AI in government isn’t failing. We’re building it wrong

Expectations of governments are rising. Citizens increasingly demand services that are faster, more responsive, and more reliable, while trust in government processes and outcomes is under strain. This creates a dual challenge: improving service performance while strengthening legitimacy.

A central part of this response lies in GovTech: the use of digital technologies to run government operations and deliver public services. Unlike Civic Tech, which focuses on citizen engagement, GovTech concerns the systems that underpin the state.

These systems include case management tools, compliance dashboards, procurement platforms, and shared data infrastructures. Increasingly, GovTech includes AI-enabled systems such as automated welfare eligibility assessments, predictive risk models in social services and policing, fraud detection in tax administration, and algorithmic decision-support in migration and regulatory contexts.

Shaping how decisions are made, resources allocated, and services delivered, these processes often operate at scale and with limited visibility. In AI-enabled contexts, challenges are compounded by issues of data quality, model opacity, and integration into existing workflows. Systems designed in top-down or vendor-driven ways, with limited input from users, risk misalignment, low adoption, and contested outcomes.

This is where co-production becomes critical.

Our new research project from the Centre for Artificial Intelligence in Government (CAIG) on ‘Co-Producing AI in Government’ examines how collaborative approaches can support the design, implementation, and assessment of AI-enabled GovTech.

In AI-enabled GovTech, co-production refers to structured collaboration between government agencies, frontline administrators, business users, digital teams, and technology providers across the service lifecycle. It reframes systems as outcomes of interaction between actors with different forms of context and expertise.

Participation across commissioning, design, delivery, and assessment is particularly important in AI systems, which depend on institutional data and context-specific decision environments. Early engagement shapes problem definition and data use; continued involvement supports alignment, monitoring, and adaptation over time.

Co-production can strengthen usability, accountability, and trust, but it is not without limits. Power asymmetries, resource constraints, and institutional inertia often shape who participates and how. In AI contexts, these challenges are intensified by technical complexity. The question is not simply whether co-production occurs, but how it is structured and under what conditions it leads to meaningful outcomes.

Our research questions how co-production can be embedded in AI-enabled GovTech, focusing on lifecycle participation and the conditions that shape outcomes such as efficiency, compliance, and trust. The answers to our questions have implications beyond academic debate.

For governments, co-production offers pathways to implement AI systems that are effective and accountable, whilst technology providers can reduce implementation failure by grounding system design in institutional context. For citizens, it speaks directly to trust, fairness, and legitimacy in AI-mediated public decisions.

This research project will contribute to ensuring that AI systems in government are effective, safe, accountable, and aligned with public value. It aims to answer a simple but urgent question: if AI is to shape government, who gets to shape AI?

Akshara Baru, Dr Martin Waehlisch, and Professor Nicole Curato