Experiment
AI Policy Sandbox
Compare alternative policy designs using connected evidence, behavioural assumptions and explicit uncertainty before full-scale implementation.
01
Problem to address
Policy teams need to compare trade-offs across uptake, cost, operations, distribution and outcomes, but evidence lives in separate datasets and assumptions are often implicit.
02
What the evidence tells us
Strategy teams need to know where interventions work and whether outcomes were achieved. Services cannot always measure how they are performing. People also want to see likely consequences before acting.
03
The proposition to explore
The idea
Create a policy sandbox where teams can compare alternative interventions or rules using connected evidence, explicit assumptions and visible uncertainty.
04
Hypothesis
If policy teams can explore scenarios with assumptions and uncertainty visible, they will make more informed choices and know which evidence gaps must be tested before implementation.
05
How it could work
Combine historical evidence, simple behavioural assumptions and validated models. Let teams vary a policy parameter and compare scenario outputs, while the system makes uncertainty, excluded evidence and sensitivity explicit.
06
Example: bring it to life
Test a 20% peat-restoration payment increase under low, expected and high uptake. Compare indicative spend, workload, affected cohorts and environmental evidence confidence before deciding what to pilot.
07
AI opportunity
Evidence synthesis, scenario generation, model orchestration and uncertainty explanation. AI is the interface and reasoning assistant; validated analytical models provide the quantitative backbone.
08
Potential value
- Faster comparison of options
- Transparent trade-offs
- Earlier identification of evidence gaps
- Bridge from qualitative policy design to modelling
09
Reduce uncertainty without overcommitting
Smallest meaningful experiment
Build a deliberately small sandbox around one policy variable using existing data and explicit assumptions. Test whether it changes the questions policy teams ask, not whether it perfectly predicts outcomes.
10
Open questions and risks
What we still need to learn
- Models can create false precision
- Scenario assumptions need independent scrutiny
- Historical data may not represent future behaviour