UK Planning AI Cuts Officer Time 58%, but Reports No Figure for Decision Weeks── The 58% measures officers' working time, so applicants' waiting weeks and decision quality have to be checked with other numbers.

On October 31, 2026, Google DeepMind said PlanAssist, a planning AI built with Gemini, had cut officers' assessment time by 58% in pilots at three UK councils. Does that number show that decisions became faster and sounder? No. It describes officers' working time; the report gives neither the weeks applicants wait nor the quality of decisions, and those need separate numbers.

01PlanAssist's 58% counts working time and does not yet show review outcomes

Anyone in England who wants to extend a house needs planning permission from the local council. PlanAssist drafts the officer's assessment report for that application. DeepMind developed it with the UK government.

Two numbers are firm in the October post: officer time on these assessments fell by 58%, and more than 70% of the drafted reports were approved without edits. Both come from DeepMind itself, and I could not find an independent evaluation.

In June, DeepMind had set a different goal, a 50% cut in decision times. The October post does not tell us whether that goal was met.

The reduction in work deserves to be taken at face value. What someone deciding on AI adoption lacks is the next set of figures. When a vendor or a sponsor says "time fell by X percent," ask separately for weeks to decision and for a measure of quality.

02The October 31 post: two figures, 58% and 70%, and no sample size

Having established that the 58% concerns working time, the next step is to see where it was counted. The post says the pilot has run since June at three councils: Barnet, Camden and Dorset.

The tool works in four stages. It validates documents and drawings, checks the application against heritage constraints and against local and national policy, looks up issues raised earlier by residents and others, and then drafts a fully cited report that an officer can edit.

The post phrases the headline number in two ways. Its opening says "up to 58%," while the body gives the average time falling "by 58%." I keep both rather than rounding one into the other.

Absence from a blog post does not prove absence from DeepMind's records. The company may hold data on weeks; outsiders cannot tell. All I can say is that the post a reader can consult today does not contain them.

03The tool covers extensions and loft conversions, and grows from three councils to 15

The 58% has no stated sample, and it also has a narrow scope. DeepMind counted householder applications, the category that covers home extensions and loft conversions. It says nothing about review of large housing developments.

The weight of this category depends on which document you open.

These numbers do not contradict one another, because they count different things at different times. A writer who cites one without its denominator will misstate the size of what was tested.

The scope is about to widen. DeepMind says twelve more councils will join in three cohorts, for 15 in total. PlanAssist then enters a 12-month Beta using live applications, beginning with a subset of those councils.

04Applicants' waits include a 21-day consultation that officer work cannot shorten

Even inside the householder category, officer work is only part of what an applicant experiences. In June the government said it aimed to halve the time to process these applications, from 8 weeks to 4 in an average case. Those 8 weeks contain time no officer controls.

Consultation is the clearest example. Government guidance says the formal consultation will normally last 21 days, with 14 days where notice appears in a newspaper. The guidance does not single out householder applications, so I would not apply the figure uniformly. Still, the wait does include days set aside for neighbours and others to respond.

If officers finish each case sooner, queues may shrink, and I do not rule that out. But the sources I read do not say how much of the 8 weeks is spent in a queue.

So weeks cannot be calculated from working time. No published formula links the two, and the government's target of 4 weeks does not follow from DeepMind's October figure.

05Speed and quality can both be counted with existing public measures

Since working time cannot reveal the wait, the question is who already counts it. The Ministry of Housing, Communities and Local Government publishes the numbers each quarter, so no new yardstick has to be invented.

In April to June 2026, 58% of "other" applications were decided within the statutory 8 weeks, and 90% within 8 weeks or an agreed time. That 58% has nothing to do with PlanAssist's 58%; the match is a coincidence. The "other" category also holds more than householder cases.

The distance between those two shares matters too. A decision past 8 weeks still counts toward the 90% if the applicant agreed to the later date. Anyone reading a performance claim should check which of the two shares is being shown.

An appeal is a challenge by an applicant to the council's decision. The 10% in the last row is one of the thresholds government uses to judge an authority's record. I could not confirm whether Beta results will be reported against it.

06A shifted unit, the reading of 70% unedited, and the fate of saved time change the verdict

With no weeks and no appeal figures in the post, readers are left with the 58% and the 70%. Three facts decide how much those numbers can carry.

The unit changed between June and October. In June, DeepMind wrote of cutting "decision times" by 50%. In October, it reported a 58% cut in time spent on "assessment tasks." The second number is larger, but it counts a different clock, so it should not be read as the goal met.

The share of reports passed without edits can be read two ways. The drafts may have been right, or the review may have been thin, and the post offers nothing to tell these apart.

Goddard and colleagues, in a 2011 systematic review, found that the tendency to over-rely on automation is shaped by workload and time pressure. That review centers on health care and did not measure planning officers, and I am not accusing anyone of cutting corners. I simply do not count the 70% as evidence of accuracy. Showing officers drafts with planted errors would test the question directly.

Saved time also has to go somewhere. The UK Department for Business and Trade evaluated a pilot of Microsoft 365 Copilot, an AI assistant for office documents, and did not find evidence that time savings had led to improved productivity. That is a different tool, productivity was not the main aim, and the report itself says its data are limited. It does show that less time and more output are separate claims.

07If the 12-month Beta publishes decision weeks, the 2027 choice has its evidence

The changed unit and the ambiguous 70% will not settle themselves, so attention moves to the 12-month Beta on live applications. In June the government said that, if successful, the technology would be rolled out nationwide by 2027.

Should the Beta produce weeks and appeal figures, government and councils would have what they need to decide. That is my inference. Within what I could check, I found neither a commitment to publish such numbers nor a stated definition of success. I also cannot say how the end of the Beta will line up with the rollout decision.

Without those figures, the announcement stays a report about working time. Lighter work for officers would still have value, though it would not establish that applicants wait less.

The sequence in the figure is not limited to planning. When a sales deck for an AI product promises that a task will take half the time, I ask the same questions in the same order.

A 58% cut in officers' assessment time does not yet show faster or sounder decisions. Until decision weeks and appeal outcomes appear, treat it as a report about working time.