Beyond Model Readiness: Institutional Readiness for AI Deployment in Public Systems

Technical Summary: Institutional Readiness for AI Deployment in Public Systems

Problem Statement
Public-sector artificial intelligence systems frequently fail not due to deficiencies in model development or technical performance, but due to a lack of institutional readiness at the point of deployment. While existing responsible AI frameworks (e.g., Model Cards, Datasheets for Datasets, NIST AI RMF) effectively assess artifact-level properties such as accuracy, robustness, fairness, and developer-side processes, they are ill-equipped to evaluate the receiving institution's capacity to govern, operate, and sustain a system. Consequently, technically viable prototypes often stall or fail to scale because of unresolved institutional dependencies, including approval pathways, data-sharing protocols, staffing constraints, fiscal continuity, and legal clarity. This gap is particularly acute in resource-constrained settings where the friction between technical viability and operational reality is highest.

Methodology
The paper introduces Institutional Alignment Readiness (IAR), a pre-deployment framework designed to assess whether a specific institution is prepared to deploy a specific machine learning system responsibly. The framework is grounded in two anonymized operational cases from a large national public education system in the Philippines:

1. Case A: An image-based anthropometric screening tool for early-grade learners.
2. Case B: A speech-analysis system for early learning risk identification.

Both projects reached technically promising stages but stalled during the transition from prototype to broader rollout due to institutional, rather than technical, barriers. The authors analyzed these cases to identify concrete institutional conditions that, if missing, preclude responsible deployment. The analysis focused on factors such as data representativeness, referral pathways, workflow integration, and legal authorization.

Key Contributions
The paper proposes the Institutional Alignment Readiness (IAR) framework, a five-dimensional model for assessing deployment readiness. Unlike maturity models that evaluate an organization's general AI capability, IAR is system-specific and deployment-specific. The five dimensions are:

1. Institutional and Operational Compatibility: Assesses approval pathways, workflow fit with frontline practice, operator training plans, and logistical feasibility across sites.
2. Data Ecosystem Maturity: Evaluates dataset representativeness for the target population, data access/sharing protocols, annotation capacity, and retention policies.
3. Human Oversight Capacity: Determines the availability of qualified reviewers, clear override authority, referral pathways, and mechanisms for non-technical challenge.
4. Fiscal Sustainability: Examines budget clarity beyond the pilot phase, maintenance and retraining plans, and continuity across leadership or funding transitions.
5. Regulatory Alignment Readiness: Verifies privacy compliance, legal basis for data collection/sharing, ethics review, consent procedures, and contestability pathways.

The framework is designed as a staging tool rather than a scoring instrument. It supports decisions to proceed with "no-go," "pilot-only," or "broader deployment" based on the maturity of specific dimensions, acknowledging that some deficits may be absolute blockers while others may only require scoping adjustments.

Results and Case Findings
The analysis of the two operational cases demonstrates that deployment trajectories are often dictated by institutional constraints invisible to standard artifact evaluation:

• Case A (Anthropometric Screening): Despite technical readiness, broader rollout was delayed by the time required to negotiate approvals and data access across multiple sites, the misalignment of data collection with school calendars, and the lack of established referral pathways for identified learners.
• Case B (Speech Analysis): The project trajectory was shaped by data infeasibility, forcing a pivot before deployment. Subsequent deployment was gated by the need to formalize ethics requirements, consent procedures, and referral protocols to ensure outputs were not stigmatizing and were actionable by qualified staff.

In both instances, the systems remained in internal validation or limited pilot phases because the receiving institution lacked the specific conditions required to move to scale, despite the models performing well in internal testing.

Significance and Claims
The paper makes three modest claims regarding the value of IAR:

1. Explicitness: IAR makes recurring deployment questions explicit and assessable before resources are committed, surfacing issues (e.g., data access, referral capacity) that often emerge too late in the process.
2. Decision-Relevance: Institutional readiness is a decisive factor in whether technically viable systems scale. The cases show that deployment decisions are shaped by institutional conditions that standard model evaluation cannot resolve.
3. Non-Redundancy: IAR is not redundant with existing responsible AI tooling. While current frameworks assess the artifact and the developer, IAR assesses the receiving institution. It addresses questions that existing tools are architecturally unable to answer, such as whether the specific cross-agency data sharing required is legally permissible or whether qualified human reviewers actually exist.

The authors position IAR as a complement to, not a replacement for, established AI evaluation tools. It serves as a necessary second layer of assessment to ensure that public-sector AI systems are not only technically sound but also institutionally viable for responsible deployment.