AI in government: the stakeholders, roles, costs and safeguards
Who Decides What Government Automates?
Five recent publications examine the choices behind public artificial intelligence (AI): which services deserve investment, who scrutinises adoption, how sensitive data travel and whose account of progress reaches the public record.
In this week’s edition
The five publications examined in this edition were published from 6 to 8 October 2026. None featured in an earlier edition.
Before the decision is settled
Who can still change the proposal?
The central issue is when public judgement enters the process. A legislature can question a proposed purchase before a supplier becomes indispensable. A clinician can help define a useful application before a pilot determines the workflow. A person whose data are reused can need protection long before any harm becomes visible. These decisions shape the terms on which government adopts technology.
A hearing, a clinical evaluation and a privacy review examine different parts of a public decision. The five readings show why each needs evidence that can change what happens next.
Five different responsibilities inform the public decision. Editorial analysis of this edition’s readings.
Parliamentary scrutiny
Institutional practice paper
Scrutiny belongs before the purchase
Published by United Nations Development Programme ·
Parliaments and Artificial Intelligence: Governing AI in the Public Interest
The United Nations Development Programme (UNDP) sets out parliamentary routes into AI governance through legislation, executive scrutiny, budgets and public representation. Its account draws partly on a November 2025 meeting involving more than 200 parliamentarians and staff from 65 countries. Participants raised concerns about concentrated technological power and the gap between parliamentary ambition and expertise.
Why it matters for governance
Kenya
A useful parliamentary or county assembly hearing would begin with the service affected and the decision that officials propose to automate. Members need independent technical advice, access to contract terms and a costed explanation of the alternatives. Scrutiny must distinguish a national policy mandate from the county's ability to implement a particular service safely.
Global South
Where a legislature has a small research service and suppliers command far greater expertise, shared parliamentary research capacity can reduce dependence on vendor explanations. Regional co-operation is useful when it equips members to interrogate a concrete contract or regulatory proposal, rather than simply multiplying declarations.
Wider world
Budget approval and procurement scrutiny can expose consequential choices before a specialised AI law is complete. The question for democratic institutions is whether they can obtain evidence independently and hear affected people while a proposal can still change.
Governance analysis drawing on UNDP. The graphic explains a responsibility, rather than claiming a measured effect.
The public decision is already under way when the budget and specification are written.
African AI adoption
Regional economic report
A useful application must justify its place in the budget
Published by World Bank ·
Africa Economic Update, October 2026: Building AI-Readiness
The World Bank favours affordable, locally adapted AI applications supported by electricity, connectivity, skills, usable data and effective institutions. It identifies Kenya, Nigeria and South Africa as centres of regional activity, while most countries remain at an early stage of adoption. The immediate opportunity is to adapt practical applications to pressing development needs, with attention to the foundations required to operate them.
Why it matters for governance
Kenya
Reported activity in Kenya should not be mistaken for uniform readiness across county services. A county considering an AI application should compare it with simpler improvements to the same problem, such as better records or changes to staff workflow. The appraisal needs actual service demand, running costs and the capacity of the office expected to use it.
Global South
Adapting an existing system may preserve scarce resources, but it still creates choices about language, access to data and supplier dependence. Regional purchasing and technical support can improve bargaining power if participating institutions retain the ability to change provider and judge whether the service works.
Wider world
The policy value of AI adoption lies in what it improves for users and workers. Countries should compare alternative ways of achieving that improvement, including a non-AI option. The size of a model or a computing facility is a poor substitute for evidence about service performance.
Governance analysis drawing on World Bank. The graphic explains a responsibility, rather than claiming a measured effect.
The investment case begins with a public problem and a credible comparison of solutions.
AI in health services
Conference proceedings and action plan
Permission to pilot is only the first decision
Published by Organisation for Economic Co-operation and Development ·
Proceeding report from Madrid conference on the responsible scale of AI in healthcare: Scaling fast while doing no harm
The proceedings draw on a Madrid meeting of 142 experts from 34 countries and 35 partner organisations, held on 28–29 May 2026. The resulting ten-action plan covers service design, staff capability, oversight across the AI lifecycle and economic evaluation. It connects expansion to the ability to assess use and effects in real health systems.
Why it matters for governance
Kenya
For county health services, the first practical question is which task would change and who would remain responsible for it. A diagnostic aid, appointment tool and automated clinical note have different consequences. Before expansion, county teams need local evidence on errors, additional workload, patient access and the total cost of continuing the service.
Global South
A pilot supported by a donor or vendor can conceal the staffing, connectivity and support costs that fall to a health system later. Procurement should make renewal conditional on locally relevant evidence and secure access to performance information when external support ends.
