Leeds City Council and Xylo: transforming planning with AI

Leeds City Council is one of the largest metropolitan authorities in England, processing over 6,000 planning applications per year.

Having undertaken multiple internal reviews and streamlining existing processes, the planning team understood where issues lay in their systems, processes, and digital capabilities. They had already completed much of the troubleshooting internally, utilising tools like Power BI and PowerApps extensively, but had hit a limit in their internal IT capacity. 
 

Xylo builds purpose built AI-powered workspaces for LPAs and prior to working with Leeds had conducted interviews with over 200 officers nationwide and shadowed two other medium size authorities to understand the challenges faced at service level.
Xylo Core is an AI workspace that is designed to help planning officers do their best work and focus on the human elements of planning. It uses AI to pull together the content and context from planning applications, suggesting the most relevant information which the officer can review. It supports officers with key tasks from post-validation through to determination. It frees officers from administrative work and allows them to focus on exercising their professional judgement. 

Whilst the product has core functionalities and features, it is not simply an off the shelf solution. Xylo and the Leeds planning team undertook a process lasting more than six months to iteratively design and pilot a platform that worked for them. 

The planning team identified three priority issues that they hoped to address:

• Time: growing backlogs, staffing crisis and reliance on extensions of time
• Accuracy and quality: complexity of job and impact context switching puts pressure on consistency 
• Wellbeing: admin burden and officer burnout.

To begin to detangle some of the challenges, and understand how Xylo Core could address those, the product team remotely shadowed officers, mapping workflows and processes. This was then put back to officers for feedback and further refinement. An opportunities report then highlighted the areas with the greatest potential for automation and improvement using AI and software. This report then formed the basis for the pilot. 

After securing backing from leadership, the team began building and undertook compliance checks with the IT and AI teams. Prior to the pilot launch, early prototypes were shared for user testing with officers. This meant early issues could be identified and resolved, allowing the pilot to more accurately reflect realistic, real‑world use cases.

The pilot then launched in mid-October with 8 participants, focusing on continuous feedback and product iteration loops. By December, the product reached a critical mass, and the successes and challenges of the pilot began to emerge. The pilot has now expanded to 20 participants.

No technology implementation is without its challenges, and the pilot surfaced four key challenges that the team had to navigate carefully.

• Fear: Officers were initially concerned about what this AI use might mean for their future roles. Addressing this required sustained, open communication and repeated reassurance that Xylo Core was designed with human in the loop principles at its core.
• Trust: The process required officers to work closely with the Xylo team, sharing their workflows and the challenges they faced. This, as well as concerns around sharing data and moving to an unfamiliar system, meant that building trust between partners was vital to making the project a success. 
• Fatigue: Officers were already stretched, and asking them to invest time in learning something new, even where it would ultimately save them time in the medium term, was difficult. Securing that initial buy-in required patience and a recognition that the short-term cost of adoption had to be made as light as possible.
• Integration: Integration limitations meant that some internal data could not be surfaced within the platform, as gaps between Xylo and other third-party providers placed constraints on what officers could see in one place. Work to address this is ongoing.
   

Leeds took a structured, compliance-led approach to AI governance from the outset, drawing on existing information governance expertise and adapting it for the specific risks posed by AI in a public service context.

• Organisational readiness and leadership alignment: Before the pilot began, the planning team secured backing from senior leadership and engaged the council's IT and AI governance teams. The Head of Digital Innovation led the compliance workstream, ensuring that AI adoption was assessed against the council's existing data ethics and information governance frameworks. This included involvement from the council's information governance, IT security, and digital teams to evaluate the proposal against Leeds's internal policies and standards.
• Supplier due diligence and compliance review: Xylo was required to demonstrate compliance with a range of standards before any data was shared or the pilot could proceed. This included ISO 27001 certification, Cyber Essentials, UK and EU data residency for all processing, and contractual commitments to zero data retention by third-party AI model providers. Xylo prepared and shared a comprehensive privacy and AI compliance whitepaper, a Data Protection Impact Assessment, and a Responsible AI Impact Assessment response aligned to the council's own AI screening framework. These were reviewed by the council's governance team ahead of launch.
• Algorithmic transparency: An Algorithmic Transparency Recording Standard (ATRS) record was completed collaboratively between Xylo and the council, providing a publicly accessible account of the tool's purpose, technical architecture, data processing, human oversight arrangements, and identified risks. This is consistent with central government guidance on algorithmic transparency for public sector bodies.
• Human-in-the-loop by design: A core governance principle throughout the pilot was that the AI workspace supports officers — it does not make decisions. Every AI-generated suggestion, whether a policy recommendation, or draft text, must be reviewed and approved by a qualified planning officer before it is used. The system maintains a full audit trail of all actions taken, preserving accountability and enabling review.
• Ongoing oversight: Governance was not treated as a one-off pre-launch exercise. Weekly check-ins between the council's pilot lead and the Xylo team provided a regular forum for raising issues, reviewing performance, and adjusting the approach. Semi-regular check-ins with the Head of Digital Innovation ensured continued alignment with the council's broader AI governance objectives. Xylo also provided weekly written updates to the wider Leeds planning team, including leadership, and held weekly product update calls with the pilot officer group to maintain transparency throughout.
• Officer training and onboarding: Each participating officer received a one-to-one onboarding session covering the tool's functionality, its limitations, and the governance expectations around its use - including the principle that officers remain responsible for all decisions. This was supplemented by ongoing support and feedback channels throughout the pilot.

