PolicyLaw, regulation & governance
Google DeepMind's PlanAssist planning tool expands from 3 to 15 English councils
PlanAssist is a Gemini-based tool from Google DeepMind. It drafts householder planning assessments that an officer signs off. The tool is moving from three pilot councils to a 12-month beta across 15 English councils, and national availability is planned for 2027.
From the Google DeepMind article of October 30, 2026 and cited sources: In early pilots at Barnet, Camden and Dorset, DeepMind says officers spent 58% less time on householder assessment tasks. [1] [2]
SOURCES[1] Google DeepMind: Speeding up planning decisions with AI deepmind.google[2] Techgenyz: PlanAssist pilot findings techgenyz.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- In early pilots at Barnet, Camden and Dorset, DeepMind says officers spent 58% less time on householder assessment tasks. [1] [2]
- Original source
- Google DeepMind article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 30, 2026 article published
- Covered
- Nov 2 edition
- Evidence
- 9 sources across 9 sites
- 1Google DeepMind: Speeding up planning decisions with AIdeepmind.google
- 2Techgenyz: PlanAssist pilot findingstechgenyz.com
- 3Google blog: UK government harnesses Geminiblog.google
- 4GOV.UK: AI tool to slash planning decision timesgov.uk
- 5PublicTechnology: MHCLG hopes to halve householder timespublictechnology.net
- 6Fish4News: Planning rules change as AI trials expandfish4news.co.uk
Reach: not measured yet
In early pilots at Barnet, Camden and Dorset, DeepMind says officers spent 58% less time on householder assessment tasks. It also says officers approved more than 70% of the drafted reports with no edits. These are vendor-reported figures. No sample sizes, error types or baselines have been published. [1] [2]
Twelve more councils are joining a 12-month beta on live applications, and their usual processes stay in place. The work is part of an £8.2m government contract with Google Cloud, Google DeepMind and Faculty. The government's target is to cut householder decisions from about eight weeks to four. [1] [4] [5] [7] [8]
The system checks documents, heritage and site constraints, and compliance with planning policy. It also brings up earlier objections and drafts a report. The officer approves the report and keeps the final decision. PlanAssist builds on Extract, an earlier Gemini tool that digitises old planning records and maps. [1] [3] [4]
The expansion comes as new rules send about 70% of applications to officers instead of elected committees. The Local Government Association warns that this weakens democratic oversight. The Royal Town Planning Institute has raised two risks: planners could lose skills over time, and consultation views could be averaged out. [6] [7] [9]
So far the evidence covers faster officer task times in three councils. The eight-to-four-week cut is still a government target.
Read the full assessment
If the beta results hold, an AI-drafted report with human sign-off could become a model for regulated public-sector casework.
Executive brief
Google DeepMind says PlanAssist, a Gemini-based assistant for English planning officers, cut the average time officers spent on householder assessment tasks by 58% in early pilots at Barnet, Camden and Dorset. It also says more than 70% of the reports it drafted were approved with no edits (Google DeepMind). The government is now widening testing from 3 councils to 15 and plans national availability in 2027. No independent evaluation has been published with these figures (Techgenyz). The expansion also coincides with new rules that move most routine decisions from elected committees to officers.
What changed and event timeline
Extract unveiled
Government and Google launch Extract, a Gemini tool that digitises old planning records, including maps and handwritten notes. They say it processes a document in about 40 seconds versus roughly 2 hours by hand ().
APD alpha begins
Augmented Planning Decisions (APD), the prototype later renamed PlanAssist, starts alpha trials in Barnet, Camden and Dorset ().
Contract and target announced
The ministry for housing (MHCLG) confirms an £8.2m APD contract with Google Cloud, Google DeepMind and Faculty. The stated aim is to cut householder decisions from about 8 weeks to 4 (;).
LGA objects to delegation rules
The Local Government Association (LGA) warns that the national scheme of delegation could weaken planning committees' democratic role. It asks for the government's review to come forward to October 2027 ().
30–31 Oct 2026: PlanAssist results and expansion
DeepMind publishes the 58% and 70% pilot figures. Twelve more councils join a 12-month beta on live applications, with their usual processes kept in place (;).
Capabilities and access
- No Gemini version is named (Google DeepMind).
- Functions: validates documents, checks heritage and site constraints, reviews compliance with planning policy, surfaces earlier objections, and drafts an assessment report for an officer to approve.
- Access: limited to 15 English councils. The new ones include Ashford, Bristol, Lambeth, Newcastle, Nottingham, Westminster and others.
Read the full section
- Model: "built with Gemini." No Gemini version is named (Google DeepMind).
