Methodology — Automated Intelligence Reporting (LLM)
The methodology treats LLM output as a first draft that must remain grounded in verified indicator data.
1. Data validation
Start with structured indicators, targets, actuals, risk notes, and recommendations.
2. Narrative planning
Map indicator status to report sections such as progress, risks, mitigation, and next steps.
3. Controlled generation
Generate concise summaries from only the available demo evidence.
4. Human review
Require M&E or programme review before donor submission.
Evaluation approach
- Compare generated narrative against source indicators for factual alignment.
- Track editing time saved by reporting officers.
- Monitor recurring risks and recommendation follow-through.
- Require reviewer sign-off before external use.
Responsible AI note: The sandbox demonstrates assistive decision support. In production, human review, audit logging, access control, model monitoring, and documented exception handling would be required.