Field Overview

Core Puzzle

The field is organised around a persistent gap:

• task-level and micro studies often show large gains from AI tools
• firm, sector, and aggregate evidence still shows modest, noisy, and contested effects

That gap is not a single puzzle. It can arise because capabilities do not yet cover the binding tasks, because firms adopt shallowly, because complementary intangible investment is missing or unmeasured, because rents accrue upstream, because labour and demand effects offset productivity gains, or because some welfare gains sit outside GDP. The core notes separate those mechanisms rather than treating “AI productivity” as one object.

Reading Sequence

Note	Core question	Use it for
AI Capabilities	What can AI do now and later?	Task capability, exposure, jagged capability, agentic trajectories, research automation, physical bottlenecks
Adoption Dynamics	How does capability become organisational change?	J-curve, weak links, adoption depth, management quality, task-to-workflow orchestration, cross-country adoption gaps
Market Dynamics	Who captures value?	Chips, cloud, models, applications, data, concentration, pricing power, open source, capex and financial risk
AI and Labour Markets	What happens to workers?	Task frameworks, levelling up, entry-level erosion, automation vs augmentation, wages, distribution, job quality
Capital Deepening	How much is capital accumulation?	Tangible capex, cloud/API accounting, intangible capital, depreciation, capital services, capital-vs-TFP measurement
TFP and Innovation	How much is residual efficiency and innovation?	Direct TFP channels, AI in science, method of inventing, recursive AI-for-AI, TFP measurement limits
Supply-Side Synthesis	What is the aggregate supply-side effect?	Growth accounting, output ladder, task-to-aggregate attenuation, projections, consumer welfare gap
Macroeconomic Implications and Policy	What happens after supply meets demand, finance, and policy?	Inflation, r*, asset markets, fiscal capacity, tax base, stabilisation, state capacity
International Economics of AI	Where does value accrue across countries?	AI value chains, trade, services, sovereignty, export controls, development strategy, UK position

Field Map

1. Capability Is Necessary, Not Sufficient

AI Capabilities is upstream of the rest of the field. It asks what AI systems can do at the task, workflow, research, recursive, and physical layers. The important distinction is between capability in isolation and reliable substitution inside production. The micro-evidence is strong for bounded cognitive tasks, but the jagged frontier, tacit knowledge, verification, liability, and physical execution remain material constraints.

2. Adoption Is the Translation Layer

Adoption Dynamics explains why strong task-level results do not automatically become measured productivity. The main candidate mechanisms are the productivity J-curve, weak-links/O-ring bottlenecks, shallow adoption, management quality, data and compliance constraints, and the task-to-workflow orchestration gap. Adoption is also where cross-country and cross-firm divergence first becomes visible.

3. Market Structure Determines Value Capture

Market Dynamics asks who captures the surplus AI creates. The stack can be concentrated upstream in chips, compute, cloud, frontier models, data, and distribution even while applications proliferate downstream. This matters because welfare, wages, fiscal receipts, and investment incentives depend on whether value flows to adopters and workers or is extracted by upstream suppliers and platforms.

4. Labour Effects Are Compositional Before They Are Aggregate

Labour Markets treats jobs as bundles of tasks. The near-term evidence is not a broad aggregate employment collapse; it is compositional change. The important signals are levelling-up effects in micro studies, entry-level erosion in exposed occupations, contracting-platform weakness, possible expertise erosion, and job-quality changes through monitoring, work intensity, and residual-task concentration.

5. Growth Accounting Splits The Supply Side

The supply-side notes separate three channels that are often conflated:

• labour: quantity, quality, allocation, and task composition
• capital deepening: more tangible and intangible AI capital services per worker
• TFP and innovation: residual efficiency gains and faster idea production

Supply-Side Synthesis recombines them. The same headline labour-productivity number has different implications depending on the channel mix: capital-deepening-heavy growth is more exposed to depreciation, obsolescence, upstream rents, and investment cycles; TFP-intensive growth is potentially more durable but harder to measure; labour-displacing growth has sharper distributional and demand consequences.

