I’ve been watching the UK’s housing crisis for years, and the bottlenecks in planning approval have always struck me as the kind of problem AI should theoretically solve. Now, with PlanAssist, we’re seeing that theory become measurable reality.
Google DeepMind just announced that their planning assistance tool, built on Gemini, is expanding from initial pilots across Barnet, Camden, and Dorset to 12 additional councils over the next year. But the headline number isn’t really the expansion itself. It’s that 58% reduction in assessment time per application, combined with the fact that over 70% of AI-generated reports needed zero edits from planning officers.
That’s not incremental improvement. That’s transformational.
The Bureaucratic Problem AI Can Actually Fix
Here’s what makes PlanAssist interesting to me as someone who thinks about how AI gets deployed: it solves a genuinely boring problem that has massive downstream consequences. Routine householder applications (loft conversions, extensions, conservatories) make up to 70% of planning officers’ workloads. These are mostly straightforward approvals. They don’t need expert judgment in the way complex commercial developments do. They just need someone to check that the application is complete, that it complies with local and national policy, that there are no heritage constraints violated, and that no objections have been raised.
That’s exactly the kind of work that AI excels at: pattern matching against policy documents, validating documentation completeness, cross-referencing constraints. It’s mechanical, repetitive, and doesn’t require the nuanced judgment we’re supposed to reserve for human experts.
What’s happened historically is that planning officers have been drowning in this routine work, leaving them too depleted to properly handle the genuinely complex cases that actually benefit from human expertise and judgment. The backlog then feeds itself: applications pile up, costs rise, projects get delayed.
PlanAssist breaks that cycle by automating the boring parts while keeping humans in the loop for final decisions. Every application still gets reviewed and approved by a qualified officer. The AI isn’t replacing judgment; it’s handling the paperwork so judgment can actually happen.
What This Reveals About Enterprise AI
I think there’s a lesson here for how we should be thinking about enterprise AI deployments more broadly. There’s this tendency to chase the most impressive-sounding applications: replacing lawyers, replacing radiologists, replacing entire departments. But PlanAssist’s success comes from being humble about what the tool does.
It validates documents. It checks policy compliance. It drafts reports with proper citations. Those are all genuinely useful things that would be tedious to do manually at scale, but they’re also things where you absolutely need human oversight before any actual decision gets made.
This isn’t a tool that’s trying to make planning officers obsolete. It’s a tool that respects the complexity of their actual job while eliminating the busywork that prevented them from doing that job well. The 58% time savings means that same officer can now handle more applications or spend meaningful time on applications that genuinely need careful consideration.
For developers thinking about AI in their own organizations, that distinction matters. The most defensible AI deployments aren’t the ones trying to replace human decision-makers. They’re the ones that free humans from drudgery so their expertise can actually be applied where it’s needed. More on this in my earlier analysis of practical AI implementation patterns.
The Scale Question
The UK government is planning to roll this out to all councils in England by 2027, which would be remarkable if it works at scale. Right now, we’re still in the pilot phase with 15 councils total. The real test is whether the tool generalizes across different local contexts, different policy frameworks, and different types of applications.
Rural districts will have different planning constraints than metropolitan areas. Some councils will have much larger backlogs. Some will have been managing their queues more effectively. The diversity of the pilot councils is smart in that regard; it’s testing the generalization problem.
There’s also a crucial question about how the tool adapts over time. Policy changes. Court precedents shift planning guidance. Heritage designations get updated. A tool that was trained on historical planning decisions needs to stay current with how planning actually evolves. That’s the kind of maintenance challenge I’ve written about before.
The Real Impact
If this works at scale, the impact extends far beyond faster application processing. The UK’s housing shortage is a genuine economic and social crisis. Even a 58% improvement in processing speed for routine applications could meaningfully affect whether the government hits its 1.5 million new homes target by 2029. For individual families, it means your loft conversion approval might come in weeks instead of months.
That’s the kind of concrete benefit that makes AI policy work worth paying attention to. It’s not transformative in the way people sometimes claim AI will be, but it’s genuinely useful in ways that matter to real people.
The question that lingers for me is whether we’re getting good at automating the genuinely boring parts of government, or just automating the parts we found easiest to productize.