Unit 4: The Economics of AI Decisions

Unit 4 of 13

Whether an institution should deploy an AI system is an economic question in the broad sense: a choice under scarcity, with trade-offs, alternatives, and other people's responses. Betsey Stevenson and Justin Wolfers organize introductory economics around four core principles: cost-benefit, opportunity cost, the marginal principle, and interdependence (Stevenson & Wolfers 2020). They are general tools of economic reasoning, not an AI-specific framework. Applied carefully, they cut through most AI proposals.

Principle 1: Cost-Benefit

Do something if, and only if, its benefits exceed its costs — counting every cost and benefit, including the ones that don't appear in a budget. AI proposals tend to quantify benefits ("claims processed 10× faster") and leave costs vague. A full ledger looks like this:

Category	Costs to count	Benefits to count
Financial	Licences, compute, integration, maintenance, monitoring, exit costs	Labour time saved, throughput, reduced losses
Errors	False positives and negatives at your base rate (Unit 3), and their downstream consequences	Errors of the status quo that the system avoids
Organizational	Redesign, training, review workload, deskilling	Consistency, faster feedback, freed expert time
Legitimacy & legal	Contestability, perceived fairness, liability, regulatory duties	Documented, auditable decision rules
Distribution	Who bears the errors and costs	Who receives the gains

Key Insight

Count the status quo's costs too. A proposal is not 'costly AI vs. costless humans'. Human decision processes have error rates, delays, inconsistencies, and biases that are usually unmeasured. An honest analysis prices both.

Principle 2: Opportunity Cost

The true cost of something is what you give up to get it: the best alternative use of the same money, time, and attention. For AI this means two comparisons. The first is AI versus the best non-AI use of the same resources. The second is AI versus the realistic status quo, including the cost of not acting. Both directions matter. A risk tool that absorbs a court's reform budget may cost a cheaper intervention with stronger evidence (Unit 5). Conversely, refusing an effective screening tool where specialists are scarce has its own cost: undiagnosed patients (Unit 7).

Intuition Check

For any proposal, write three columns: the AI option, the best non-AI option with the same budget, and doing nothing new. If the proposal only compares the first with the third, it has not done the analysis.

Principle 3: The Marginal Principle

Decisions about quantities are best made incrementally: ask whether one more unit is worth it, rather than whether the whole is good or bad. AI deployment is full of quantity decisions:

How much accuracy to buy: raising sensitivity from 85% to 90% may double false alarms at a low base rate. Is the extra catch worth the extra review cost?
How much to automate: which share of cases should be decided automatically, and which routed to a person? The answer lies where the marginal cost of review equals the marginal value of the errors it catches.
How many decisions change: a system's value comes from the marginal cases where it alters the decision (Unit 3), not from its performance on cases everyone already gets right.

Note what the marginal principle is not. Asking 'accurate for whom?' — whether a system that works on average fails particular groups — is essential, but it is a question about distribution and fairness (Unit 10). Keeping the two ideas separate prevents muddled analysis. One asks whether the next increment is worth its cost. The other asks who bears the costs and benefits.

Principle 4: Interdependence

Your best choice depends on your other choices, on others' choices, on developments in other markets, and on expectations about the future. AI systems are embedded in exactly this kind of web, and four patterns recur:

Pattern	What happens	Example
Strategic response (gaming)	People learn what the system rewards and change their behaviour or reported data, eroding the signal	Applicants tune CVs to screening software; sellers game ranking algorithms (Hardt et al. 2016)
Performativity	The prediction changes the outcome it predicts	A student labelled 'at risk' is treated differently, which changes the risk
Feedback loops	Decisions shape the data used for the next round of predictions	Patrols sent where past arrests were recorded generate more recorded arrests there (Lum & Isaac 2016)
Correlated decisions (monoculture)	Many organizations using the same model make the same errors on the same people	A shared hiring screen can shut the same applicant out of every employer at once (Kleinberg & Raghavan 2021)

Beyond the Four Principles: Incentives and Complements

Two further pieces of economics explain most of the gap between AI's promise and institutional results.

Principal–agent problems. The vendor is paid for adoption and headline metrics, not for your outcomes. Staff are evaluated on throughput, not on the errors a system causes. A 'human in the loop' may be there to absorb blame rather than to exercise judgment (Unit 11). Ask whose incentives each design choice serves.
Complementary investments. General-purpose technologies deliver gains only after organizations invest in redesigned processes, skills, and data. Those investments are costly and largely unmeasured, so measured productivity can fall before it rises — the 'productivity J-curve' (Brynjolfsson, Rock & Syverson 2021). Budget for the complements, and don't judge a deployment only on its first months.

The Principles as an Evaluation Checklist

Principle	Key question	Watch for
Cost-benefit	Have we priced all costs and benefits — including the status quo's errors?	Benefits quantified, costs hand-waved; status quo treated as costless
Opportunity cost	What is the best alternative use of these resources, and what does inaction cost?	False binary of 'AI vs. nothing'
Marginal principle	Is the next increment of accuracy or automation worth its cost? Which decisions change?	Averages over cases the system doesn't affect
Interdependence	How will people, competitors, and other systems respond?	Evaluating the system as if nobody will adapt to it
Incentives & complements	Whose incentives does this serve? What complementary investment is needed?	Vendor metrics standing in for outcomes; no budget for redesign

Key Takeaways

The four core principles (Stevenson & Wolfers) are general economic reasoning, and they apply directly to AI proposals.
Cost-benefit and opportunity cost must include the realistic status quo and the cost of not acting, not just AI's risks.
The marginal principle is about increments — how much accuracy, how much automation, which decisions change. 'Accurate for whom?' is a separate, distributional question.
Interdependence shows up as gaming, performativity, feedback loops, and correlated errors across organizations.
Misaligned incentives and missing complementary investments explain much of the gap between AI's promise and institutional results.

Evidence & Further Reading

Stevenson, B., & Wolfers, J. (2020). Principles of Economics, ch. 1: The Core Principles of Economics. Worth Publishers / Macmillan Learning.
Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The Productivity J-Curve: How Intangibles Complement General Purpose Technologies. American Economic Journal: Macroeconomics, 13(1), 333–372.
Hardt, M., Megiddo, N., Papadimitriou, C., & Wootters, M. (2016). Strategic Classification. Proceedings of ITCS 2016.
Lum, K., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.
Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. Proceedings of the National Academy of Sciences, 118(22), e2018340118.