The Zambezi Institute Policy Opinion Paper No. 09/2026
Abstract
Governments across the developing world face persistent operational failures — slow
services, unreliable data, and weak oversight — alongside deeper crises of legitimacy rooted in corruption and unaccountable leadership.
This paper examines whether artificial intelligence (AI) can address either problem. We argue that AI offers measurable gains inoperational effectiveness — bureaucratic automation, predictive planning, service delivery, and anomaly-based oversight — but has no independent capacity to resolve the political conditions that produce dysfunction in the first place: weak political will, thin accountability, and institutional capture.
AI is best understood as an execution technology that amplifies the intent of whoever controls it, for better or worse. We propose a policy framework centered on targeted, transparent, and institutionally insulated AI deployment, paired with
human capacity-building and citizen feedback mechanisms. We conclude that AI cannot substitute for institutional reform, but can reduce the human cost of dysfunction while reform proceeds.
Keywords: artificial intelligence, governance, public administration, institutional reform, government effectiveness, accountability.
1. Introduction
The question of whether artificial intelligence can improve government performance has moved from academic speculation to active policy debate. Citizens in low-capacity states increasingly ask a version of the same question: if AI can automate a bank’s fraud detection
or a company’s logistics, why can’t it fix a passport office, a tax authority, or a hospital
supply chain?
This paper takes that question seriously while resisting a common conflation in publicdiscourse — the assumption that “government dysfunction” is a single problem with a single technological fix. We distinguish between two analytically separate failures: low effectiveness (poor execution of agreed policy) and low legitimacy (the absence of political accountability). We argue that AI has real, demonstrable purchase on the first and
essentially none on the second, and that policy design should proceed from this distinction rather than obscure it.
The paper proceeds as follows. Section 2 sets out the conceptual distinction between
effectiveness and legitimacy. Section 3 identifies the domains where AI deployment
produces measurable operational gains. Section 4 examines the structural limits of AI as a governance tool. Section 5 sets out policy recommendations. Section 6 discusses the risks of misapplied AI deployment. Section 7 concludes.
2. Conceptual Framework: Effectiveness Versus Legitimacy
Government dysfunction is commonly treated as a single undifferentiated condition, but it typically manifests through two distinct channels. Effectiveness failures are operational: slow services, lost records, poor planning, weak execution of otherwise sound policy. Legitimacy failures are political: corruption, patronage, unresponsive or unaccountable leadership, and institutions captured by narrow interests.
This distinction matters because the two failure types respond to entirely different
interventions. Effectiveness failures are, in large part, engineering problems — solvable
through better systems, better data, and better process design. Legitimacy failures are power problems — they persist not because a solution is unknown, but because those with the power to fix them frequently benefit from leaving them unfixed.
Treating AI as a solution to both problems at once leads to two predictable failure modes:
policymakers overpromise what AI can deliver, and evaluators later judge AI pilots against a standard — improved legitimacy — that the technology was never capable of meeting. We propose that policy design, and evaluation, should explicitly separate the two.
3. Where AI Delivers Measurable Gains
Evidence from early government deployments and adjacent private-sector applications points to four domains where AI produces measurable operational improvement.
3.1 Bureaucracy and verification.
Automated processing of identity documents, licenses,
and tax filings can shorten processing times and reduce the number of discretionary human touchpoints where informal payments are typically extracted. This does not eliminate corruption, but it narrows the surface area available for it.3.2 Data and planning. Predictive analytics can convert fragmented administrative records
in health, education, and infrastructure into decision-relevant information, allowing resource allocation ahead of a crisis rather than in response to one.
3.3 Service delivery. Continuously available AI systems — chatbots, automated routing for emergency services and utilities — can extend basic service access without requiring a prior overhaul of the surrounding institution.
3.4 Oversight. Anomaly-detection systems applied to procurement and payroll data can identify ghost workers, duplicate vendors, and inflated pricing considerably faster than manual audit processes.
The common feature across all four domains is that none presupposes a well-functioning or high-integrity government as a precondition. Each requires digital infrastructure and political permission to deploy — not institutional reform.
4. Structural Limits of AI in Governance
4.1 Political will. AI cannot compel a leader to act in the public interest. Where the
governing objective is the entrenchment of power rather than public service, the same tool are readily redirected toward surveillance, censorship, or targeted patronage. The technology is indifferent to which objective it serves.
4.2 Accountability.
An algorithmic system can identify a problem; it cannot prosecute, dismiss, or remove the official responsible for it. In the absence of independent courts, oversight bodies, or a press able to act on algorithmic findings, those findings carry no enforcement weight.
4.3 Capacity and infrastructure.
AI deployment presupposes reliable electricity, connectivity, usable data, and trained personnel — none of which can be assumed in many low-resource settings. Where underlying data is poor, automation compounds rather than corrects the problem, producing flawed decisions at greater speed and with greater apparent authority than a human analyst would typically claim.
4.4 Value judgments.
Resource-allocation questions — a rural clinic versus an urban highway — involve trade-offs that a model can describe but cannot resolve. Presenting an algorithmic output as a resolution to what is, in substance, a political choice obscures rather
than removes the underlying politics.
5. Policy Recommendations
Based on the preceding analysis, we propose five design principles for AI deployment inlow-capacity governance settings:
1. Start small and start visible. Pilot AI deployment in two or three high-frequency citizen services — business registration, social grants, hospital supply chains — with success measured in time and cost savings rather than announcement volume.
2. Mandate algorithmic transparency.
Decision logic, error rates, and audit outcomes should be published and subject to the same civil society and media scrutiny applied to public budgets.
3. Institutionalize independence.
Core AI systems, particularly in revenue collection, statistics, and anti-corruption functions, should sit within agencies holding statutory protection from political interference rather than under direct ministerial control.
4. Invest in human capacity alongside
technology. Civil servants require training to work alongside AI systems rather than be displaced by them; every deployment should be paired with a dedicated data and audit function.
5. Build citizen feedback mechanisms.
Every AI-administered service requires a
functioning channel through which users can appeal decisions, report errors, and rate
outcomes. Effectiveness gains that citizens do not trust are unlikely to prove durable.
6. Discussion
The central risk identified in this analysis is not that AI will fail to improve government
operations — the evidence suggests it frequently can — but that operational gains will be mistaken for, or substituted for, institutional reform. A government can become measurably faster and more efficient while becoming no more accountable, and in some cases less so, if
the same tools are deployed toward surveillance or control rather than service.
This is not a technical risk but a governance risk, and it cannot be engineered away. It canonly be managed through the institutional safeguards outlined in Section 5 —independence, transparency, and functioning feedback loops — each of which depends on political actors choosing to accept constraints on their own use of these tools.
This dependency is itself the paper’s central finding: AI’s value to governance is entirely
conditional on the political environment into which it is introduced.
7. Conclusion
AI will not resolve the underlying conditions that produce dysfunctional government. It can, however, reduce the day-to-day cost of that dysfunction for the citizens who depend on public services, while slower institutional reform proceeds in parallel.The relevant policy choice is not “AI or reform” but “AI while reform continues.”
Deployedwith appropriate safeguards, AI can save time, save money, and rebuild incremental public trust through improved service delivery. Deployed without those safeguards, it becomes an unusually efficient instrument for the same dysfunction it was meant to address — because it is, at its core, very good at doing precisely what it is directed to do.
The determining factor is not technical capability but political choice. Author’s Note: This is a policy opinion paper reflecting analysis and argument by the
Zambezi Institute rather than an empirical study; it does not report original data collection and should be read as a conceptual and policy framework rather than a quantitative evaluation.
The Zambezi Institute Research for policy. Policy for people.
