AI for Decision Makers
Unit 9: Economic Systems
AI's economic effects run through four channels: which tasks people do, how productive they are, how markets are structured, and how gains are shared. Public debate tends to jump to mass unemployment. The evidence so far points to something more specific: large, uneven, task-level effects whose distribution depends heavily on choices by firms and governments.
The Institutional Landscape
| Level | AI effect | Key question for decision-makers |
|---|---|---|
| Tasks | Changes which tasks within jobs are done by people, by software, or by both | Are we automating tasks or augmenting the people who do them? |
| Firms | Algorithmic management, measurement, and coordination of work | Who gains from productivity improvements, and who bears new risks? |
| Markets | Large fixed costs in frontier models; platforms that set market rules | How concentrated are the inputs we depend on, and how locked in are we? |
| Policy | Shifts in demand for skills; new forms of work | Which policies have evidence behind them, and which are slogans? |
Exposure Is Not Displacement
'Will AI take this job?' is the wrong unit of analysis. Jobs are bundles of tasks, and technology changes tasks. Using that lens, Eloundou and colleagues estimate that about 80% of US workers could have at least 10% of their tasks affected by large language models, and about 19% could see at least half their tasks affected (2024). Unlike earlier automation waves, exposure is higher for higher-wage occupations. Exposure measures potential change, though, not job loss. Whether exposed tasks are automated, augmented, or untouched depends on cost, reliability, regulation, and management choices.
| Task type | Examples | Earlier automation (computers, robots) | Generative AI |
|---|---|---|---|
| Routine cognitive | Data entry, standard reports, form processing | High automation | Further automation |
| Non-routine cognitive | Writing, analysis, coding, advising | Complemented — largely protected | Among the most exposed tasks |
| Routine manual | Assembly, sorting, packing | High automation via robotics | Limited direct effect |
| Non-routine manual | Care work, trades, repair | Low automation | Limited direct effect, though work around these tasks (scheduling, records) is exposed |
Task framework adapted from Autor, Levy & Murnane (2003), updated with Eloundou et al. (2024)
What the Productivity Experiments Show
| Study | Setting | Finding |
|---|---|---|
| Noy & Zhang (2023) | Professional writing tasks, randomized access to ChatGPT | About 40% less time and 18% higher quality; gaps between workers narrowed |
| Brynjolfsson, Li & Raymond (2025) | 5,172 customer-support agents, staggered rollout of an AI assistant | About 15% more issues resolved per hour on average; largest gains for less experienced workers; small gains in speed, with small quality declines, for the most skilled |
| Dell'Acqua et al. (2023) | 758 consultants on realistic tasks | Inside the AI frontier: 12.2% more tasks, 25.1% faster, over 40% higher quality. On a task outside it: 19 percentage points less likely to be correct |
| METR (2025) | 16 experienced open-source developers, 246 tasks in their own codebases | Tasks took 19% longer with AI tools, although developers believed AI had made them about 20% faster |
Displacement, Reinstatement, and Choice
Acemoglu and Restrepo decompose technology's effect on labour into three forces (2019). Displacement: machines take over tasks. Productivity: cheaper production raises demand. Reinstatement: new tasks are created for people. Past automation waves were followed by new tasks that absorbed workers, which is why automation anxiety has often proved overblown. They also warn about 'so-so automation': technology that displaces workers without raising productivity much. It delivers the costs of automation without its benefits.
Whether AI differs from earlier technologies is an open question. Its reach across cognitive tasks could speed up displacement relative to reinstatement. For decision-makers, the actionable point is that automation versus augmentation is largely a design choice. The same model can replace a junior analyst or make a junior analyst far more productive. That choice determines who captures the gains (Acemoglu & Johnson 2023).
Market Structure: Compute, Capital, and Dependence
A common claim is that 'data network effects' make AI markets winner-take-all. The evidence is more nuanced: for many applications, additional data shows diminishing returns (Bajari et al. 2019). The stronger concentration forces in today's AI are the very large fixed costs of training frontier models — compute, capital, and specialized talent — together with control over cloud infrastructure and distribution. Open-weight models and competition among providers push the other way. For an institution, the practical issue is dependence. Ask how many of your critical workflows rely on a single provider's model, what happens to them when that model changes or is withdrawn, and whether your data and evaluations would let you switch.
