England put PlanAssist in 15 councils. Preserve the officer's decision trail

THE SIGNAL IN ONE SENTENCE

A loft conversion is not the moon landing. It can still disappear into a planning system for weeks while an officer checks drawings, policies, maps, heritage constraints, public comments and whether the applicant remembered the document that everybody assumed somebody else had uploaded. England is giving more councils an AI assistant to help with that pile. On October 31, the Ministry of Housing, Communities and Local Government said PlanAssist would expand from pilots in Barnet, Camden and Dorset to 12 additional local planning authorities. That brings the total to 15 councils for a 12-month beta using live applications. PlanAssist was developed with Google DeepMind, Google Cloud, Faculty and local authorities. It brings application material, planning policy and site constraints into one workspace. The system checks whether required documents and drawings are present, looks for relevant constraints such as tree preservation orders, reviews local and national policy, surfaces objections and drafts a cited assessment for a qualified planning officer. The officer still makes the decision. That sentence is the hinge. It is also too small to carry the entire accountability system by itself. Government and Google DeepMind say early testing cut the average time officers spent on key householder assessment tasks by 58 percent. DeepMind also says more than 70 percent of generated reports in the initial pilots were approved without edits. Routine householder applications can make up as much as 70 percent of planning workloads, so even a modest reduction in repetitive checking could release meaningful time. Those are promising early figures. They are not yet a public evaluation. The announcements do not provide the underlying sample, task mix, comparison method, error categories or performance by council. A report accepted without edits may be excellent. It may also concern an easy application with clean data. A report that needs one edit could contain a trivial punctuation change or a serious policy correction. The headline percentage does not tell us which. PlanAssist is moving from a limited pilot into a larger and more varied test. That is the right moment to preserve a decision trail for every application. The trail should begin before the model does anything. Record the exact application package received by the council: drawings, forms, ownership certificates, design statements, photographs, maps and revisions. Give each file a version and timestamp. If an applicant replaces a drawing halfway through review, the record must show which version the assistant examined and which version the officer approved. Next, keep the retrieved evidence. PlanAssist can bring together national policy, local plans, conservation areas, listed buildings, tree preservation orders and other site constraints. Each surfaced rule should carry its source, publication date, geographic coverage and the time it was retrieved. A policy quotation without a version is a trap wearing a sensible tie. Planning rules change. Boundaries move. Local documents are revised. Data can be incomplete or stored in awkward formats. If the system misses a recently adopted policy because an index is stale, the final report may look beautifully cited and still rest on yesterday's map. The decision trail should make missing evidence visible. It should say when a dataset was unavailable, when a scanned document could not be parsed, when a boundary produced an uncertain match and when a policy had to be checked manually. Then preserve the public voice. Planning objections are not votes, and an objection does not automatically defeat an application. Officers must distinguish material planning considerations from concerns that planning law does not allow them to use. That makes summarization useful and risky at the same time. An assistant can group repeated concerns, identify the policy topics they touch and save an officer from reading the same sentence in 47 slightly different forms. It can also flatten the one unusual objection that identifies a drainage problem, a disability-access issue or a factual error in the application. Keep every original submission linked to the summary. Record which comments were included, grouped or treated as non-material, and let the officer inspect the source before relying on the synthesis. Do not turn public participation into a sentiment score. The generated assessment needs its own receipt. Store the PlanAssist version, prompt or task configuration, data sources, retrieved passages, draft report and confidence or warning signals available at the time. Link every factual statement and policy conclusion to evidence the officer can open. If the system says a proposed extension is inside a conservation area, show the boundary record. If it says the design conflicts with a local policy, show the policy passage and the facts used to connect it to the application. If it says no objection raises a material issue, show the classification and the officer's review. This is not a demand to expose private reasoning or publish personal information. Councils already have duties around data protection, confidentiality and public records. The useful approach is a layered trail: a complete internal audit record for authorized review, plus a public decision record that explains the evidence and human judgment without exposing protected data. Human review should be more than a green button. Record what the officer accepted, edited, rejected or added. Separate cosmetic edits from material ones. A material edit changes a fact, policy interpretation, recommendation, condition or reason. Those changes are valuable evaluation data because they show where the assistant actually needed professional judgment. The final record should name the responsible role, not merely say that a human was involved. It should show when the review occurred, which evidence was checked and why the final decision followed or departed from the draft. That matters for applicants too. If an application is refused, the applicant should receive reasons tied to the adopted policy and the submitted facts. If the decision is appealed or corrected, the council should be able to reconstruct what PlanAssist produced, what the officer changed and what evidence was available at the time. Without that trail, accountability becomes an archaeological dig through screenshots and browser history. The beta should measure more than speed. Publish results by council, application type and stage of review. Track missing-document detection, retrieval accuracy, unsupported statements, material officer edits, time saved, total processing time, appeal outcomes and corrections. Compare applications processed with and without the tool where a fair design allows it. Look for uneven performance. Historic neighborhoods, rural sites, converted buildings and properties near administrative boundaries may produce different data problems. Councils organize documents differently. Some local plans are clean and structured. Others are a geological layer of PDFs. A national average can hide a tool that works wonderfully in one authority and creates extra checking in another. The government also says key planning documents will be digitized and published as open data by the end of 2027. That work may matter as much as the model. An assistant cannot reliably retrieve a constraint that exists only as a fuzzy scan with no stable identifier. Better data helps PlanAssist, competing planning tools, applicants, researchers and the public. Open data should not become a quiet subsidy for one assistant. Publish common standards, update schedules, provenance and correction routes so different tools can use the same civic infrastructure. The government intends to make PlanAssist available to every English council by the end of 2027, subject to successful testing. It also says the software will not be mandatory. Good. The beta should earn expansion rather than treat expansion as proof. The practical promise is easy to understand. Let software assemble the paperwork so scarce planners can spend more time on judgment, complex development and conversations with communities. The practical risk is just as clear. A fast, polished draft can become the default answer before anyone notices that the map was stale, the objection was compressed or the policy citation did not support the conclusion. PlanAssist does not need to replace the officer to shape the decision. It shapes the file the officer sees, the order of the evidence and the first recommendation placed on the desk. Preserve that path. For every application, the public system should be able to answer five questions. What did the council receive? Which rules, maps and comments did the assistant use? What did it draft? What did the officer change and why? What happened after the decision? If England can answer those questions while cutting routine delay, PlanAssist could become a useful piece of public infrastructure. If it can only report that a human clicked approve, the planning system will have gained speed and lost the map of how it got there.