Build a testable pathway for an AI benefits adviser
Create a contribution pathway with mechanisms, assumptions, equity indicators, alternatives and stop or adapt decisions.
35 minutes
Scenario
A public-interest organisation plans an AI benefits adviser. The draft pathway says deployment will improve household security and measures only completed chats.
- Your role
- Programme and evaluation lead accountable for access, evidence and adaptation
- Method
- Theory of Change
Evidence pack
Need
Eligible households miss benefits because rules are difficult to understand and applications are incomplete.
Prototype
Users with supported languages complete guidance more often; unsupported-language users abandon at twice the rate.
Accuracy
Policy rules change monthly; the content pipeline has a 10-day update lag.
Agency capacity
Application decisions take 3–8 weeks and vary by region; the adviser cannot control processing.
Safeguard
Users must retain phone and in-person routes, human escalation and an appeal path.
External factor
A national outreach campaign will start during the pilot and may change application volume.
Constraints
- Frame long-term impact as contribution, not promise.
- Distinguish chats from outcomes.
- Test language access and policy currency as assumptions.
- Analyse the national campaign as an alternative contributor.
Case steps
Work through each prompt using the evidence pack. These guided cases support self-directed practice; server-scored attempts are not available yet.
Build backwards from secure access to benefits through observable outcomes, outputs and activities.
Identify critical assumptions, external factors, feedback and rival explanations.
Define indicators and continue, adapt or stop decisions before the pilot.