, and therefore cannot be automated. As foundation models trained on large corpora extend their capability, more of that pattern recognition is becoming substitutable: drafting, summarisation, diagnosis, customer-facing communication.

• Job Polarization Core

  The job-polarization literature documented the hollowing of middle-skill work on the assumption that non-routine cognitive work at the top of the distribution was protected. As foundation models substitute more of that work, the wage distribution is shifting from middle-hollowing toward top-compression.

• Social-Skill Complementarity Core

  The social-skill complementarity hypothesis assumed conversational interaction was a stable human moat - the part of work where soft skills earn a premium. As conversational LLMs improve, they are handling more of that interaction at scale - support, sales-development, coaching first-drafts - eroding the moat from inside the bundle.

• Task Model of Automation Partial

  The task model of automation assumes a stable partition between automatable and non-automatable tasks. As foundation models gain capability, the partition keeps shifting into work previously classified as non-automatable - making it a moving variable, not a fixed parameter of the model.

• Augmentation Partial

  As the capability frontier extends, AI increasingly takes over more of the judgment role assigned to the human in centaur-style augmentation theory (human leads, AI assists).

Prior automation technologies were specified and deployed at a fixed capability state. The labour-economic models that analyse their effects - the task model of automation, GenAI productivity measurement, and the long-run dynamics in the job-polarization literature - assume a stationary capability frontier. Foundation models exhibit continuous capability gain on a six-to-twelve-month cycle, driven by scale, training improvements, RLHF, and architectural advances. The frontier is not a stationary state.

Any empirical study of AI labour effects is undertaken against a moving target. Productivity studies report on a frontier that has materially shifted by publication. Firm-level optimization is over a frontier that no longer exists by implementation. The standard methodology of the labour-economic literature - empirical estimation at a fixed point - is mismatched with the phenomenon. Centaur-style augmentation theory assumes a stable human-AI balance that asymmetric capability gain prevents from holding. The Reshuffle thesis treats the trajectory of the constraint, not the snapshot, as the relevant unit.

• Task Model of Automation Core

  The task model of automation, in its firm-level optimization form, assumes the capability frontier is stable across the optimization horizon. Six-to-twelve-month capability cycles mean the frontier is no longer stable across that horizon - the firm is optimizing over a moving target.

• AI Productivity Core

  GenAI productivity-measurement methodology assumes the capability being measured persists from data collection through publication. Capability shifts within that window, so reported productivity estimates anchor to a frontier that has already moved.

• Augmentation Core

  Centaur-style augmentation theory (human leads, AI assists) presupposes a stable complementarity between human and machine capability. AI improves on a six-to-twelve-month cycle; human capability does not. Asymmetric capability gain progressively destabilizes the new normal.

• Job Polarization Partial

  Long-run polarization dynamics assume capability shocks are bounded and that recovery periods can absorb them. On the trajectory observed so far, continuous capability gain compresses the recovery window the framework implies.

Prior automation technologies operated inside workflows that had already been codified. The labour cost was front-loaded - specification, integration, sign-off - and ongoing labour was marginal. AI deployed into tacit knowledge work runs the other way. The workflow is not codifiable in advance, so the system requires continuous labour at scale: RLHF feedback to align outputs, ground-truth labelling for training data, content moderation, output verification, monitoring, red-teaming, fine-tuning for domain adaptation. The ghost-work literature documents this for back-stage labour. The structural fact extends to front-stage roles, where verification and attestation are increasingly the human work that remains in AI-enabled bundles.

Labour-saving estimates that assume discrete deployment systematically understate the labour requirement of AI systems and misidentify where the labour now sits. The augmentation framework reads the remaining human work as collaboration; the structural fact is that the remaining work is increasingly correction, verification, and attestation - labour categories the augmentation framework does not theorize as binding constraints. The ghost-work literature names the back-stage version of this but does not theorize its spread to front-stage roles as a structural feature of every AI-enabled bundle.

• Ghost Work Core

  The ghost-work literature named the hidden labour of annotators and moderators as a back-stage support category sitting behind the AI product. The continuous-feedback architecture of generative AI - RLHF, in-context correction, output verification - is pulling more of that same work into the front-stage role bundle.

