The Ministry of Housing, Communities and Local Government deployed Extract, an AI tool that automates the processing of complex planning documents, to every local council in England. The system is expected to save the average council around 255 hours of manual work annually.
The ministry is simultaneously running a pilot of Augmented Planning Decisions with three London boroughs, a tool designed to help planning officers interpret local policy. Both systems use Google Cloud infrastructure and Gemini models hosted in a protected environment designed to limit prompt-injection risks and maintain data sovereignty.
The Extract tool was developed by MHCLG and the Incubator for AI. Augmented Planning Decisions was built by Google Cloud, Google DeepMind and delivery partner Faculty. The ministry plans to roll APD out to all 300+ local authorities by 2027.
What this reveals about government AI adoption
The announcement is notable for what it does not claim. There is no talk of transformation. No promises that AI will revolutionise planning decisions. No suggestion that the system will replace planning officers or make judgments about whether a development should be approved.
Extract does one thing: it extracts information from planning documents. It saves time on data entry and document review. A planning officer still reads the summary. Still makes decisions based on that summary.
APD does one thing: it helps officers understand local policy. It does not make the policy decision. The officer does. The AI provides context and interpretation that would otherwise require manual research across dozens of local policy documents.
The model is administrative efficiency, not autonomous decision-making. That distinction matters because it is how governments are likely to deploy AI successfully without triggering the kind of backlash that has met other government AI initiatives.
The infrastructure question
Both systems run on Google Cloud rather than on-premises government infrastructure. This creates dependency on a US vendor. The ministry addressed this by requiring data protection mechanisms and protected environments that limit Google's access to the underlying data.
The compromise is pragmatic. Government IT infrastructure is typically years behind cloud providers in capability and cost-efficiency. Running on Google Cloud means faster deployment and lower maintenance burden. The trade-off is vendor dependency.
The rollout to 300+ councils will test whether that trade-off is sustainable at scale. If the system works, other governments will follow. If it does not, the blame will be distributed across four parties: the government, the vendor, DeepMind and the delivery partner.
That distribution of responsibility is probably intentional. No single entity can be blamed if the system underperforms.
