In 2025, 49% of innovations submitted to the Gartner Eye on Innovation Awards for Government included an AI component. That figure was 8% in 2020. The jump is significant, but the real news is not the growth in numbers: it is the quality of what those numbers represent. Increasingly, these are operational projects with measurable KPIs and documented impact, not proof of concepts.
What separates the projects that work from those that stall? The cases recognized in 2025 offer a concrete answer.
Real efficiency, not generic savings
The Banco Central do Brasil developed Axis, an LLM-based tool that analyzes audit reports in natural language from over 1,400 organizations every six months. Before implementation, the process took several weeks. With Axis, the same work is completed in 120 minutes, with a validated accuracy of 99.03%. Staff can now focus on interpreting data patterns and intervening early on anomalous situations.
The Ile-de-France region used an LLM with access to a complete IT technical documentation library to handle support requests from employees and high school students. Average resolution time dropped between 15% and 50% depending on incident complexity, and total incident volume decreased between 5% and 20%. Costs related to incident escalation fell between 10% and 25%.
Cafcass, the UK family court support service, was manually personalizing 80,000 letters per month. A slow process with risk of delays and generic communications. Integrating a letter-writing assistant into the case management system reduced production time by 28%, saving approximately 1,365 person-hours per year. ROI is expected within five years, with an average annual return exceeding 10% over a decade.
Public mission with measurable results
The city of Scottsdale, Arizona, faced a concrete problem: the growth of short-term rentals was generating community tensions over noise, waste, and parking. The Map Tool and Resource Center, built on a GIS platform, uses machine learning to automatically identify non-compliant properties on major rental platforms. Results: unauthorized short-term rental identification increased by 30%, complaint resolution times reduced by 20%, unresolved complaints decreased by 15%.
The UK's MHRA — the regulatory agency for medicines and medical devices — developed three distinct tools: an AI system for clinical trial review, an automated system for counterfeit drug detection, and a RAG assistant for regulatory sciences. Combined results include a 35% reduction in clinical trial assessment times and a drop in time spent on counterfeit drug investigations from 30 hours to 3 hours. Annual savings amount to 2.4 million pounds.
The Georgia Department of Corrections developed GHOST, a system that integrates data from over a dozen detection systems to identify and track illegal phones in prisons. The system contributed to the identification and deactivation of nearly 8,000 contraband devices, reducing analysis times from months to 20 minutes.
Hybrid technology, not LLM monoculture
One of the most relevant findings from the analysis of recognized cases is the combined use of different AI techniques. None of the most successful projects relies exclusively on LLMs or generative AI. Non-generative machine learning models, rule-based systems, graph databases, and optimization algorithms frequently appear in combination with generative components.
The city of Taoyuan, Taiwan, replaced static traffic control systems with an AI-based adaptive network. Cameras with YOLO object recognition and edge computing analyze traffic in real time and adjust signals accordingly. Emergency response times dropped from 4 minutes to 45 seconds. The system saves 49,000 liters of fuel per year and reduces CO2 emissions by 111 tons.
The Dubai Police developed a training program based on VR crime scenes with interactive forensic tools and automated performance analysis. Post-training assessment scores increased by 32.8%, real-case performance improved by 57.2%, and the total cost over five training cycles was 826,558 AED less than the traditional method.
What makes these projects different
The projects that deliver results share several structural characteristics. They start from a specific operational problem with defined KPIs, not from a generic "efficiency" goal. They integrate AI with foundational organizational and technological improvements: data consolidation, process unification, integration of systems that previously worked in silos. They incorporate responsible AI practices from the design phase, which accelerates adoption and reduces legal challenges. They are designed to be scalable and replicable: the Taoyuan case, for example, inspired the national program and led to the creation of a knowledge-sharing platform among cities.
The "build vs buy" model in successful cases is almost always hybrid: commercial cloud services combined with open source tools and internal development, calibrated to the specific needs of the public mission. Projects developed entirely in-house often do so for reasons of digital sovereignty or data security, not for lack of commercial alternatives.
The main lesson is as simple as it is hard to apply: AI in the public sector does not deliver value when used to do the same thing as before slightly faster. It delivers value when it solves a real bottleneck, with clear metrics, in an organizational context ready to change how it works.