• Augmentation Core

  Centaur-style augmentation theory names the remaining human work as partnership - judgement, oversight, collaboration. Increasingly, it functions as the continuous-feedback loop the model needs to keep working: RLHF, output verification, correction. A different category of labour from the one the framework theorises.

• Task Model of Automation Partial

  The task model of automation implicitly assumes discrete deployment - the technology is specified, installed, and operated. AI deployment is iterative; the ongoing-labour requirement is structural to the production function, not transitional.

Prior automation was domain-specific by construction: CRM systems for sales, design tools for engineering, payroll systems for HR. The modularity-and-design-rules framework and team-topology theory presupposed that automation respected functional boundaries. Foundation models do not. The same model is deployed across functions - substituting for distinct tasks in each, without functional adaptation. The model is the cross-functional infrastructure layer the prior literature did not anticipate.

The team-topology principle inverts. Organizational structure no longer codifies functional specialization at the team level once the deployed model is general across functions. The strategic-bottleneck logic in the modularity-and-design-rules framework, which located the bottleneck at module interfaces in the product architecture, now locates it at the orchestration layer above all modules - a layer the prior literature did not theorize as a strategic asset. Job-polarization analyses that assumed function-specific routine partitions miss the cross-functional reach of foundation-model substitution.

• Team Topologies Core

  Team-topology theory assumes team boundaries codify functional specialization - sales tooling for sales, design tools for design, payroll for HR. As general-purpose models are deployed across more functions, they behave less like team-level artifacts and don't respect the boundaries the literature treats as structural.

• Design Rules Core

  The modularity-and-design-rules framework places the strategic bottleneck at the interface between modules in a product, with the platform architect as the implicit orchestrator. As foundation models substitute across more modules cross-functionally, the bottleneck is rising above any single product's modular architecture toward a general-purpose orchestration layer.

• Job Polarization Partial

  Job-polarization analyses partitioned routine and non-routine tasks within specific functions. Foundation models are general-purpose and cross-functional, so the partition does not scope cleanly onto the wage distribution it predicted.

• Industry Architecture Partial

  Industry architecture theory treats value chains as composed of separable functional layers - each can be vertically integrated or disintegrated by firm choice. Foundation models extend across those functional boundaries with one general-purpose capability, which the framework doesn't natively theorize. The architectural redraw now happens across all functions simultaneously rather than function-by-function.

Prior automation technologies were specified, procured, and operated by the deploying firm. The firm internalized the technology choice and the resulting surplus. AI capability is produced by upstream actors and rented by the deploying firm via API access. The provider sets the safety policy, the rate limit, the permitted use cases, the fine-tuning constraints, the deprecation schedule, and the data-handling terms. The relevant degrees of freedom for the deploying firm are narrower than the technology choice the literature models.

The firm-as-optimizer framing in the task model of automation presupposes that the firm chooses among feasible technology configurations. With AI, the configuration is determined upstream; the firm chooses adoption within a constrained menu. Three implications follow. (i) Surplus migrates outward to the provider, partially exiting the firm's production function - the firm-level optimization estimates do not capture this. (ii) The automate-versus-informate choice framed by labour process theory as a power move by capital is partially pre-resolved by the provider's design choices; the two-actor political economy under-counts the provider as an exogenous actor. (iii) Strategic bottlenecks in the modularity-and-design-rules framework shift from product modularity, where they were endogenous to the firm, to the orchestration layer, where they are set by provider API design - exogenous to the firm.

• Task Model of Automation Core

  The task model of automation, in its firm-level optimization form, presupposes that the firm chooses among feasible technology configurations. With AI, the configuration is set upstream - by the provider's training data, safety policy, rate limit, fine-tuning constraints, and deprecation schedule - and the firm chooses adoption within a constrained menu.

• Labour Process Core

  Labour process theory modelled the firm choosing between automating a job and informating it - capital vs labour, two actors. The model provider now sets parts of that choice upstream through training data, fine-tuning constraints, and API design - a third exogenous actor the framework did not theorize.

• Design Rules Core

  The modularity-and-design-rules framework places strategic bottlenecks inside a product's architecture, where the firm has design authority. Foundation-model API access shifts a meaningful share of the architectural choice to the provider - what the model exposes, what it refuses, what it deprecates - and the firm's modular choices are constrained by that upstream design.

