Deeper AI integration links to stronger monitoring-and-evaluation systems and greater evidence use in African public management, but evidence comes from a small cross‑sectional practitioner survey and cannot prove causation.

Main Finding

Deeper, organization-wide AI integration among African M&E practitioners is positively associated with stronger M&E system effectiveness—especially System Quality and Information Quality—and stronger M&E effectiveness is in turn strongly associated with higher Organizational Evidence Capability and Net Benefits. These relationships persist after adjusting for M&E-system maturity. AI use also corresponds with faster reporting, better data quality, and greater evidence use in decisions; it was not associated with shorter data-processing time. The study interprets AI as a complementary capability whose value depends on integration into institutional evidence processes rather than mere access to tools.

Key Points

• Theoretical framing: adapts DeLone & McLean information‑systems success model to M&E (System Quality, Information Quality, Service Quality) and positions AI-use intensity as a reinforcement/complementarity layer, not a separate success dimension.
• Focal construct: AI-use intensity measured on a 0–5 ordinal scale (no use → organization-wide integration); M&E maturity measured separately (1–5).
• Significant bivariate/adjusted associations:
  • System Quality: Kruskal–Wallis H = 42.26, ε2 = .553, p < .001; adjusted cumulative odds ratio (OR) = 2.75 (95% CI [2.00, 3.78]).
  • Information Quality: H = 26.98, ε2 = .338, p < .001; adjusted OR = 1.94 (95% CI [1.34, 2.82]).
  • Service Quality: H = 12.73, ε2 = .137, p = .005; adjusted OR = 1.60 (95% CI [1.01, 2.55]).
• M&E System Effectiveness → Organizational outcomes:
  • Organizational Evidence Capability (RBM, KIM, EBDM): OR = 10.82.
  • Net Benefits: OR = 7.15.
• Operational outcomes: deeper AI use corresponded with improved reporting efficiency, higher data quality, and increased evidence use in decisions; no statistically significant effect on data-processing time.
• Measurement quality: multi-item scales with high internal consistency (Cronbach’s alpha .896–.979).
• Main caveats: cross-sectional design, purposive non-probability sample (n = 75 across 19 African countries), self-reported measures—associations not causal.

Data & Methods

• Design: Cross-sectional analytical web survey (English/French), June–July 2026; purposive sampling of M&E professionals; n = 75 observations.
• Unit of analysis: a specific M&E system, programme, project, or professional assignment referenced by each respondent.
• Key variables:
  • AI-use intensity (0–5 ordinal: no use → organization-wide integration).
  • M&E system maturity (1–5 ordinal).
  • Outcome constructs: System Quality (8 items), Information Quality (9 items), Service Quality (5 items); composites for M&E System Effectiveness, Organizational Evidence Capability (RBM, KIM, EBDM), and Net Benefits.
  • Operational indicators: processing time, reporting time, data quality, evidence use.
• Analytical approach:
  • Preserved ordinal measurement properties.
  • Nonparametric Kruskal–Wallis tests and rank-based associations for bivariate analyses.
  • Adjusted cumulative‑logit (ordinal logistic) models controlling for M&E maturity to estimate cumulative odds ratios.
  • Qualitative open-ended responses were coded and harmonized to characterize tool use and functional scope (ML, NLP, generative AI, computer vision, geospatial AI, dashboards, etc.).
• Limitations noted by authors: non-representative sampling, self-report bias, potential reverse causality (mature systems may enable deeper AI integration), small sample size limiting external generalizability.

Implications for AI Economics

• Complementarities matter: The study supports the view from AI-economics that returns to AI investments depend heavily on complementary organizational capabilities (processes, governance, skills). Simple access to AI tools is unlikely to produce maximal returns without institutional integration.
• Measurement and metrics: “Use intensity” (depth of integration) is a more informative variable than binary adoption for estimating productivity/benefit relationships in empirical work on AI. Economists should prefer measures that capture operational integration across functions.
• Heterogeneous returns and thresholds: The large ORs for System Quality and Organizational Evidence Capability suggest non-linear or threshold effects—deep integration may yield disproportionate benefits. Modeling production functions for AI should allow interactions between AI capital and organizational (human/institutional) capital.
• Policy and investment priorities: For public-sector settings (and likely many low-/middle-income contexts), investments should pair AI tools with investments in institutionalization (M&E maturity), training, data governance, and decision‑use processes to capture value. Subsidizing hardware/software alone risks low returns.
• Task- and outcome-specific effects: The lack of association with data-processing time but positive association with reporting efficiency and data quality highlights that AI effects vary by task and outcome. Economic evaluations should disaggregate effects (speed vs. accuracy vs. decision quality).
• Research agenda for AI economics:
  • Causal studies (RCTs, difference-in-differences, instrumental variables) to identify returns conditional on complementarities.
  • Larger, representative datasets across sectors and countries to estimate heterogeneity and general equilibrium effects.
  • Cost-benefit and cost-effectiveness analyses that incorporate governance, risk (e.g., misinformation/fabrication from generative models), and capacity-building costs.
  • Structural models of adoption that include threshold/complementarity parameters between AI capital and organizational capital.
  • Consider labor-market implications where AI redistributes expertise (skill-biased amplification vs. task automation) in public management and evaluation occupations.

Limitations to keep in mind when applying these implications: the evidence is cross-sectional, self-reported, from a small purposive sample in Africa; causal magnitudes and broader external validity remain to be established.