Wider world
Health regulators, purchasers and providers judge different things. Product approval does not answer whether an application fits a particular workflow, improves care or merits wider public funding. Each institution needs evidence appropriate to its decision, with a continuing route for staff and patients to report problems.
Governance analysis drawing on OECD. The graphic explains a responsibility, rather than claiming a measured effect.
Expansion deserves a fresh decision based on local performance and continuing cost.
Sensitive data and secondary use
Policy paper with company evidence
A device can pass review while its data travel further
Published by Organisation for Economic Co-operation and Development ·
Neurodata governance: Evidence from innovators for public policy
Neurodata are signals and information derived from the brain and nervous system. The Organisation for Economic Co-operation and Development (OECD) examines practices reported by 16 companies headquartered in seven countries. It describes difficulties controlling downstream use by partners and uneven engagement with data-protection authorities, alongside medical-device oversight. Intended treatment, research and later analytical uses can therefore raise different governance questions.
Why it matters for governance
Kenya
If a public facility considers equipment or software that processes this information, procurement and privacy review should follow the data beyond the device. Officials should identify each recipient, purpose, retention period and permitted onward use. The relevant national authorities and the facility must agree responsibilities before deployment, including how a patient can question a later use.
Global South
Imported devices can leave public institutions dependent on overseas processing and contractual information. Purchasers need intelligible data-flow descriptions and enforceable restrictions on reuse, with the expertise to assess what those terms mean. Supplier assurances are insufficient when the institution cannot see subsequent handling.
Wider world
Regulatory co-ordination needs to follow the use of information across organisations. A finding about product safety answers a different question from whether later data processing is legitimate. Oversight should identify where one body's examination ends and another body's responsibility begins.
Governance analysis drawing on OECD. The graphic explains a responsibility, rather than claiming a measured effect.
The full path of the data belongs in the public decision to buy and use the technology.
Participation and the public record
Review of national reporting
Consultation counts do not show who changed the conclusion
Published by United Nations Development Programme ·
SDG 16: 2026 Voluntary National Reviews through the Lens of Peaceful, Just, and Strong Institutions
Among 36 national reviews, UNDP records civil-society and private-sector engagement in 35 each, and parliamentary engagement in 11. It also describes Norway's inclusion of separate government and civil-society assessments. The study concerns reporting on Sustainable Development Goal 16, which addresses peace, justice and strong institutions.
Why it matters for governance
Kenya
National reporting can make county experience more useful by preserving disagreement and tracing submissions into final findings. A public participation record should show what was proposed, the responsible body's response and which decision changed. County assemblies need access to that record as part of scrutiny, rather than relying on a list of meetings held.
Global South
Reporting requirements often reward a finished document while leaving local evidence systems under-resourced. Administrative data, accessible consultation and the ability to publish a dissenting assessment deserve institutional support. Countries can learn from the form of a review without assuming that its reported arrangements work equally well elsewhere.
Wider world
The same test applies when governments report digital transformation: who can question the official account, and where is that challenge preserved? A record of adoption should distinguish the number of systems introduced from changes in service access, errors and remedies.
Reported engagement in 36 national reviews. Groups overlap; participation is not a measure of influence. United Nations Development Programme, Figure 10.
Public reporting becomes more useful when readers can see how evidence and disagreement shaped the decision.
The governing question
Public judgement needs a place in the decision
This week's publications put the right to judge technology beside the capacity to use it. Sovereignty has an institutional dimension: public bodies need sufficient knowledge and authority to choose an application, set its terms and decide whether it deserves to continue. A locally hosted system does not settle those questions by itself.
The evidence also exposes different objects of scrutiny. Legislators examine authority and spending. Health teams test service performance. Regulators follow data uses and product risks. Citizens and their representatives question the official account. These roles can reinforce one another only if each receives information that is useful for its own decision.
The practical discipline is to make the consequential choices visible before they become expensive to reverse, and to preserve the evidence needed to revisit them.
Reflections for policy work
Three questions to carry into the institution
Which decision remains open?
Before committing funds, can the institution explain the problem, compare an AI proposal with other solutions and identify who may reject or change it?
Who can examine the evidence?
Do legislators, service staff, regulators and affected people have the information and expertise needed to judge the aspects for which each is responsible?
What would justify continuation?
Which results, costs or harms would trigger renewal, modification or withdrawal, and how will the public record show that the institution acted?
Original publications
Sources and publication dates
The research accounts report the publications’ claims. The governance readings, interpretive diagrams and synthesis are Ahmednadhir.com’s analysis. The sources differ in method and reach; none establishes that an AI programme improved public services.
