Government AI That Actually Works: Lessons from Five Deployments That Scaled

The failure pattern is consistent across jurisdictions. A government department commissions a pilot with a motivated team, a curated dataset, and a vendor with strong incentive to make it work. The pilot produces impressive demo results. The department then attempts to scale to full deployment and encounters: legacy IT infrastructure incompatible with the AI system, data quality that deteriorates when the curated pilot dataset is replaced by live operational data, procurement rules that prevent the vendor relationships needed for ongoing model maintenance, and a workforce that was not consulted during the pilot and responds to mandatory adoption with the passive resistance that characterises entrenched bureaucratic change. The technology rarely fails; the ecosystem around it does.

HM Revenue and Customs deployed an AI document classification system for processing tax correspondence at scale — routing incoming letters, forms, and supporting documents to the correct processing team without manual review. The deployment succeeded for three reasons. First, the problem was narrow and well-defined: document classification has clear correctness criteria and does not require AI to make consequential decisions — it routes, humans decide. Second, the training data was native to HMRC: millions of historical documents already correctly classified by experienced staff, providing strong signal without data curation effort. Third, the implementation team included frontline HMRC staff from the start, designing routing logic with their input and addressing their concerns before implementation. The system now processes 4 million documents per month with 94% first-pass accuracy, with human review for the 6% classified with low confidence.

Singapore's Ministry of Social and Family Development deployed Jarvis — an AI system that assists case workers in assessing eligibility for social benefits — across 24 Social Service Offices. The key design decision was advisory rather than autonomous: Jarvis presents a recommended eligibility assessment with the evidence behind it, but a case worker reviews and approves every decision. This design choice was made after extensive consultation with front-line staff who expressed concern about algorithmic decision-making affecting vulnerable citizens. The advisory architecture addressed those concerns while still delivering the efficiency benefit: case worker processing time fell 47%, freeing capacity for the casework interviews and family support activities that cannot be automated.

The United States Air Force's deployment of AI predictive maintenance for the F-35 fleet represents one of the largest government AI programmes that achieved operational status. The programme succeeded where previous defence AI initiatives failed because it was built around operational data that already existed — sensor data from aircraft health management systems — rather than requiring new data collection infrastructure. The ML pipeline was co-developed with the maintainers who would use it, who specified what information they needed from the system to make maintenance decisions (remaining useful life estimates for specific components, not generic risk scores). Integration with the existing maintenance scheduling system was prioritised from day one, ensuring predictions translated directly into actionable work orders. The programme reduced unscheduled maintenance events by 32% in its first operational year.

The Dutch RVO (Netherlands Enterprise Agency) automated the triage and pre-processing of environmental permit applications — a notoriously slow process averaging 26 weeks from application to decision. The AI system performs three functions: completeness checking (identifying missing documents or information at submission, reducing the back-and-forth that previously extended timelines), technical pre-screening (flagging applications that require specific expert review versus those that can be handled by generalist reviewers), and precedent retrieval (identifying previous permit decisions for comparable applications to support consistent decision-making). The system does not make permit decisions — those remain with trained reviewers — but by the time an application reaches a reviewer, it is complete, appropriately routed, and supported by relevant precedents. Average processing time has fallen from 26 weeks to 11 weeks.

The most sensitive of these deployments is New South Wales Family and Community Services' child safety triage system, which assists intake workers in prioritising among the 300,000+ contacts received annually — more than can receive comprehensive human assessment within safe timeframes. The system scores the urgency of investigation response based on structured intake data, with the explicit design constraint that it only elevates priority and never recommends de-prioritisation — reflecting the asymmetric consequences of errors in each direction. The system was built with extensive involvement from frontline workers, an independent ethics review, and a governance structure that requires quarterly auditing of decision patterns for demographic bias. Early outcome data shows a 28% improvement in time-to-investigation for high-risk cases, driven by the system surfacing cases that intake workers would otherwise have triaged as lower priority based on incomplete information.

Across these deployments, five factors consistently distinguish those that scaled from those that did not. First, narrow problem definition: all five targeted a specific workflow task, not a broad capability. Second, advisory architecture: none gave AI systems autonomous decision authority over citizen-affecting outcomes — humans remain in the decision loop. Third, native data: all used operational data that already existed in government systems rather than creating new data collection requirements. Fourth, workforce co-design: front-line staff were involved in designing the system's outputs and workflow integration from early stages, not consulted after the design was fixed. Fifth, measurement from day one: each programme defined specific, measurable success criteria and collected outcome data continuously, enabling evidence-based iteration rather than advocacy-based persistence.

Found this useful?

Share the key finding on X.