The UK government wants 1.5 million new homes built by 2029. The math is not working out. Official data shows that as of late 2025, only about 18% of that target had been reached while 27% of the parliamentary term had already passed. At the current rate, England needs to add around 822 new homes every day to get there.
Planning is one of the clearest bottlenecks. At the worst-performing councils, applications take over 400 days to decide, more than four times the legal limit. Even by the government's own statistics, only about one in five major applications gets decided within the statutory 13-week deadline. And the picture on smaller applications, such as home extensions and loft conversions, is only marginally better.
The deeper problem is people. A survey of England's district councils found 84% of planning departments were struggling to recruit and retain staff. About one in nine planning posts nationally is currently unfilled, and 70% of officers say understaffing is a primary barrier to getting development approved. The government's pledge of 300 additional planners when it took office last year sounds meaningful until you see that the same research says at least 600 more are needed on top of that number.
This is the context in which Google DeepMind's planning tool matters. The tool, built using Google's Gemini AI, is being tested at three councils: Barnet, Camden, and Dorset. It handles what officers call the heavy lifting: reading through stacks of policy documents and historical files, flagging the relevant rules, summarising public consultation responses, and producing a first draft of the final assessment report, complete with citations the officer can verify.
The goal is to cut decision times on standard applications by 50%. Householder applications, things like extensions and loft conversions, make up close to 70% of all planning applications every year. If a 50% time reduction holds, the same officer handling two cases a week could handle four. That is the closest thing to a structural answer the housing backlog has seen.
Critically, the tool does not decide anything. The officer reviews every line, edits the reasoning, and retains full authority to approve or reject. The system also keeps a step-by-step record of how it worked through each case, so there is a clear paper trail if a decision is ever challenged. That accountability design matters: planning decisions affect property values, neighbourhoods, and legal rights, and councils cannot outsource responsibility to software.
If the three-council trial goes well, the government plans to make the tool available to all councils nationally from 2027. That is the real test. Councils vary enormously in their local policies, data quality, and the age of their document archives. An AI that works smoothly in Camden may struggle with a rural council whose planning records date back decades and exist only as scanned paper.
The broader picture is that AI is increasingly being applied to the administrative layer of government, the part that is slow not because decisions are hard but because the paperwork is overwhelming. Planning is a particularly clear example of this. The decisions themselves often require professional judgment. The preparation for those decisions, reading documents, cross-referencing policies, summarising objections, is exactly the kind of structured, repetitive task that AI handles well.
For anyone operating a business that depends on planning approvals, or simply waiting to build an extension, a faster system is plainly good news. Whether this prototype survives contact with the full complexity of England's planning system is the question that 2027 will answer.