• Industry Architecture Partial

  Industry architecture theory explicitly addresses vertical disintegration - firms outsourcing functions to specialized providers as market dynamics evolve. The framework's logic carries: AI capability sitting with an upstream provider is a vertical disintegration. What it doesn't natively theorize: the provider isn't a market participant the firm chose, and the configuration is set upstream rather than negotiated.

Prior automation required formal specification, IT involvement, and organizational decision-making. The literature assumed deployment authority sat at the firm or team level. Natural-language interfaces let individual contributors deploy AI capability without formal specification, IT involvement, or organizational approval. The decision moves to the individual level, ahead of the firm-level decision-making process.

Team-topology theory presupposed that team and tool boundaries were set by engineering choice at the firm level. They are now set by individual choice at the IC level. Bundle reconfiguration is occurring below the formal decision-making structure of the firm, often without firm-level visibility. Firm-level optimization in the task model of automation misses this because the firm is not the relevant unit of choice. The modularity-and-design-rules framework, which assumed deliberate architectural choice at the design level, has the same misspecification: the cuts are now occurring via individual prompting rather than via architectural decision.

• Team Topologies Core

  Team-topology theory assumes engineering authority codifies team-tool boundaries - IT decides which tools the team uses, the team adopts them. Natural-language interfaces let individual contributors deploy AI tools without engineering involvement, so deployment authority decentralizes below the team level.

• Design Rules Core

  The modularity-and-design-rules framework assumes architectural choices are deliberate decisions taken at the design level by an architect. Individual contributors prompting general-purpose models are making modular cuts inside the workflow without any explicit architectural decision being recorded.

• Task Model of Automation Partial

  Firm-level optimization in the task model of automation misses individual-level deployment that is occurring ahead of and outside any firm-level decision. Increasingly, the unit of choice is not the firm.

Prior automation produced deterministic outputs that could be certified once and trusted. AI outputs are probabilistic: identical prompts yield different outputs, different model versions yield different outputs. Verification of an AI output is not a one-time certification but a per-output check, conditional on the output. Critically, the verification carries liability - the human who attests to the output is liable for it. Verification is not a task AI can perform on its own outputs in a way that resolves the liability constraint; the liability requires an accountable human.

This aspect explains the empirical pattern of where the value-capturing human role ends up after rebundling. Across roles where AI is most capable - radiology, legal drafting, code generation, medical diagnosis, financial analysis - the remaining human role consistently centers on signature, attestation, sign-off, or risk-bearing. The pattern is not explained by what AI cannot do at the task level (it can produce the read, the filing, the diagnosis). It is explained by what verification structurally requires: an accountable human. The Reshuffle thesis claim that value migrates to whoever holds the new constraint reads through this lens. The new constraint is verification-with-liability; the holder of that remaining constraint is the human signatory. This is the load-bearing aspect for the survivor pattern in the empirical record.

• Routinization (ALM) Core

  The routinization hypothesis assumes output trust - once a task is automated, the output is reliable enough to feed the next step in the workflow. Probabilistic generation strains that assumption; verification of each output is becoming a structural labour category the framework does not name.

• Job Polarization Core

  Wage analyses in the job-polarization literature do not register verification as a distinct labour category, so they cannot account for it as an emerging high-wage role that survives rebundling around AI.

• Task Model of Automation Core

  The task model of automation does not theorize verification as a distinct task in the production function. With probabilistic outputs, verification is becoming the task that remains after substitution and the one that carries the liability.

• Ghost Work Core

  The ghost-work literature names verification labour as back-stage support - annotators checking outputs after the fact. Front-stage attestation and sign-off carry legal liability, so the same labour is migrating into the value-capturing role in the bundle, not the support tier.

• Augmentation Core

  Centaur-style augmentation theory treats verification as part of the partnership - the human checks, refines, collaborates. The framework does not see verification-with-liability as the constraint that defines which part of the rebundled role captures value.

Earlier AI systems produced text, code, or analysis; a human read the output and decided what action to take. Agentic AI takes the action directly: books the flight, sends the email, deploys the code, executes the trade, schedules the meeting, reads the email and replies. The decision-execution gap that the labour-economic literature implicitly assumed would remain human-mediated is collapsing into the agent itself.

