Google DeepMind's PlanAssist Cuts Housing Review Time by 58% Across England
Google DeepMind launches PlanAssist , a Gemini-powered tool for UK planning officers. Pilots in Barnet, Camden and Dorset cut householder assessment time by up to 58%. Over 70% of generated reports were approved by officers with no edits required.
Takeaways
- Google DeepMind launches PlanAssist , a Gemini-powered tool for UK planning officers.
- Pilots in Barnet, Camden and Dorset cut householder assessment time by up to 58%.
- Over 70% of generated reports were approved by officers with no edits required.
- Built with Google Cloud and Faculty; handles document checks, policy matching, objections, and cited drafts.
- 12 more councils joining a 12-month live beta, bringing total to 15.
- Government targets rollout to all English councils in 2027, human officers retain final say.
Google DeepMind is expanding PlanAssist , a Gemini-powered assistant for householder planning applications, after pilots in Barnet, Camden and Dorset reduced officers’ average assessment time by 58%. The pilot has run since June, and the tool is entering a 12-month beta that will process live applications.
PlanAssist targets the document checks and policy reviews behind applications for loft conversions, extensions and conservatories. It drafts cited assessments for qualified planning officers, who can edit the reports and retain responsibility for every decision.
Small applications, heavy workload
Householder applications involve relatively modest changes to existing homes, but their volume creates a substantial administrative burden. Routine cases can account for up to 70% of planning officers’ workloads, with each application requiring document validation, site checks, policy review and a written assessment.
Automating parts of that process could free officers to handle complex developments and larger housing schemes. That additional capacity may support the government’s target of building 1.5 million new homes in England by 2029 while reducing delays for homeowners awaiting decisions on smaller projects.
Four checks, one cited report
Google DeepMind developed PlanAssist with the UK government, Google Cloud and Faculty. The system runs a sequence of tasks that mirrors the checks performed by a planning officer:
- Validate the submission: confirm that application forms, supporting documents and architectural drawings contain the required information.
- Review site constraints: identify relevant heritage designations, tree preservation orders and other local restrictions.
- Apply planning policy: assess the proposal against local and national rules.
- Draft the assessment: summarize findings, incorporate recorded issues or objections and cite the supporting policies and documents.
Gemini generates drafts linked to specific policy clauses and application materials. Those citations let officers trace each conclusion to its source, correct errors and approve the final report within the existing decision process.
Pilot results: faster drafts, fewer edits
| Measure | Reported result | What it indicates |
|---|---|---|
| Pilot authorities | Barnet, Camden and Dorset | Testing covered two London boroughs and a largely rural council area. |
| Average officer time | 58% lower | Time spent on householder assessment tasks fell by more than half. |
| Reports requiring no edits | More than 70% | Most generated reports were accepted on the first officer review. |
The no-edit figure covers generated assessments such as checks for policy compliance and submission completeness. Because qualified officers reviewed every report, the metric reflects first-pass usefulness within a supervised workflow. Application approval rates are a separate measure.
Beta widens the policy test
The 12-month beta will add 12 local planning authorities in three cohorts, bringing the total number involved to 15. The phased rollout begins with a subset of councils and expands as testing progresses.
The participating authorities include metropolitan and rural districts, giving the system a broader test across communities with different local plans, site constraints and application patterns. Local variation is a central scaling challenge because a proposal that complies with one council’s rules may require a different assessment elsewhere.
Subject to successful testing, the government plans to make PlanAssist available to every council in England in 2027. A qualified planning officer will continue to review and approve each assessment before a decision is issued.
A workable pattern for regulated agents
PlanAssist forms part of DeepMind’s broader UK government partnership on public-service AI and follows an earlier planning announcement . Its implementation offers several practical choices for teams building domain-specific agents:
- Choose a bounded workflow with explicit rules and repeatable inputs.
- Ground generated conclusions in source documents and policy citations.
- Give accountable specialists control over edits and final decisions.
- Measure processing time and first-pass acceptance separately.
- Test across jurisdictions where policies and edge cases vary.
The live beta will test whether the pilot’s time savings and no-edit rate hold across higher volumes, varied local policies and a wider range of applications. Evidence from those 15 authorities will inform the planned national rollout in 2027.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