Algorithmic Management and Platform Work
Ride-hailing and delivery platforms use algorithms to assign tasks, set pay, and evaluate and deactivate workers. Through the market-design lens of Unit 1, the platform both runs the market and writes its rules. It sees aggregate data on demand, pricing, and every worker's behaviour. Each worker sees only their own jobs and earnings. That information asymmetry is a large part of the platform's bargaining power.
- Pay set by opaque, frequently changing algorithms makes income volatile and hard to evaluate.
- Ratings and acceptance metrics can function as discipline without the procedural protections of employment.
- Algorithmic task assignment can exert substantial control while workers are classified as independent.
Regulation is catching up. The EU Platform Work Directive (Directive (EU) 2024/2831) creates a rebuttable presumption of employment where a platform directs and controls work. It also sets rules on algorithmic management, including transparency about automated monitoring and decision systems and human oversight of significant decisions. Member states must transpose it by 2 December 2026.
The Four Principles Applied
| Principle (Unit 4) | Application | Insight |
|---|---|---|
| Cost-benefit | Algorithmic management raises measured efficiency while shifting income risk to workers | Count costs that land off the firm's balance sheet |
| Opportunity cost | Automating a role vs. augmenting the people in it with the same tools | The same budget can buy very different distributions of gains |
| Marginal principle | How far to automate a process: the next task automated may save little and cost much in errors or lost expertise | Automate incrementally, measuring each step |
| Interdependence | When every firm adopts the same tools, wages, skills, and market power shift economy-wide | Individually rational adoption can produce collective outcomes no firm chose |
Policy Options and the Evidence Behind Them
| Option | What the evidence suggests | Caveat |
|---|---|---|
| Retraining | Broad programmes have a mixed record; sectoral programmes tied to employers' needs have shown large earnings gains (Katz et al. 2022) | Design and employer links matter more than spending levels |
| Taxing automation ('robot taxes') | Theory supports at most modest, temporary taxes; a stronger case exists for removing tax biases that favour capital over labour (Acemoglu, Manera & Restrepo 2020) | Hard to define 'automation' for tax purposes; risks discouraging productive investment |
| Labour law for algorithmic management | Transparency and human review of significant automated decisions address documented harms | Enforcement and worker access to data determine whether rules bite |
| Competition policy | Scrutiny of concentration in compute, cloud, and model partnerships | Fast-moving markets make durable remedies hard to design |
| Safety nets | Income support and portable benefits cushion task-level disruption and volatile platform income | Fiscal cost; design for transitions, not only unemployment |
Key Takeaways
- Analyse tasks, not jobs. Generative AI exposure is highest for non-routine cognitive work and for higher-wage occupations — a reversal of earlier automation patterns.
- Experiments show large but task-specific gains, often largest for less experienced workers, and occasional losses; measure productivity rather than assume it.
- Whether AI automates or augments is largely a design and management choice that determines who captures the gains.
- Concentration in AI comes mainly from compute, capital, and distribution rather than data alone; manage your dependence on single providers.
- Algorithmic management concentrates information and power; new rules such as the EU Platform Work Directive target it directly.
- Prefer policies with evidence behind them; several popular proposals rest more on intuition than on results.
Evidence & Further Reading
- Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science.
- Autor, D. H., Levy, F., & Murnane, R. J. (2003). The Skill Content of Recent Technological Change: An Empirical Exploration. Quarterly Journal of Economics, 118(4), 1279–1333.
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
- Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942.
- Dell'Acqua, F., McFowland, E., Mollick, E., et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School Working Paper 24-013.
- METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. Research report, July 2025.
- Acemoglu, D., & Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.
- Acemoglu, D., & Johnson, S. (2023). Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. PublicAffairs.
- Bajari, P., Chernozhukov, V., Hortaçsu, A., & Suzuki, J. (2019). The Impact of Big Data on Firm Performance: An Empirical Investigation. AEA Papers and Proceedings, 109, 33–37.
- Katz, L. F., Roth, J., Hendra, R., & Schaberg, K. (2022). Why Do Sectoral Employment Programs Work? Lessons from WorkAdvance. Journal of Labor Economics, 40(S1), S249–S291.
- Acemoglu, D., Manera, A., & Restrepo, P. (2020). Does the US Tax Code Favor Automation? Brookings Papers on Economic Activity, Spring 2020.
- Directive (EU) 2024/2831 of the European Parliament and of the Council on improving working conditions in platform work (2024).