Google DeepMind cuts renovation application review time by 58% with Gemini

If you have ever built a veranda or altered a roof in the UK, you know that you first need permission from the local council. Behind every application is a person: a planning officer. They read drawings, check them against local rules, examine neighboring plots and write an assessment. The work is repetitive, the paperwork extensive and the backlog growing. This is where Google DeepMind has put Gemini to work.

What DeepMind built

DeepMind announced PlanAssist, a Gemini-powered tool for UK planning officers. It helps assess householder applications: submissions from private homeowners for extensions, loft conversions, garages and alterations. These make up the system's largest and most routine stream of applications. Complex commercial projects and residential developments fall outside its scope.

The headline figure from the announcement: the tool reduces the time an officer spends assessing such an application by up to 58%. The post says "up to." That is the best result, not the overall average. The end of the post in the source material is truncated: it says the tool is becoming available to 12—apparently organizations—but the text does not tell us exactly which ones. Those details should be checked against DeepMind's full announcement.

Why this is more interesting than another "chatbot for public services"

Most government AI projects end with a demo, a polished presentation and silence. This one shows three signs of maturity.

First, it targets a specific routine task with clear inputs and outputs. It is not an "assistant for everything," but a tool for assessing a particular type of application.

Second, its impact is measured in staff working time. For a customer, whether a council or a bank, that is the most readily understood metric. It can be translated into money saved and a shorter backlog.

Third, the tool has moved beyond the pilot stage and is being rolled out more widely. For a public-sector AI product, that is already a significant step.

What this means in practice

Imagine a typical application: a homeowner wants to add a second story above a garage. The officer needs to find the dimensions in the submitted documents, check them against local limits on height and distance from the property boundary, check for any protected status and compile the findings into a report. Most of the time goes into searching and cross-checking, rather than making the decision itself. A model with a long context window and multimodal input—plans and drawings are documents too—is well suited to precisely this kind of paperwork.

For developers, the takeaway is simple: the LLM applications with the strongest returns are often boring. They are found wherever people do a lot of reading, check information against regulations and write standardized text.

For product managers and founders, there are three lessons:

• Choose a narrow, repetitive use case rather than "AI for an industry."
• Agree on a metric with the customer from the outset. "Minutes per application" is an easier sell than "answer accuracy."
• Plan the rollout: a pilot without a plan to scale will remain a pilot.

Limitations and questions

There are caveats. First, 58% is the upper limit, and without the measurement methodology, that figure cannot be applied to other processes. Second, the post does not explain how oversight works: who makes the final decision, how model errors are recorded or what happens in disputed cases. In a regulated environment, these issues matter more than speed. Third, the post comes from DeepMind itself, so there is no independent assessment yet.

Still, the direction is telling. Major labs are increasingly showcasing deployments with quantified savings rather than benchmarks. If you are building a B2B product for a document-heavy field—law, insurance, real estate or compliance—PlanAssist offers a ready-made example of what a pitch should look like: a specific routine task, a measurable gain and a clear path to scale.