Business Context: The Cost of Planning Uncertainty
UK property developers make major capital commitments before they know whether planning permission will be granted. A site acquisition, architect fees, and pre-application engagement can represent six-figure expenditure before a planning application is submitted, let alone decided. The ability to assess planning risk more accurately and faster has direct financial value. Planning consultants offer this expertise, but good planning consultants are expensive and have limited availability. The information underpinning their assessments, local development plans, national planning policy, and prior appeal decisions published by the Planning Inspectorate, is publicly available and structured enough for AI to process. An AI planning analysis tool does not replace a planning consultant for complex or contested applications. It does provide a rapid first assessment that helps developers and their architects decide which sites to pursue, which proposal types to test, and what policy arguments to lead with. For smaller developers and proptech platforms that want to offer planning risk data at scale, this is the tool that makes it possible.
Architecture: Policy Retrieval, Precedent Search, and Assessment Generation
The system is built in three layers. The policy data layer ingests local development plan documents (PDFs published by local planning authorities) and national policy documents including the National Planning Policy Framework. These are chunked, embedded, and stored in a vector database (Pinecone or Supabase pgvector) indexed by local authority, policy area (housing, commercial, heritage, green belt), and document date. The precedent search layer queries the Planning Portal API for comparable application decisions filtered by local authority, application type, and proximity to the subject site. Decision notices are retrieved, summarised by Claude, and indexed for retrieval alongside the policy data. The assessment layer accepts a description of the proposed development (site address, use class, scale, key design features) and generates a structured assessment: applicable policies, policy compliance summary, comparable decisions with outcomes, and a qualitative approval likelihood narrative. The output is formatted for export as a PDF report or structured JSON for integration into proptech platforms.
AI Components: Document Retrieval and Policy Interpretation
Claude's large context window makes it well suited to planning policy analysis. UK local development plans can run to hundreds of pages, and relevant policies are scattered across multiple chapters. The RAG pipeline retrieves the most relevant policy chunks for the proposed development type and presents them to Claude alongside the development description. Claude identifies which policies are directly applicable, how the proposal aligns with or conflicts with each policy, and what weight each policy should carry in the decision-making context. The system prompt is structured around the five-step planning officer assessment framework used in UK planning decision reports: principle of development, design and character, impact on amenity, transport and highways, and other material considerations. This mirrors the structure of actual planning reports, which makes the AI output more immediately useful to planning consultants and less likely to miss material policy considerations.
Challenges: Policy Currency and Local Variation
UK planning policy changes frequently. Local plans are reviewed and updated, supplementary planning documents are added, and national policy evolves. The system must track document currency and alert users when an analysis is based on a local plan that has since been superseded or substantially modified. SpeedMVPs builds a document versioning layer that records the date each policy document was ingested and surfaces a currency warning when analysis relies on documents older than 12 months. Local planning authorities vary enormously in their interpretation of national policy, their appetite for development, and their precedent decision patterns. An analysis calibrated on one local authority may not apply to another even for similar proposal types. The system presents precedent decisions with explicit local authority attribution so users can distinguish between precedents from the same authority (highest relevance) and comparable decisions from other authorities (lower relevance). Planning decisions always involve discretion. The AI assessment is explicitly positioned as an evidence-based input to professional judgment, not a definitive prediction. This is the correct framing both practically and for managing user expectations.
Outcomes: What Proptech Teams and Consultants Report
Planning consultancies that pilot AI planning analysis tools consistently report the same primary benefit: speed of initial site assessment. A standard first-cut planning assessment that previously took a consultant 3-4 hours to produce (reading the local plan, searching for comparables, drafting the assessment) compresses to 30-45 minutes of reviewing and annotating an AI-generated report. This allows consultancies to assess more sites, produce higher-quality initial assessments for lower-value commissions, and focus senior consultant time on applications where judgment and relationship capital matter. For proptech platforms offering planning data as a feature, the AI assessment layer transforms static planning data into a structured analysis product that can be offered at scale across the UK property market without a consulting team behind each query.
Lessons: Calibrate Against Real Decisions, Not Theory
The most important calibration step for an AI planning analysis tool is backtesting against real decisions. Take a sample of 50 past planning applications from the target local authorities, run the AI assessment for each one, and compare the AI's approval likelihood assessment against the actual decision. This identifies where the model is systematically miscalibrated: perhaps it underweights design objections from conservation area policies, or overweights transport policy compliance because it cannot read the local planning officer's culture around highways concerns. Backtesting before launch is the difference between a tool consultants trust and one they dismiss after one bad prediction. The second lesson is to be explicit about what the tool does not do. It does not predict political decisions by elected planning committees. It does not account for a developer's relationship with the local authority. It does not replace pre-application engagement. These limitations should be stated clearly in the product.