PlanAssist cut UK planning work by 58%, while humans keep the final decision

Google DeepMind and the UK government are expanding PlanAssist from three pilot authorities to 15 English councils. Initial tests cut average officer assessment time by 58%, and more than 70% of generated reports were accepted without a single edit.

Gemini assembles the evidence while an officer signs the conclusion

PlanAssist checks documents and architectural drawings, finds local and national policy and tests constraints such as heritage protections and tree preservation orders. It also considers earlier objections and drafts a cited report that a qualified planning officer can edit before making the final decision.

The pilot has run since June in Barnet, Camden and Dorset. DeepMind says average assessment time fell by 58%, while officers accepted more than 70% of generated reports unchanged. The UK government links that result to a goal of cutting routine householder application processing from 8 weeks to 4.

Another 12 councils will join in three cohorts. All 15 authorities will test the tool on live applications for 12 months. Every assessment must still be reviewed and approved by a qualified officer.

The largest gain sits inside the dull work of government

Householder applications account for nearly 70% of planning work. They include extensions, conservatories and loft conversions. These repetitive cases are where a system can save time by finding policies, checking completeness and assembling evidence.

The important design choice is that PlanAssist does not take legal responsibility away from the officer. It moves work from collecting documents to checking an argument. That is a more practical use of AI than an automated approval stamp, but it raises the standard for audit: officers must spot cases where citations are correct and the conclusion is still wrong.

The pilot measured task time, not the whole statutory process

The 58% figure covers selected assessment tasks, not automatically the full period from submission to decision. Statutory consultation, site visits, neighbour notices and committee schedules still affect the timeline. The government’s claim of a reduction from 8 weeks to 4 therefore depends on broader process reform, not the model alone.

The figure of more than 70% of reports needing no edits also says little about the severity of errors in the remainder. In routine administration, one missed heritage constraint can matter more than dozens of correctly copied details.

Live applications will test whether 15 councils keep the same accuracy

The 12-month beta spans metropolitan and rural areas. Error rates by application type, the number of substantive report edits and officer review time will matter most. Without those measures, a faster first draft could merely move work into later correction.

The government intends to make PlanAssist available to every English council by the end of 2027 if testing succeeds. Use will not be mandatory. Publishing planning documents as open data will matter just as much, because no Gemini model can repair a bad source record.

Lilith's verdict

PlanAssist earns its keep when an officer checks a conclusion instead of hunting for attachments. If the 58% saving disappears during corrections, the faster stamp remains a polished demo.