Module 3 · Cross-Cutting Challenges
AI for Decision Makers
Unit 13: Synthesis: A Decision Protocol for AI
Unit 13 of 13
The course has made one argument from many angles. AI lowers the cost of prediction, and increasingly of generation and action. Whether that improves an institution's decisions depends on what is predicted, how predictions become decisions, what they are compared with, how people respond, and who is accountable. This unit turns that argument into a protocol you can run on any proposal.
The Six Questions
| Question | What a good answer looks like | Red flag |
|---|---|---|
| 1. Is it a prediction problem? | Clear statement of whether the decision needs a forecast or a causal effect, and evidence for the intervention if it is the latter | A risk score used to allocate a programme with no evidence about who benefits (Unit 3) |
| 2. What is predicted, and is the label valid? | Explicit target variable, with an account of how it is measured and for whom it may be mismeasured | Convenient proxies — arrests for crime, spending for need, scores for learning (Units 5–7) |
| 3. How do predictions become decisions? | A documented decision rule with stated error costs and a named owner | A vendor default threshold nobody chose (Units 2–3) |
| 4. Compared with what? | Measured performance of the current process and of a simple baseline, in your population | Comparison with an idealized human or with 'nothing' (Units 3–4) |
| 5. How will the system respond? | A plan to measure adherence, overrides, gaming, and feedback effects | An assumption that people will use the tool exactly as designed (Units 4–5, 11) |
| 6. Who is accountable, and how would we know it's failing? | Named owners, monitoring of outcomes by group, a working appeal route, and a stop rule | 'Human in the loop' as the only safeguard (Unit 11) |
A Staged Protocol
- Screen. Define the decision, its volume, and its stakes. Identify which decisions would change, and whether the problem is predictive or causal. Kill weak proposals here.
- Specify. Fix the target label, the decision rule and its error costs, the owner of each, and the comparison set: current process, simple baseline, best non-AI alternative.
- Validate. Test locally and, where possible, prospectively. Measure calibration and error rates by group at the intended threshold. Check vendor claims with the questions in Unit 3.
- Pilot. Roll out in a way that allows causal evaluation (randomized or staggered). Measure decisions and outcomes, not only model metrics, and measure how people actually use the system.
- Decide. Weigh costs and benefits against the status quo and the alternatives (Unit 4), including distributional effects and legitimacy. Record the decision and its rationale.
- Operate. Monitor performance, drift, overrides, and outcomes by group. Maintain an appeal route and incident process. Set a review date and the conditions under which you would stop.
Key Insight
Most failures in this course would have been caught at stages 2–4: an invalid label (Obermeyer), a group distribution imposed on individuals without recourse (A-levels), no independent validation (Watson, Epic), or no measurement of real-world use (Kentucky). The protocol is cheap compared with any of those failures.
Red Flags, in Priority Order
- No measured baseline for the current process.
- Performance evidence only from the vendor, or only from another population.
- Accuracy quoted without base rate, threshold, or positive predictive value.
- Target variable chosen for convenience, with no discussion of how it is mismeasured.
- An adverse decision about individuals with no practical route to contest it.
- No plan to measure overrides and adherence after launch.
- A contract without audit access, update control, or exit rights.
The One-Page Decision Memo
Require every AI proposal that affects people's rights, access, or safety to fit on one page under these headings. If it can't, it isn't ready.
- Decision supported, annual volume, and who is affected.
- Prediction or causal problem — and the evidence for the intervention, if causal.
- Target label, how it is measured, and known measurement gaps.
- Decision rule, error-cost assumptions, and the owner of each.
- Comparison: current process, simple baseline, best non-AI alternative — with numbers.
- Validation evidence: where, on whom, by whom, prospective or retrospective.
- Expected effects by group, and the fairness criterion chosen, with reasons.
- Oversight design: override rules, logging, and adherence monitoring.
- Recourse: how affected people contest decisions, and who can change them.
- Monitoring metrics, review date, and stop conditions.
Three Lenses, Revisited
| Lens | Core question | What it reveals |
|---|---|---|
| Structural (the decision system) | Where does AI enter, and where has judgment moved? | Labels, thresholds, decision rights, and points of failure |
| Economic (incentives and alternatives) | Is this worth it compared with the alternatives, and for whom? | Hidden costs, the status quo's costs, marginal value, and strategic responses |
| Epistemic (evidence and trust) | What evidence supports trusting this system for this task, here? | Validation gaps, the jagged frontier, and what monitoring must catch |
No single lens is enough. A well-validated model can be deployed through a broken decision rule. A cost-effective system can be illegitimate. A well-governed process can rest on an invalid label. Used together, the lenses make most failure modes visible before launch.
Your Role
- Own the judgments: targets, thresholds, and fairness criteria are leadership decisions, not technical defaults.
- Insist on comparison with the real status quo — in both directions. Don't let AI be held to a standard the current process fails, or excused from one it meets.
- Fund the complements: process redesign, evaluation capacity, and monitoring are where value and safety come from.
- Measure use, not just models: adherence, overrides, and outcomes by group.
- Keep contestability real for the people affected.
- Revisit decisions as capabilities and evidence change; the frontier moves.
Key Insight
AI is now part of how institutions decide. The question isn't whether to trust it in general. It is how to build decision systems that use it where the evidence supports it, constrain it where it doesn't, and make the remaining judgments explicit and accountable. That is a leadership task, and you now have a protocol for it.
Core Readings
- Agrawal, A., Gans, J., & Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press.
- Kleinberg, J., Ludwig, J., Mullainathan, S., & Obermeyer, Z. (2015). Prediction Policy Problems. American Economic Review, 105(5), 491–495.
- Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown Spark.
- Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1, 206–215.
- Stevenson, M. T., & Doleac, J. L. (2024). Algorithmic Risk Assessment in the Hands of Humans. American Economic Journal: Economic Policy, 16(4), 382–414.
- Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453.