Why Public Sector AI Projects Fail: They Are Designed for Committees, Not Citizens
Artificial intelligence can improve public services, but many government AI projects stall because they are shaped around internal governance, consensus and risk avoidance rather than the citizen experience they are meant to improve.
Public sector AI should make services faster, fairer and more responsive.
In the public sector, artificial intelligence holds enormous promise: faster processing of citizen applications, more accurate risk assessments in social services, predictive maintenance for infrastructure and more efficient allocation of limited resources.
Yet across governments worldwide, including in New Zealand, many AI initiatives stall, deliver underwhelming results or quietly disappear after substantial investment.
The core problem is not usually the algorithm. More often, the project has been designed around internal machinery rather than the person who needs the service.
Public sector AI failure is rarely just a technology failure.
The failure rate for AI projects is high across enterprise and public contexts. Research and industry analysis frequently point to weak problem definition, poor data quality, unclear value, governance friction and implementation barriers as repeated causes of AI project failure.
In public sector environments, those risks are amplified. AI projects must work inside fragmented legacy systems, inter-agency boundaries, privacy obligations, procurement rules, political sensitivity and public accountability expectations.
That makes decision quality essential. Before a public agency asks whether the AI model can work, it should ask whether the decision environment around the project has been properly tested.
Misaligned or diluted objectives from the start.
Public sector projects often begin with broad, aspirational goals shaped in steering groups, workshops and inter-agency consultations. By the time requirements are documented, the original citizen problem can become diluted.
A problem such as reducing wait times for benefit applications may slowly become a specification that satisfies policy, legal, operational, reporting and political requirements, while losing focus on the person waiting for help.
RAND Corporation research on AI failure has highlighted misunderstanding or miscommunication of the core problem as a major risk. In government, that risk is often magnified by the number of actors involved.
The public-facing problem is often clear at the beginning: faster decisions, better access, clearer support or less friction.
Every internal group adds constraints, preferences and controls until the original problem becomes secondary.
The project may satisfy sign-off requirements while failing to meaningfully improve the citizen outcome.
Endless layers of governance and approval.
Public sector AI projects can accumulate committees quickly: project boards, risk registers, ethics panels, procurement teams, privacy impact assessments, technical reviews and senior oversight. Each layer may add scrutiny, but not necessarily better decision insight.
When approval processes stretch for months, pilots lose urgency and the original service problem fades into governance procedure. Risk-averse cultures can end up prioritising “no surprises” over practical learning.
Internal compliance becomes more important than citizen outcomes.
Public sector AI is often optimised for audit trails, explainability to internal reviewers and defensibility in information requests. These things matter, but they can overwhelm the actual service experience if they become the dominant design logic.
Interfaces become clunky to accommodate every possible edge case. Features are stripped out to minimise perceived risk. The system technically works, but citizens experience it as another bureaucratic layer.
Governance should protect citizens, not displace them from the centre of the design.
Data and integration problems are amplified by silos.
Public data is often fragmented across legacy systems, departmental boundaries and inconsistent formats. AI needs clean, integrated and high-quality data to be useful, but public sector projects often defer the difficult work of data sharing, governance and modernisation.
Instead, they settle for narrow pilots using whatever data is easiest to access. That may be enough to demonstrate a concept, but not enough to support a reliable public service.
If the underlying data is inconsistent, incomplete or poorly governed, AI outputs will inherit those weaknesses.
When departments cannot share or connect data effectively, AI projects become narrow and fragile.
Privacy concerns should be designed into the work early, not used late as a reason to avoid difficult decisions.
Projects resist iteration and real user feedback.
Private-sector AI often improves through rapid iteration: build, test with users, learn and refine. Public sector projects shaped by committee consensus often resist change once scoped.
Citizen testing can become tokenistic or too late in the process. Negative feedback triggers more review, rather than fast improvement. The outcome is a system designed for sign-off, not adoption.
The project has to be designed around the citizen problem, not the committee process.
Fixing public sector AI failure does not require abandoning governance. It requires reorienting governance around outcomes, evidence and citizen value.
New Zealand already has useful guidance through the Public Service AI Framework, MBIE Responsible AI Guidance and the broader New Zealand AI Strategy. The challenge is turning guidance into working delivery conditions.
Citizen-centred AI needs lighter, sharper and more outcome-focused assurance.
The answer is not reckless experimentation. It is disciplined, citizen-centred testing before the project becomes too large, slow or politically exposed to change.
Define the specific service friction, delay, risk or unfairness the AI project is meant to improve. Keep that problem visible throughout the project.
Cross-functional teams should be able to define narrow, high-impact scopes without every decision being diluted through committee consensus.
Citizen and frontline feedback should be treated as evidence, not a threat to the project plan.
Data quality, privacy, integration and governance need to be solved as core project conditions, not left as late-stage blockers.
Governance should test whether the project is useful, safe, lawful, explainable and improving the citizen outcome it was created for.
Public sector AI needs decision assurance before procurement and implementation.
Ministry of Insights is built for this kind of pre-commitment testing. The MOI Lab system helps leaders test the decision environment before money, people, data, reputation or public trust are placed at risk.
The cost is not just failed technology. It is lost public trust.
When public sector AI projects fail, the cost is not only financial. It is the lost opportunity to make government services faster, fairer and more responsive at a time when public trust is fragile.
New Zealand’s public sector has strengths: pragmatism, a relatively small scale for testing and growing AI guidance from MBIE and the Government Chief Digital Office. But until AI projects are genuinely designed around citizens rather than committees, many will continue to underperform.
Decision support, responsible AI and public-sector assurance.
For broader context, review MOI’s Lab system, responsible AI position and relevant public-sector AI guidance.