Guidance for AI proof of concept to scale: scenarios

A service delivery function achieved an impressive 85% accuracy in triaging citizen complaints and feedback. The AI team was ready to deploy the model, but they hit a wall – the agency's existing IT infrastructure.

To move to production, the model needed a stable deployment environment that could be easily updated and monitored, but the traditional IT change management process required extensive documentation and manual approvals for even minor code changes, creating a bottleneck for a model that needed frequent training and updating.

The project stalled as the IT division, already strained with other priorities, could not dedicate the resource to modernise the infrastructure and change management processes to support the AI model.

Agencies should treat AI development and IT modernisation as a single, unified effort. Establish a governance framework from the start that includes both data science and IT teams. The framework should define clear pathways for data access and model deployment, allowing continuous feedback.

What worked well

• The team achieved strong model performance, demonstrating clear value potential.
• Early identification of infrastructure constraints prevented rushed deployment into an unsuitable IT environment.

Lessons learned

• The solution could not progress because IT infrastructure and change management processes could not support modern AI deployment cycles.
• Manual approvals and traditional processes created bottlenecks inconsistent with the need for frequent model updates.
• Competing IT priorities hindered timely progress, highlighting the need for coordinated planning.
• AI development and IT modernisation should be treated as a unified program, with shared governance, shared resourcing and early alignment on deployment pathways.