UK planning AI tests Google Cloud at national scale

The UK government has put Google Cloud at the center of a new test for public-sector AI: turning England’s planning backlog into a software, data and operations problem that councils can manage at national scale.

The Ministry of Housing, Communities and Local Government and the Department for Science, Innovation and Technology said June 16 that two AI tools will support planning officers. The first, Extract, helps councils convert planning documents, maps and handwritten records into usable data. The second, Augmented Planning Decisions, or APD, helps officers assess routine householder applications.

The government said officers in Barnet, Camden and Dorset started alpha trials for APD in May 2026. Officials plan to expand trials to up to 10 more councils later this year and pursue a national rollout from 2027 if the prototype works.

The GOV.UK announcement frames APD as a tool that could cut routine planning decisions from eight weeks to four weeks in an average case. Qualified planning officers retain approval authority, which gives the program a clear governance boundary: AI can collect, summarize and draft. Humans decide.

What changed

MHCLG and the government’s Incubator for AI have made Extract available to local planning authorities across England. That gives councils a shared route for digitizing planning records that often sit in paper archives, scanned PDFs or mixed document stores.

Government officials said Extract followed trials across 20 local planning authorities, including Exeter and Hillingdon. They expect an average council to save about 255 hours of manual work that staff would have spent digesting documents into digital form. Across England, officials estimate planning officers spend 250,000 hours a year checking those records.

APD tackles a different layer of the workflow. Instead of digitizing old planning data, APD helps officers handle new applications. The prototype summarizes case information, checks local rules and drafts the basis for a report that an officer can review.

Google Cloud, Google DeepMind, Faculty and local planning authorities are working with government on APD. MHCLG funds the work through an 8.2 million-pound contract with Google Cloud, Google DeepMind and Faculty as delivery partners.

That split matters for architecture. Extract addresses the data foundation. APD addresses decision support. Councils need both if they want AI to reduce planning delays without pushing officers into a black-box workflow.

Provider comparison

Google Cloud brings several strengths to this project. The platform has Vertex AI for model access, deployment, evaluation and governance. It also gives the government access to Gemini models on Google Cloud, which can handle document-heavy workflows and support multimodal inputs.

For planning departments, multimodal capability has value. Officers may need to interpret maps, scanned forms, notes, policy documents and application text. A platform that can connect those materials into one workflow has an advantage over a narrow text-only setup.

Microsoft Azure would bring strong public-sector credentials in the UK, mature identity controls through Microsoft Entra ID and broad uptake across councils that already run Microsoft 365. Azure OpenAI Service gives public bodies a familiar route into generative AI, and Microsoft Fabric could support data estate modernization.

Amazon Web Services would bring breadth, cost controls and a long track record in government infrastructure. Bedrock offers model choice, and AWS gives teams mature options for data lakes, event processing and audit-heavy workloads.

Google’s edge in this case comes from the combination of DeepMind research, Gemini models and Google Cloud delivery. The APD project also gives Google a public proof point in a domain that demands policy interpretation, evidence handling and workflow controls.

Councils should still compare providers on five practical tests before they commit to a wider migration. They should assess data residency, model auditability, integration with case management systems, support for human review and predictable cost at peak demand.

Pricing and cost control

The public contract gives APD an 8.2 million-pound starting point, but councils should look past the contract value and model the unit cost of each application.

AI planning tools create cost in four places. Teams pay to ingest documents, store data, run model inference and keep officers inside review workflows. They also pay for integration work when existing planning systems lack clean APIs.

Provider price pages can mislead public-sector buyers because planning demand comes in bursts. A district may see steady householder applications for weeks, then face a spike after local policy changes or seasonal renovation cycles. Councils need budgets that account for inference spikes, document reprocessing and retention rules.

A good cost model starts with case volume. England receives about 350,000 planning applications a year, and householder applications make up much of that workload. Councils should estimate cost per application, cost per document page and cost per officer review. They should separate one-time digitization costs from recurring application support.

Google Cloud can reduce operational burden if MHCLG centralizes platform operations and lets councils consume shared services. That model can give smaller councils access to stronger security, monitoring and model governance than they could fund alone.

A local-by-local procurement model would raise duplication costs. Each council would need its own integration plan, support model and risk review. National rollout works better if the government provides common rails and lets councils adapt local policy rules inside those rails.

Migration considerations

Councils face a data migration problem before they face an AI problem. Planning data often spans scanned maps, document management systems, GIS platforms, committee reports and case management software. Officers understand local quirks that no model can infer from a clean schema.

The first migration task should map the records that officers check during a routine application. Councils should identify which records Extract can digitize, which records need manual cleanup and which local policy documents APD must use during assessment.

The second task should define review checkpoints. Officers need to see the source documents behind a model’s summary. They need version history when policy changes. They need escalation routes when the system gives weak evidence or misses local context.

The third task should address procurement and exit rights. Councils should avoid tying planning records, prompts, evaluation data and workflow state to one provider in a format they cannot export. A sound contract should give councils a clean path to move data, logs and evaluation results if prices rise or policy changes.

Security teams should test prompt injection, data leakage and role-based access. Planning files can include personal data, property details and commercially sensitive development plans. A model that handles those files must respect council access controls and record-retention duties.

Business impact

The government wants AI to help it build 1.5 million homes this Parliament. Planning reform will need policy change, more staff capacity and better data. AI can help on the administrative work that slows routine cases.

For councils, the main gain comes from officer time. If Extract saves an average council 255 hours and APD cuts routine decision time, planning teams can shift scarce expertise toward major developments, enforcement issues and complex local objections.

For residents, faster householder applications could reduce waiting time for loft conversions, extensions and home adaptations. For developers, better planning data could cut uncertainty during site assessment and reduce repeated document checks.

For cloud buyers, the program gives a useful pattern. The government did not start with a general AI assistant. It selected specific workflows, retained human signoff and tied the tools to measurable targets. That gives procurement teams a firmer basis for evaluating value.

Google Cloud now has to prove that its platform can handle national use without turning council planning into a fragmented AI estate. The next test will come when MHCLG expands APD beyond the first three councils and exposes the tool to more policy variation, more document types and more edge cases.

Public-sector AI projects succeed when teams treat models as one part of a service. Councils will need trained officers, clean data, audit trails, cost controls and clear accountability. Google Cloud can provide the infrastructure, but government leaders must run the operating model.