Risk identification and mitigation were embedded into the pilot from the earliest stages, using a combination of formal assessments and ongoing operational safeguards.

Formal impact assessments. The council's governance team conducted an AI screening assessment, which identified the pilot as requiring enhanced due diligence given the use of AI in a public-facing decision-support context. This was followed by a Responsible AI Impact Assessment, completed jointly by Xylo and the council, covering areas including fairness and bias, transparency, accountability, data protection, and the potential for unintended consequences. A draft Data Protection Impact Assessment was also prepared, mapping the data processing lifecycle from capture through to deletion and identifying specific risks and mitigations at each stage.

Key risks and mitigations identified:

• Accuracy and hallucination risk: The risk that AI-generated outputs could contain incorrect planning law references or policy interpretations was identified as the most significant concern. Mitigations include requiring officers to verify every output against source material, displaying source links and reasoning alongside all suggestions, and running automated and manual evaluation checks for bias and regression using curated planning datasets.
• Data protection and privacy: Planning applications contain personal data, primarily addresses, and may occasionally include special category data such as health or religious information. Mitigations include processing all data within UK-based infrastructure (AWS UK facilities), enforcing zero data retention agreements with all third-party AI model providers, ensuring no council data is used for model training, and applying data minimisation principles so that only necessary information is processed.
• Officer trust and adoption: The risk that officers might over-rely on AI suggestions or, conversely, resist the technology entirely was managed through the human-in-the-loop design, sustained communication about the tool's purpose and limitations, and a phased rollout that allowed officers to build confidence gradually.
• Integration and data access: Limitations in integration with legacy back-office systems were identified early as a constraint on what the pilot could achieve. These were managed transparently with workarounds put in place.
• Security: Xylo's ISO 27001 certification, Cyber Essentials Plus accreditation, and role-based access controls provided the security baseline. Regular security reviews were part of the ongoing assurance process.

Continuous risk management. Risk was not treated as a static, pre-launch exercise. The weekly check-in cadence between the council and Xylo allowed emerging risks to be identified and addressed in near real-time. Officer feedback — gathered through interviews, usage data, and the weekly product calls — fed directly into product adjustments, ensuring that the risk profile was actively managed throughout the pilot.

The results from the pilot, drawn from officer interviews and usage data, were striking across all three of the priority areas the team had set out to address.

The success of the project was supported by a strong foundation for partnership. Leeds had the right degree of scale, digital maturity, and leadership commitment, whilst Xylo had the research and technical skills, AI expertise, and quick development needed to make it work. The success of the pilot was underpinned by four aspects of the process:

• Cultural and behavioural shift: a different sort of relationship promoted success. Officers worked in partnership with Xylo to codesign the product, utilising feedback loops to constantly iterate. 
• Shaped directly by officers: Officer feedback was directly reflected in changes to the product, not just noted down. Weekly product calls, new feature demos and highlighting successes contributed to successful collaboration.
• Success became self-reinforcing: As early adopters began to experience the benefits, word spread and momentum built, with more and more officers wanting to get involved.
• Leadership and organisational buy-in: securing buy-in across the organisation and drawing on expertise beyond the planning department meant that the conditions for that momentum were in place from the outset and having a champion for the technology within the team, helped to foster collaboration between Xylo and the team.

The focus now turns to developing further integration. Plans are in place to expand Xylo Core across the wider planning team and to explore how it can support adjacent teams who face similar pressures. Some challenges remain. The lack of full integration with the council’s planning systems continues to limit what can be achieved, something which will hopefully be overcome in time.

Scaling the partnership without losing the co-design ethos that made the pilot work is another priority. Providing compelling evidence of financial return on investment, within budget cycles, is also an evolving challenge.