The finding
Public institutions remain answerable for decisions shaped by AI
Together, the seven readings point to a practical requirement: public institutions need more than strategy language to govern AI. Once a system drafts, searches, classifies, predicts or recommends within government, choices about evidence, uncertainty, cost, energy, bias, data and redress enter an administrative file. What matters is whether the institution can test the system, supervise its use and remain answerable for the decision it shapes.
A procurement standard should therefore begin with the public task. Factual performance and the willingness to admit insufficient knowledge are distinct capabilities. Kenyan English, Kiswahili, local records and actual service conditions matter more than a general commercial leaderboard. Evaluation must continue after award because models, data, vendors and operational conditions change.
Decision required
Before approval, assign who can test, reject, halt and replace the system
Before deployment, the approving office should identify the lawful data basis, minimum human capability, locally relevant validation, incident route and officer authorised to stop use. County readiness cannot be inferred from a national strategy. Facilities, connectivity, records, staff, budgets and the ability to correct a harmful output are governing conditions.
Operational sovereignty is similarly more demanding than domestic hosting. It concerns infrastructure dependence, skilled personnel, access to logs and documentation, audit rights, continuity when a supplier fails and credible exit terms. Regional sharing may reduce duplication, but shared capability needs public governance of its own. Across all four questions, responsibility should remain with the authority that accepted the consequential decision.