AI-assisted data audits in results-based programmes
 |
note May 2026 ; 11 pages
Ed. Uptime - Oxford
Téléchargeable sous format: PdF
(1 240 ko)
Téléchargeable chez l'éditeur
Abstract:
Results-based programmes depend on rigorous verification, yet auditing evidence at scale remains costly and complex. This represents a persistent barrier to national implementation. AI-assisted data auditing offers a practical way forward. This paper examines the cost and practicality trade-offs of applying AI within large-scale results-based funding programmes to analyse diverse, unstructured datasets spanning image files, spreadsheets, PDFs, and handwritten records in multiple languages.
Since 2023, Uptime Global has applied this approach across two functions in audits of rural drinking water services within a multi-country results-based contract portfolio: data extraction and discrepancy confirmation. Using a combination of specialised document intelligence services and general-purpose Vision Language Models (VLMs), Uptime extracts over 200,000 performance records from thousands of pages of source documentation per audit cycle, at a total AI processing cost of under USD 350. Retrieval-Augmented Generation (RAG) offers further efficiency gains, enabling auditors to query source documents directly and receive evidence-grounded responses with traceable citations.
Cost is not a barrier to scale. Across a twentyfold increase in programme size, AI processing costs at current rates rise by less than a few hundred dollars per year and remain well under USD 1,000 dollars in total. Expressed as a share of total recurring programme costs, AI processing represents a fraction of one percent, even for programmes covering hundreds of thousands of sites.
Realising this potential requires clear expectations about what AI can and cannot do. It is one component of a broader audit pipeline, and many efficiency gains also reflect improvements in sample design, data systems, and script-based processing. AI cannot be a substitute for human judgement in interpreting ambiguous findings, nor for the governance structures that uphold sanctions arising from verification results. AI also depends on consistent, credible source data, an area that continues to require sustained investment in many low- and middle-income country contexts. For governments and development banks designing results-based programmes, AI-enhanced auditing is best understood not as an added cost, but as a tool that protects the integrity and credibility of their investments.
Editeur/Diffuseur: |
|
Uptime - Oxford - Royaume Uni |
En cas de lien brisé, nous le mentionner à communication@pseau.org