- Functions: validates documents, checks heritage and site constraints, reviews compliance with planning policy, surfaces earlier objections, and drafts an assessment report for an officer to approve.
- Access: limited to 15 English councils. The new ones include Ashford, Bristol, Lambeth, Newcastle, Nottingham, Westminster and others. A national rollout is planned for 2027, subject to testing (Superpower Daily).
- Scope: householder applications only.
Technical analysis for researchers and developers
- Design pattern: retrieval and synthesis across application files, local policy and constraint layers.
- Related pipeline: Extract (Gemini on Vertex AI) pulls polygons out of maps and georeferences documents, using OpenCV and Segment Anything (Google blog).
- Evaluation: the 58% figure measures officer task time, not end-to-end decision time.
Read the full section
- Design pattern: retrieval and synthesis across application files, local policy and constraint layers. The output is a draft report that an officer signs off (GOV.UK).
- Related pipeline: Extract (Gemini on Vertex AI) pulls polygons out of maps and georeferences documents, using OpenCV and Segment Anything (Google blog). Digitised records like these feed the kind of constraint checks PlanAssist performs.
- Evaluation: the 58% figure measures officer task time, not end-to-end decision time. The "no edits" rate shows whether officers accepted a report, not whether the report was correct. Sample sizes, error types and baselines are not published, so the results cannot be reproduced.
Claims and evidence
- 58% less assessment time. More than 70% of reports unedited: reported by DeepMind from 3 pilot councils, with no published methodology (Google DeepMind).
- "8 weeks to 4 weeks": the government framed this as a target (GOV.UK). Some coverage presents it as already achieved (Newshub).
- Extract speed: about 40 seconds per document, as reported by Google (Google blog).
Read the full section
- 58% less assessment time; more than 70% of reports unedited: reported by DeepMind from 3 pilot councils, with no published methodology (Google DeepMind). These figures have not been independently verified (Techgenyz).
- "8 weeks to 4 weeks": the government framed this as a target (GOV.UK). Some coverage presents it as already achieved (Newshub). Any speed-up would also reflect the broader reforms, not the AI alone.
- Extract speed: about 40 seconds per document, as reported by Google (Google blog).
Context and prior work
- PlanAssist builds on Extract, which is now available to all English councils (GOV.UK).
- It runs alongside a planning-staff push: 420 new planners have been hired against a manifesto target of 300 (Newshub).
- From 31 October, the national scheme of delegation routes about 70% of applications, mostly householder cases, to officers rather than committees (Newshub).
Read the full section
- PlanAssist builds on Extract, which is now available to all English councils (GOV.UK).
- It runs alongside a planning-staff push: 420 new planners have been hired against a manifesto target of 300 (Newshub).
- From 31 October, the national scheme of delegation routes about 70% of applications, mostly householder cases, to officers rather than committees (Newshub).
- The Royal Town Planning Institute (RTPI) has flagged two risks: AI could erode planners' skills over time, and consultation views could be "averaged out" (Specification Online).
Limitations, safety and contested findings
- Officers keep final decision authority on every application (GOV.UK).
- A high "approved with no edits" rate could partly reflect automation bias, where reviewers accept machine output too readily.
- The LGA disputes the governance changes that will route more PlanAssist-assisted cases to officers.
Read the full section
- Officers keep final decision authority on every application (GOV.UK).
- A high "approved with no edits" rate could partly reflect automation bias, where reviewers accept machine output too readily. No published data separates that effect from genuine report quality.
- The LGA disputes the governance changes that will route more PlanAssist-assisted cases to officers. This is a disagreement about the system's oversight, not its accuracy (Fish4News).
- The pilots cover only three councils so far. The beta adds rural and metropolitan councils to test whether results generalise (Google DeepMind).
Business and practitioner implications
- Procurement model: a frontier lab plus a delivery partner (Faculty) under a modest £8.2m contract, scaling one council at a time.
- Builders: "drafted report plus human sign-off" is the deployable pattern in regulated decisions.
- Developers and homeowners: faster householder decisions are plausible but unproven. The 12-month beta results will matter more than the pilot figures.
Read the full section
- Procurement model: a frontier lab plus a delivery partner (Faculty) under a modest £8.2m contract, scaling one council at a time. This is a template for AI in public-sector casework.
- Builders: "drafted report plus human sign-off" is the deployable pattern in regulated decisions. Metrics need to go beyond acceptance rate, for example error audits and appeal outcomes.
- Developers and homeowners: faster householder decisions are plausible but unproven. The 12-month beta results will matter more than the pilot figures.