If you have ever waited months for a council to approve a loft conversion or an extension, the bottleneck PlanAssist targets will be familiar. Google DeepMind's Gemini-based tool for checking planning applications is expanding to 12 more local authorities in England for a 12-month beta on live cases, bringing the total to 15. It matters because routine householder applications, the simple ones, account for up to 70% of a planning officer's workload yet take a disproportionate share of their time.
The tool handles the mechanical checks and leaves the decision to a person. It validates that an application's documents and drawings are complete, checks the proposal against heritage constraints such as tree preservation orders and against local and national policy, and flags whether residents have raised objections before. It then drafts an assessment recommendation. The design keeps a qualified officer with the final say on every application, which is the right safeguard for decisions that affect neighbours and streets.
The pilot numbers are the draw. In early testing across Barnet, Camden and Dorset, DeepMind says PlanAssist cut the average time an officer spent on householder assessment tasks by 58%, and that officers approved more than 70% of the reports it generated with no edits at all. Those are the kind of figures that would let a stretched planning department clear a backlog, and the context is a government target of 1.5 million new homes by 2029, where slow approvals carry a real economic cost. The tool was built with the UK government, Google Cloud and Faculty.
For the people waiting, the appeal is obvious: a conservatory or extension approved in weeks rather than months. Routine householder work is a sensible first target precisely because it is high-volume and low-controversy, the kind of task where speeding up checks does the most good with the least risk.
The honest caveat is that the headline figures come from the pilot partners themselves, not an independent evaluation, so the 58% and 70% should be read as encouraging early results rather than settled fact. The 12-month beta on live applications is the real test, because performance on a controlled pilot and performance on the messy stream of real cases are not the same thing. A national rollout is floated for 2027, but it should depend on how the beta holds up, and on the officers who review every output continuing to catch what the model misses.