Learning objective
Distinguish activities, outputs and outcomes, then state mechanisms, assumptions, rival explanations and decisions for critical causal links.
Why it matters
An intervention can deliver every activity and still fail because people cannot, do not or should not follow the assumed pathway.
Core concept
Build backwards from meaningful change. Treat each arrow as a claim that can fail for identifiable reasons.
Scenario
An AI benefits adviser counts completed chats as success, but users with limited language access abandon applications and have no human route.
Worked decision
Chats are outputs. Outcomes include correct understanding, completed eligible applications, timely decisions and equitable access. Language coverage, trust, rule currency and agency capacity are assumptions. Preserve human escalation and measure drop-off and appeals by group.
Decision rule
Prioritise evidence for links that are uncertain, consequential and necessary. Define in advance what result would adapt or stop the intervention.
Ethical boundary
Affected groups must be able to challenge the desired outcomes, assumptions and acceptable trade-offs. Participation is not a checkbox after the pathway is fixed.
AI lens
AI can trace claims and find missing links; it cannot settle contested values or infer attribution.
Quick check
Deployment completed on time. What has been established?
A. The impact occurred.
B. An activity or output occurred; outcome evidence is still needed.
C. The programme caused any later change.
D. All assumptions held.
Answer: B.
Next step
Complete the AI benefits-adviser case and identify the one assumption whose failure should stop scale-up.