Three implications. (i) The substitution surface expands from generation to execution. Roles whose value derived from being the decision-and-action layer are substituted at both ends. (ii) The locus of value capture accelerates toward whoever owns the agent's policy surface - typically the upstream provider, who defines what the agent can do, with what oversight, under what guardrails. Capture acceleration is faster than what generative AI alone produced. (iii) The augmentation framework breaks in a way that even tacit substitution did not. Augmentation presupposed the human is the decision-maker; under agentic action, the human is the policy-setter at best, the auditor at worst. The centaur new normal (human leads, AI assists) becomes machine-led-with-token-human-oversight at faster pace than the augmentation literature anticipated.

• Augmentation Core

  Centaur-style augmentation theory puts the human in the routine decision loop alongside AI - query, output, decide, act. As agentic AI increasingly books the flight, executes the trade, sends the email, the human is shifting toward policy-setter or auditor - and away from the in-loop decider role the frame presupposes.

• Team Topologies Core

  Team-topology theory treats humans as the action-takers across organizational interfaces - the people who execute handoffs between teams. AI agents are becoming a new cross-cutting actor at that layer, taking actions across team boundaries the framework theorises as human-only.

• Design Rules Core

  Modular interfaces in the modularity-and-design-rules framework presuppose human action between modules - a person passing artifacts from one module to the next. Agent execution is making the inter-module actor non-human in more places, which changes which interfaces are strategic and where coordination cost concentrates.

• Labour Process Core

  Labour process theory framed the firm's choice as automate vs informate - replace the worker or augment them. Agency adds a third option - delegate execution to the agent under a policy surface set by the provider - which the framework did not theorize.

• Ghost Work Partial

  The ghost-work literature documents the humans behind AI outputs - annotators, moderators, reviewers. The framework does not account for actions agents take that no human sees or verifies, which expands the back-stage labour category in directions the framework does not name.

Coase explains firms as solutions to transaction costs: internal hierarchy is cheaper than coordinating each task through arms-length contracts. AI lowers BOTH costs - internal coordination AND market coordination - but unevenly. The algorithmic platform becomes a new locus of coordination, sitting between firm and market. It dispatches work, ranks workers, sets prices, monitors performance, and absorbs the coordination tasks that historically defined the firm. Workers below the algorithm (whose work is dispatched, ranked, priced) populate this new structure. Workers above the algorithm (who design, configure, own) capture its surplus.

The labour-economic literature treats labour as flowing between two states - firm-employed or market-contracted. Neither captures the dominant new structure. As bundles unbundle, fragments are absorbed into the algorithmic platform, not returned to the firm. Reading displacement through firm-vs-market boundaries misses where work ends up. The above-vs-below-the-algorithm distinction is the right diagnostic - it tracks who captures the surplus. Regulation, taxation, and labour protection designed for firm-employment relationships have no purchase on this structure.

• Task Model of Automation Core

  The task model of automation treats substituted tasks as either staying inside the firm with different workers or being eliminated. It doesn't theorize platform-mediated reorganization, where unbundled fragments are absorbed by an algorithmic intermediary that sits between firm and market.

• Team Topologies Core

  Team-topology theory assumes work flows through teams inside firms. Algorithmic platforms organize work that no team owns - dispatching, ranking, pricing, and monitoring across worker pools without team boundaries.

• Design Rules Core

  The modularity-and-design-rules framework treats modularity as letting firms outsource modules across the boundary. It doesn't theorize the algorithm itself as a coordination infrastructure replacing both internal hierarchy and external market interfaces.

• Labour Process Core

  Labour process theory frames capital vs labour at the firm level. Algorithmic platforms are a third actor that organizes labour outside any single firm's authority - capturing the coordination surplus the framework attributes to the firm.

• Ghost Work Partial

  The ghost-work literature anticipated platform-mediated labour but treated it as a back-stage support category sitting behind the AI product. The structural fact is broader: the algorithmic platform is becoming the dominant new structure of work, not a support tier.

• Industry Architecture Core

  Industry architecture theory's central claim is that value migrates to whoever holds the new architectural bottleneck when the architecture shifts - and the framework names firms as the actors who control the shift through entry, exit, M&A, and vertical disintegration. AI-mediated algorithmic platforms are exactly such a new bottleneck, but they sit in a third coordination mode the framework doesn't natively theorize: not a firm in the traditional sense, not a market, but an algorithmic intermediary that organizes labour outside any single firm's authority.