Principles
Structure before dashboards
A governed data layer and agreed definitions come before any tool choice. Dashboards are only as good as the structure beneath them.
Systems the team can run
Built and documented so the in-country team can maintain and extend them, and so they serve field staff rather than burden them.
Equity by design
Women’s access, phone access, literacy and landholding are analysed from the start, and data is never collapsed without a reason.
Honest limits
Contribution, not attribution. Limits are set before the analysis, and early signals are reported as early signals.
Change control and reproducibility
Snapshots before changes, written change records and reproducible sampling draws, so the work can be checked and taken over.
Definitions decided by the client
Where a definition is ambiguous, the options go to the client for a decision rather than being fixed silently.
Where AI helps
AI entered the work in 2025. It supports specific tasks, and people review its output before anything is used.
- Field-visit advisory: AI reviews agronomists’ visit records and suggests actions; the agronomy team decides.
- Bilingual interview transcription, translation and qualitative coding support, checked against the source.
- Scripting: a person designs the logic and checks the outputs; AI writes the code.
- An independent check on analysis before results go out.
Safeguards in place today
- Read-only data access by default
- A snapshot before every change, verification after
- Explicit approval before anything touches live data
- Every delivered figure verified by a person
Drafted · not yet in force
Governance direction
The practice has drafted an AI data policy and client disclosure. Its principles: tiered rules on what AI may touch, AI working on pseudonymized data, no identified individual-level records sent to AI tools, and clients told how AI is used and free to restrict it.