AI recommendations in Bengkulu’s provincial budgeting are linked to less budget padding and higher employee motivation, suggesting algorithmic advice can both tighten fiscal oversight and boost engagement; however, evidence derives from a single cross‑sectional survey and cannot firmly establish causality.

Main Finding

AI-based budget recommendations in the Bengkulu Provincial Government significantly reduce budget slack and increase employee motivation. Employee motivation both directly reduces budget slack and partially mediates the effect of AI recommendations on slack—so AI operates as a technical control and a behavioral governance mechanism.

Key Points

• Hypotheses tested and supported:
  • H1: AI recommendations → negative effect on budget slack (significant).
  • H2: AI recommendations → positive effect on employee motivation (significant).
  • H3: Employee motivation → negative effect on budget slack (significant).
  • H4: Employee motivation mediates the AI → budget slack relationship (significant mediation).
• Theoretic framing: Agency Theory, Goal‑Setting Theory, Technology Acceptance Model / Algorithmic Trust, and Self‑Determination Theory are combined to explain how AI reduces information asymmetry, clarifies targets, and influences motivation/ethics.
• Measurement: Three constructs (AI Recommendation, Employee Motivation, Budget Slack) measured with multi‑item 5‑point Likert scales. Indicators showed good convergent and discriminant properties (outer loadings ≥ 0.792; AVE: AI 0.676, Motivation 0.712, Slack 0.665).
• Reliability: strong internal consistency (Cronbach’s α: AI 0.884, Motivation 0.901, Slack 0.875; Composite Reliability: AI 0.912, Motivation 0.928, Slack 0.906).
• Sample and context: 140 valid respondents (from 160 distributed; 145 returned), purposive sampling of civil servants involved in budgeting across OPDs in Bengkulu. High exposure to digital budgeting: 87.9% reported using AI/digital budgeting systems. Balanced participant roles (planning, finance, structural, technical) and experience levels.
• Analysis approach: Structural Equation Modeling (SEM; covariance‑based or PLS conditional on data distribution). Mediation tested via bootstrapping/Sobel as described. Results reported as statistically significant at p < 0.05 (direction consistent with hypotheses).
• Practical nuance: while AI recommendations lower slack overall, authors note possible resistance or new adaptive behaviors if AI is perceived as unfair or reduces perceived autonomy—hence the behavioral channel (motivation/trust) matters.

Data & Methods

• Design: Cross‑sectional survey of government employees directly involved in budgeting in Bengkulu Province.
• Sampling: Purposive sampling with inclusion criteria: involvement in budgeting, ≥1 year experience, and exposure to digital/AI budgeting systems. Final N = 140.
• Variables and indicators:
  • AI Recommendation (X): accuracy, ease of use, clarity, trust, perceived contribution (5 indicators).
  • Employee Motivation (M): intrinsic/extrinsic motivation, involvement, willingness to meet targets (4 indicators reported).
  • Budget Slack (Y): inflated cost estimates, lowered performance targets, safety margins, justifications for manipulation (4 indicators reported).
• Measurement: 5‑point Likert scales.
• Validity & reliability: Outer loadings ≥ 0.792; AVE values > 0.50; Cronbach’s α and Composite Reliability > 0.70.
• Inferential analysis: SEM to estimate direct effects and mediation; mediation assessed via bootstrapping / Sobel test. Hypotheses accepted at p < 0.05. (The paper reports significant path estimates consistent with the hypotheses; exact path coefficients and p‑values are reported in the full paper.)

Implications for AI Economics

• Public finance efficiency: AI recommendations can reduce deliberate over‑budgeting, improving allocative efficiency and lowering wasteful slack in public budgets—supporting arguments for digitization investment in public finance systems.
• Behavioral channel matters: Economic models of algorithmic governance should incorporate behavioral responses (motivation, trust, perceived autonomy). AI’s effect on outcomes is not purely informational; it also changes incentives and norms.
• Design and governance of recommendation systems: To realize efficiency gains, policymakers should prioritize transparency, explainability, perceived fairness, and usability to build trust—these features increase motivation and adoption and help avoid resistance or compensatory slack behaviors.
• Complementary policies: Combine AI recommendations with participatory processes and incentives (recognition, performance evaluation) rather than replacing human participation. Preserving elements of autonomy and involvement helps sustain intrinsic motivation, which further reduces slack.
• Caution on generalizability and dynamics: Results are cross‑sectional and context‑specific (Bengkulu provincial government). Future research/policy pilots should:
  • Test causality with longitudinal or experimental designs.
  • Evaluate heterogeneity by role, tenure, and local institutional settings.
  • Monitor for emergent strategic behaviors as algorithms are adopted (new forms of slack or gaming).
• Cost‑benefit and scaling: Potential fiscal savings from reduced slack should be weighed against implementation, training, and oversight costs; economic evaluations of scaling AI across governments should include behavioral adjustment costs and governance investments.

If you want, I can (a) extract the reported path coefficients and p‑values from the full results tables, (b) produce a short policy checklist for implementing AI budgeting tools focused on trust and motivation, or (c) draft possible extensions for a replication study in another region. Which would you prefer?