Bath and North East Somerset Council has confirmed it used Microsoft Copilot to process more than 5,500 public comments on a contentious plan to build an 18,000-seat rugby stadium in the city’s historic center—a decision that lifts the veil on how local authorities are deploying AI to manage democratic participation at scale. The revelation, buried in a 121-page planning officer’s report, marks one of the most visible tests yet of generative AI’s role in civic decision-making, and it has reignited a fierce debate over transparency, accountability, and the very nature of public consultation in the digital age.
The council’s planning committee is due to consider Bath Rugby’s resubmitted application on September 17, though central government has already instructed councillors not to grant approval without “special authorisation,” signaling the high political stakes. At the heart of the storm is the Recreation Ground—or the Rec—a cherished green space in the middle of a UNESCO World Heritage city that has been the subject of a three-year tug-of-war between the club, conservationists, and thousands of vocal residents.
According to the planning officer’s report, 5,590 representations were lodged during the latest consultation, with 5,086 in support, 368 objections, and 136 uncategorized. The sheer volume presented a familiar headache for an under-resourced planning team: how to fairly distil the substance of thousands of comments—ranging from a few words to lengthy technical submissions—into a manageable summary for elected members. The council’s answer was Microsoft Copilot.
“Given the very large number of representations submitted via the council’s on-line comments form, these have been reviewed and summarised by Microsoft Copilot,” the report states. “This is an artificial intelligence tool that was instructed to identify reasons for objection/support. The programme reviewed all comments received since the application was submitted. The topics below are derived from those generated by Copilot and subsequently refined by the case officer based on a review of a sample of comments.” The officer then individually read and summarised representations sent directly to them, again using Copilot.
That sentence—casual in tone but seismic in implication—raises a host of questions that sit squarely at the intersection of technology and governance. For Windows users and IT professionals, the use of a familiar Microsoft 365 tool in such a high-stakes public sector scenario is both a natural progression and a cautionary tale. Copilot’s integration into the council’s existing workflows speaks to its rapid mainstreaming in government, but the lack of detailed procedural disclosure exposes gaps that could undermine public trust if not addressed.
The Planning Battle for Bath’s Recreation Ground
The stadium proposal has been through multiple iterations, with the club’s latest plans reducing some heights and redesigning facades to better integrate with the Georgian cityscape. Bath Rugby argues the new stadium will deliver economic and social benefits, regenerate the riverside, and provide a multi-purpose venue. Opponents—including high-profile figures like film director Ken Loach—insist the development would permanently damage a public open space gifted to the city in 1956 and threaten Bath’s World Heritage status. The intervention by the Secretary of State for Housing, Communities, and Local Government means the ultimate decision could be taken out of local hands entirely.
In this charged atmosphere, the planning officer’s report carries enormous weight. It shapes how councillors understand the balance of opinion, and its summary of objections and support can influence the final vote. So, was AI the right tool to produce that summary?
How the Council Used Microsoft Copilot
The council’s approach was a hybrid workflow: Copilot ingested all online comments, generated a list of topics and reasons for objection or support, and the planning officer then spot-checked a sample of comments to validate and edit the machine’s output. Comments sent outside the online form were also summarised individually via the same tool. The full text of every submission remains available on the council’s planning portal, so anyone can cross-reference the AI’s work against the originals—in theory, at least.
On paper, this is a textbook example of a “human-in-the-loop” system. The officer acted as a reviewer, and the final summary was not simply spat out by an algorithm. But critics will note that “review of a sample” introduces its own risks: how representative was that sample? Were complex heritage arguments or technical flooding data given equal weight to short, emotive messages of support? The report does not say.
The Dual-Edged Sword of AI Summarization
Defenders of the approach point to genuine operational benefits. Planning departments across the UK are stretched thin, often handling thousands of applications with skeleton crews. Automating the initial triage of repetitive comments frees officers to focus on specialist assessment, site visits, and the crafting of defensible recommendations. Copilot can apply consistent topic tagging, reducing the variability that comes with multiple human readers. If the audit trail is maintained—logs of prompts, raw AI output, and officer edits—the process could theoretically be more traceable than a lone officer’s manual précis.
Yet the risks are substantial and well-documented. Large language models are prone to hallucination, confidently generating plausible but incorrect connections. Topics that appear frequently in short comments might swamp a single, technically detailed objection that raises a material planning consideration. A resident who submits a carefully researched note on flood risk could find their point submerged in a sea of generic boosterism. Moreover, the council has not disclosed whether the Copilot instance used was configured to prevent data retention, whether prompts were stored, or what contractual protections govern the handling of citizens’ personal data.
