Why validation is a bottleneck in planning systems

Planning validation is a procedural step, but its effects are systemic. Before any application can be assessed on planning merits, it must first be checked for completeness and compliance with statutory requirements. This includes verifying that all mandatory forms are completed, the correct plans are submitted, fees are accurate, and location- and policy-specific requirements are met.

Research undertaken in the UK shows that this stage absorbs a disproportionate amount of professional time relative to its complexity. Analysis led by the Alan Turing Institute found that manual validation typically takes between 30 and 60 minutes per application. At a national scale, this equates to approximately 250,000 hours of local authority staff time each year spent on validation alone .

The same research highlights that invalid applications are not an edge case but a structural feature of the system. Over one-third of planning applications in England are initially submitted as invalid. Importantly, this is not due to complex planning judgement, but to relatively mundane errors: incomplete forms, missing plans, incorrect fees, or omissions such as north arrows on drawings. A collaborative pilot study led by Agile Datum demonstrated that around 80% of invalid applications could be traced back to just twelve recurring error types .

The operational consequence is a feedback loop of delay. Applicants often wait weeks for validation, only to be told that a minor omission requires resubmission. Officers must then re-check the amended application, further extending timelines and increasing workload. Validation, in effect, becomes a bottleneck that constrains the entire planning pipeline.

What the UK research shows about AI-assisted validation

UK research and pilots over the past five years have focused narrowly on whether this bottleneck can be reduced through targeted automation. The emphasis has not been on replacing planners, but on supporting routine checks that are rule-based and repetitive.

Early evidence comes from local authority pilots. In 2019, the London Borough of Redbridge worked with Agile Datum on one of the first AI-assisted validation trials. The council reported that up to 80% of validation tasks could be automated. Average validation turnaround fell from approximately three weeks to 24 hours, and officer time savings were estimated at around 60 minutes per application, equating to roughly £250,000 in annual staff cost savings .

More recent pilots have reinforced these findings. In 2024, Enfield Council trialled an AI-driven “Planning Insights” tool focused on completeness checks and information gathering. According to reported pilot results, the system achieved 100% accuracy in retrieving site-specific constraints such as Article 4 directions and planning history. Tasks that previously took several hours were reduced to 20–45 seconds, allowing officers to redirect attention to more complex elements of casework .

At a national level, the UK government has also invested in document processing. In 2025, trials of the “Extract” tool, developed with Google DeepMind, were undertaken across three councils, including Hillingdon and Exeter. Extract digitised and classified legacy planning documents in approximately three minutes per file, compared to one to two hours of manual effort. The government has stated an intention to roll this tool out across English planning authorities by 2026, with the stated aim of significantly reducing the time spent on validation checks .

Taken together, these initiatives show that AI is being used in the UK to address a specific and well-defined problem: identifying missing, inconsistent or incorrect information at the point of submission, and extracting structured data from unstructured documents.

How human oversight is preserved in UK pilots

A consistent feature across UK pilots is the explicit preservation of human oversight and accountability. None of the documented systems make autonomous validation decisions in isolation.

This is most clearly articulated in the 2025 pilot at Leeds City Council, which introduced an AI-assisted workspace known as “Xylo Core”. The system uses large language models to analyse application documents, flag potential issues, suggest relevant policy references, and draft correspondence. However, every AI-generated output must be reviewed by a planning officer. Officers retain full authority to accept, modify or reject the AI’s suggestions, and no decision is made without human sign-off. An audit log records both AI outputs and human decisions, ensuring traceability and transparency .

This approach aligns with national governance guidance. In 2024, the Planning Inspectorate issued guidance stating that if AI is used to draft or alter content in planning submissions or appeals, this must be disclosed. Crucially, responsibility for factual accuracy remains with the human submitter, who must affirm that the content is lawful and accurate. AI assistance, under this framework, does not dilute accountability; it reinforces it through explicit attribution .

Professional commentary reinforces this stance. Planning bodies and practitioners consistently stress that planning is a discretionary system requiring contextual judgement, ethical consideration and democratic accountability. AI tools are therefore framed as decision-support mechanisms rather than decision-makers, with human planners retaining responsibility for interpretation and outcomes .

What has not been automated and why

Equally instructive is what UK pilots have deliberately chosen not to automate. None of the documented systems attempt to replace professional judgement on matters such as design quality, policy balance, community impact or material considerations. These aspects of planning involve subjective evaluation and value-based trade-offs that cannot be reduced to checklist logic.

Even within validation, automation is scoped carefully. AI systems focus on detecting absence, inconsistency or mismatch, such as whether a required plan is missing or whether the written description aligns with submitted drawings. They do not determine whether an application should be deemed acceptable in planning terms.

The research also highlights evidence gaps. Many pilots remain time-limited or geographically specific, and long-term performance data has not yet been widely published in peer-reviewed form. Adoption across the UK is uneven: some councils report extensive experimentation, while others have confirmed that they are not using any AI tools in planning workflows. As a result, there is limited national-level data on error rates, bias, or unintended consequences of AI-assisted validation .

This caution reflects an understanding that premature automation of discretionary functions could undermine trust in the planning system. The UK experience suggests that restraint has been a deliberate design choice rather than a technical limitation.

Evidence-based lessons for Ireland

For Irish planners and policymakers, the UK evidence offers lessons, but not a template to be copied wholesale.

First, the research shows that validation is a structurally inefficient stage that lends itself to targeted automation. The types of errors identified in UK applications, missing documents, incomplete forms, incorrect fees, are not unique to that system. Where such errors are repetitive and rule-based, AI-assisted checks can reduce officer workload without encroaching on professional judgement.

Second, the UK experience demonstrates that human-in-the-loop design is not an optional safeguard but a foundational requirement. Systems that preserve officer control, require review of AI outputs, and maintain audit trails are more likely to align with planning law, professional ethics and public accountability.

Third, the evidence cautions against overextension. UK pilots have succeeded precisely because they focus on narrow, well-defined tasks. There is no empirical support, at present, for automating evaluative planning judgement, and UK institutions have been explicit about the limits of AI in this regard.

Finally, the research underscores the importance of transparency and governance. Clear disclosure of AI use, explicit retention of human responsibility, and ongoing evaluation are central to maintaining trust. The absence of comprehensive long-term data also suggests that any Irish exploration of similar tools would need to proceed incrementally, with evaluation embedded from the outset.

In sum, the UK experience does not suggest that AI will “fix” planning systems. It does, however, provide concrete evidence that carefully scoped, human-supervised tools can reduce avoidable delay at the validation stage. For Ireland, the lesson is not about technological ambition, but about disciplined application: using AI where the evidence supports it, and resisting it where professional judgement must remain paramount.