6. Macro Policy Is About Pass-Through

Macroeconomic Implications and Policy asks what happens after the supply-side shock interacts with demand, inflation, asset markets, and the state. A positive productivity shock can still produce weak demand if gains accrue to low-MPC capital owners while losses fall on high-MPC workers. AI can also produce sectoral disinflation alongside bottleneck inflation in power, land, construction, and compute. Public finance becomes central if the tax base shifts away from labour income toward capital, profits, and intangible rents.

7. The Stack Is International

International Economics of AI treats AI as an international value chain. Countries can access AI without owning the stack, but they may pay recurring rents to foreign chip, cloud, model, and platform providers. The same system is also a national-security system: compute, cloud, data, models, and AI assurance can sit inside defence, cyber, public administration, and critical infrastructure. Sovereignty therefore means managed interdependence, not full-stack autarky.

Foundational Lenses

• Task-based technological change. The unit is the task, not the occupation. Autor, Levy, and Murnane (2003) and Acemoglu and Restrepo (2019) are the backbone.
• General-purpose technology and the J-curve. AI may require complementary invention and intangible investment before aggregate gains appear. Bresnahan and Trajtenberg (1995) and Brynjolfsson, Rock, and Syverson (2021) frame this.
• Prediction economics. Cheap prediction raises the value of complements: judgement, data, workflow design, and decision rights (Source Note - Prediction Economics).
• O-ring and weak links. Automating many tasks may do little if the remaining bottleneck tasks bind the production chain (Kremer (1993), Jones (2026)).
• Growth accounting. Labour productivity, capital deepening, and TFP are different objects. The distinction matters even when the channels are jointly determined.
• Market structure and rents. AI’s welfare effects depend on competition across chips, cloud, models, data, applications, and distribution, not only on total productivity.
• Managed interdependence. International AI policy is about access, ownership, adoption, and strategic dependence across a transnational stack.

Evidence Status

Settled Enough To Use As Working Facts

• AI has large task-level effects in bounded cognitive domains.
• Exposure is not realised substitution.
• Adoption is broad but uneven, and depth matters more than tool access.
• The lab-to-deployment-to-aggregate attenuation is steep.
• Compute, cloud, chips, data, and distribution are economically important concentration points.
• Growth accounting must distinguish labour, capital deepening, and TFP.

Contested But Decision-Relevant

• Whether weak macro productivity data mainly reflects delay, measurement, weak-links constraints, or genuinely modest effects.
• Whether AI primarily augments labour, substitutes for labour, or does both across different margins and horizons.
• Whether entry-level erosion is an early broad labour-market signal or still concentrated in a few exposed occupations.
• Whether application-layer adopters capture most surplus as models commoditise, or upstream providers keep pricing power.
• Whether current AI capex is productive infrastructure buildout, overbuild, or both.
• Whether frontier-model ownership or broad diffusion capacity matters more for national advantage.

Still Speculative

• Whether AI automating research changes the long-run growth regime.
• Whether agentic systems close the workflow-orchestration gap quickly enough to change adoption dynamics within a few years.
• Whether AI materially weakens classic development paths for lower-income countries.
• Whether current financial-market enthusiasm creates systemic macro risk or mainly funds useful infrastructure.

Why The Field Disagrees

• Scholars use different evidence windows: lab tasks, firm deployments, vacancy data, payroll data, capex disclosures, macro aggregates, and structural models.
• They use different units of analysis: tasks, jobs, workflows, firms, sectors, countries, and growth regimes.
• They disagree about horizons: near-term augmentation, medium-term reallocation, and long-run growth-regime change can point in different directions.
• They assign different weights to complements, bottlenecks, institutions, and market power.
• They compare projections, structural decompositions, expert priors, and realised measurements too casually.

Entry Points By Question

• What can AI actually do? Start with AI Capabilities.
• Why are aggregate productivity effects still muted? Start with Adoption Dynamics and Supply-Side Synthesis.
• Who captures the gains? Start with Market Dynamics.
• What happens to workers? Start with Labour Markets.
• Is AI growth mostly capex, TFP, or labour reallocation? Start with Capital Deepening, TFP and Innovation, and Supply-Side Synthesis.
• What are the policy implications? Start with Macroeconomic Implications and Policy.
• What does this imply for trade, sovereignty, and the UK? Start with International Economics of AI.