Data Protection and the Public’s Right to Privacy
When residents submit comments to a planning portal, they often include names and addresses. Feeding that data into a cloud-based AI service triggers immediate privacy obligations under the UK GDPR. Did the council conduct a Data Protection Impact Assessment? Does Microsoft’s commercial agreement include a “no-training” clause that prevents customer data from being used to improve underlying models? Are logs deleted after a set period? These are not arcane technicalities; they are core safeguards that differentiate responsible deployment from reckless experimentation.
In this case, the council has not proactively published any such assurances alongside the committee papers. For a local authority that already uses Microsoft 365 extensively, the temptation to switch on Copilot without rigorous governance is real—and dangerous. Other UK councils have piloted Copilot for back-office tasks, but typically with DPIA publication, governance board oversight, and phased rollouts. Bath’s apparent leap into high-stakes consultation analysis cries out for a similar level of openness.
What Good AI Governance in Planning Looks Like
Experience from early-adopter public sector bodies suggests a checklist of minimum requirements. Crucially, every AI-generated summary used in a committee paper should be explicitly reviewed, annotated, and signed off by a named officer, with a plain-English explanation of what the AI did and how human checks were performed. The audit trail should be published—ideally as an appendix to the report—including prompts, raw AI output, and the edited final version. A DPIA must be made publicly available, along with a description of contractual protections and data handling procedures.
Technical safeguards matter too. Councils should insist on EU data boundary residency, ephemeral compute where possible, strict access controls, and contractual clauses that forbid Microsoft from using ingress data for product improvement. Even then, the weighting rules applied to the AI’s topic extraction need careful calibration: frequency alone must not determine significance. Officers should manually flag and elevate submissions from statutory consultees, experts, or those containing new material facts that the AI might overlook.
Finally, the process must be appealable. If a campaigner or councillor questions the summary, they should be able to trace each theme back to representative original comments, with links or references provided in the planning portal. Without that, any veneer of transparency evaporates.
The Microsoft Angle: Copilot’s Civic Debut
For Windows enthusiasts, Bath’s experiment is a real-world stress test of technology that many organisations are now adopting. Microsoft has positioned Copilot as the intelligent assistant that can summarise emails, draft documents, and analyse data, all within the familiar Office interface. In local government, that pitch is seductive: fewer late nights for planning officers, faster turnaround for residents, and a modernised public service. But Bath’s case highlights that the tool’s output is only as trustworthy as the governance wrapped around it.
If the council cannot produce a clear audit trail or demonstrate robust privacy safeguards, the credibility of the entire planning decision is at risk—not just for the stadium, but for future AI-assisted decisions. And if Microsoft’s own contractual terms do not transparently support public sector accountability, the company could find itself named in judicial reviews or Freedom of Information disputes. Both the vendor and the customer have a shared interest in getting this right.
The Broader Implications for Local Democracy
The Bath stadium is not an isolated incident. Across the UK, councils are contemplating or already using AI to summarise consultation responses, draft committee reports, or even power chatbots that answer resident queries. The allure is understandable: the volume of public engagement is growing while resources shrink. But consultation is not a customer service function; it is a cornerstone of representative democracy. The process by which views are synthesised shapes the perceived legitimacy of decisions that affect people’s homes, environment, and heritage.
When a planning officer’s report states that AI “reviewed all comments” and a human merely “refined” the output, it raises an unsettling prospect: that the civic act of writing a planning objection has been reduced to a data point in a machine learning pipeline. Even if the outcome is identical, the process matters. Residents must believe their voice is being heard by a human decision-maker, not probabilistically interpreted by a neural network.
Conclusion: A Tool Demands Transparency
Bath and North East Somerset Council’s use of Microsoft Copilot to process 5,500 stadium comments sits at a technological crossroads. On one hand, it is a pragmatic response to a genuine operational challenge. On the other, it is a procedural choice that demands rigorous, public-facing governance. The hybrid workflow described in the report could, with the right safeguards, serve as a blueprint for other authorities. But without immediate publication of the audit trail, the DPIA, and the contractual protections behind the service, the council leaves itself open to accusations that it has substituted algorithmic efficiency for democratic deliberation.
Councillors meeting on September 17 must be equipped to probe the AI’s role directly—asking for sample comments, querying the sampling methodology, and insisting on full documentation. Every future planning committee that confronts an AI-assisted report will be watching Bath’s example. For the tech community, the lesson is equally clear: tools like Copilot are not neutral. They amplify both the best and worst of the processes they automate. In the hands of a local authority, they become instruments of governance. And governance, by its very nature, demands the light of transparency.