AI-Assisted Data Audits in Results-Based Programmes
Briefing Note | May 2026
Andrew Armstrong, Uptime Global
Results-based funding depends on credible, independently verified performance data. However, reviewing the thousands of supporting documents required to verify results can be costly, time-consuming and difficult to scale.
Published in May 2026, AI-assisted data audits in results-based programmes examines how artificial intelligence (AI) can reduce the cost and complexity of verification while maintaining the integrity of results-based payments.
Drawing on Uptime Global's experience managing results-based contracts for rural drinking water services, the paper demonstrates how AI can process large volumes of operational and financial evidence, identify discrepancies and support more efficient audits.
The findings offer practical lessons for governments, development banks and other organisations seeking to implement results-based programmes at national scale.
Applying AI to Rural Water Data Audits
Uptime's results-based contracts link payments to measurable performance, including water service reliability, water volumes, local revenue and water safety.
To verify these results, service providers submit supporting evidence ranging from handwritten maintenance logs and financial receipts to automated meter readings and digital monitoring records.
Across Uptime's international portfolio, this documentation spans multiple languages, formats and levels of completeness, creating significant challenges for traditional manual auditing.
Since 2023, Uptime has been developing AI-assisted approaches to process this evidence more efficiently.
1. Automated data extraction
Uptime uses two complementary approaches to extract information from supporting documents:
Document intelligence: Specialised tools, including Google Cloud Document AI, identify information within maintenance logs, meter readings, financial records and water safety documentation, converting it into structured datasets.
Vision Language Models (VLMs): General-purpose AI models, including Claude Sonnet, process more complex documents such as spreadsheets with multiple tabs and non-standard layouts.
These tools help transform unstructured evidence into information that can be systematically compared against reported performance.
2. Identifying and investigating discrepancies
Once information has been extracted, automated processes compare supporting documentation against the results submitted by service providers.
Potential discrepancies are flagged for further investigation, helping auditors focus attention on records that may affect payment calculations.
The paper also explores Retrieval-Augmented Generation (RAG), an emerging application that allows auditors to query original source documents directly.
Rather than manually searching through hundreds of pages, auditors could ask whether a reported breakdown, meter reading or financial transaction is supported by the submitted evidence and receive a response linked to the relevant source documentation.
3. Combining AI with human oversight
Uptime's approach combines AI-assisted document processing with rule-based automation and human review.
AI supports the extraction and investigation of evidence, while automated scripts perform structured comparisons and classify discrepancies.
Human oversight remains essential for interpreting ambiguous findings, assessing significant discrepancies and determining appropriate responses.
This combination supports a more efficient and traceable audit process without removing the accountability mechanisms required for results-based funding.
Key Findings
Processing more than 200,000 records
During Uptime's 2025 audit cycle, AI tools processed more than 200,000 performance records from approximately 6,500 pages of supporting documentation.
This represented more than twelve times the volume of records extracted during the previous audit cycle from a comparable number of files.
AI processing costs of less than USD 350
The total AI processing cost for the 2025 audit cycle was less than USD 350.
This demonstrates the potential for AI to significantly increase the volume of evidence reviewed while keeping direct technology costs low.
Making national-scale verification affordable
The paper estimates that a national results-based programme covering 100,000 waterpoints could process approximately 8,000 to 10,000 pages of supporting documentation annually at an AI processing cost of approximately USD 500 to 600.
These estimates suggest that AI processing costs need not increase proportionally with programme size.
A critical factor is the use of probability-based audit sampling, which allows verification to focus on a statistically selected subset of waterpoints rather than reviewing every site individually.
The technology is only one part of verification costs
While AI processing costs are low, comprehensive verification also requires site visits, human review, data management and programme administration.
Across Uptime's existing portfolio, the combined costs of validation, verification and payment administration average less than 9% of total results-based contract value.
The paper identifies opportunities to reduce these costs further through improvements in data systems, sample design and automated audit processes.
Implications for Results-Based Funding at Scale
The findings demonstrate how AI-assisted auditing can help address one of the practical barriers to scaling results-based financing: the cost and complexity of verifying performance across large programmes.
For governments and development banks, affordable verification systems are essential to ensuring public funding is allocated against credible results.
The paper identifies three important considerations:
Cost efficiency: AI can reduce the resources required to extract and review large volumes of supporting evidence.
Scalability: Automated processing and probability-based sampling can help verification systems accommodate national programmes without proportionate increases in audit costs.
Data integrity: AI is most effective when combined with reliable primary records, standardised procedures and appropriate human oversight.
Although the case study focuses on rural drinking water services, the underlying methodology has potential applications across development programmes where payments depend on independently verified outcomes.
Conclusion
AI-assisted data audits in results-based programmes demonstrates that artificial intelligence can substantially improve the efficiency of performance verification while maintaining the evidence and accountability required for results-based funding.
Uptime's experience shows that large volumes of operational and financial data can be processed at relatively low cost, creating opportunities to extend rigorous verification to programmes operating at national scale.
The paper also emphasises that AI cannot replace reliable primary data, human judgement or institutional accountability. Continued investment in service providers' data systems and verification processes remains essential.
Together, these approaches offer a practical pathway towards more transparent, accountable and scalable results-based funding.
Related reading: The Data Dividend | Frequently Asked Questions