DeepMind Expands PlanAssist to 12 More English Councils: Application Checks 58% Faster

Anyone who has ever waited for permission to build an extension or convert a loft knows how slowly this wheel turns. The problem is not complexity: standard applications are simple, but there are a great many of them. This is the queue Google DeepMind has decided to ease with Gemini.

What happened

DeepMind and the UK government are expanding PlanAssist, an AI tool for municipal planning authorities. Another 12 councils will join the three pilot councils (Barnet, Camden and Dorset), bringing the total to 15. Next comes a 12-month beta on live applications, and if it succeeds, the tool is meant to be opened to all English councils in 2027.

The scale of the problem is clear from the numbers: standard homeowner applications account for up to 70% of planners' workload. They are usually approved easily, but consume a disproportionate amount of time. This also undermines the government's goal of building 1.5 million homes by 2029.

How it works

PlanAssist does in four steps what a person used to do by hand:

1. It checks that documents and drawings are complete: whether everything is attached so the application can move forward at all.
2. It cross-checks the application against constraints such as tree and heritage protections, as well as local and national rules.
3. It looks at whether objections have been received previously, for example from neighbours.
4. It prepares a report with quotations and links to sources, which the officer edits and only then makes a decision.

Imagine an application for a rear extension. Previously, a staff member opened the drawings, searched for the relevant policy clauses, checked the constraints map and wrote a conclusion. Now they receive a ready draft and review it rather than assembling it from scratch.

Pilot results

Since June, across the three councils, officer time spent on assessment tasks fell by an average of 58%. More than 70% of generated reports (for example, on compliance with local rules or completeness of documents) were approved without edits. These figures are internal and come from the pilot, with no independent verification, so they should be treated with caution.

Limitations

It is still an early stage. The beta is only starting, and the sample covers three councils, so scaling to rural and urban areas will show how robust the result is. The post says nothing about the error rate among the 30% of reports that were edited, or how the risk of a missed violation is measured. An important detail: every decision is reviewed by a qualified officer, meaning the AI does not make decisions on its own.

Practical takeaway

For developers and product managers working with the public sector and regulated processes, this is a good architecture pattern. First, the AI handles preparation, not the decision. Second, every conclusion comes with citations, so a person can verify it quickly. Third, the success metric is measurable and clear: time per task and the share of reports without edits. If you are building an assistant for compliance, legal teams, or insurance or banking applications, take the same approach: document validation, rule cross-checking, a draft with references and mandatory human approval.

It is also worth watching how the project performs on the flow of live applications across 15 councils. If the numbers hold, this will be one of the most convincing public cases of AI in government services.