.png)
How they work
The input-to-output mapping for this particular technique can be broken down into 3 separate components that form a pipeline for input processing.
Versatility: These models can perform various tasks, including text summarisation, translation, content generation, sentiment analysis, and answering questions, depending on the prompt's design.
Zero-shot and Few-shot Learning: General-purpose LLMs can often generate accurate outputs with little or no task-specific examples. Zero-shot learning involves asking the model to perform a task directly from a prompt, while few-shot learning includes providing a few examples within the prompt to guide the response.
Contextual Understanding: LLMs excel in maintaining context over long passages, enabling them to generate coherent and contextually appropriate responses even for complex queries.
Generality: The generality of pre-trained LLMs is in itself a limitation. While they are trained on vast datasets, they may lack deep expertise in niche or highly specialised fields. This can lead to more incomplete or incorrect responses, the more specialised the task, the greater the likelihood of this.
Hallucinations and Fabrications: LLMs sometimes generate information that is factually incorrect or entirely fabricated, known as "hallucinations." This is particularly concerning in high-stakes applications where accuracy is critical, such as legal, medical, or governmental contexts. An LLM will always respond to a query, but guaranteeing it is grounded in truth requires complex and expensive data governance.
Generality: The generality of pre-trained LLMs is in itself a limitation. While they are trained on vast datasets, they may lack deep expertise in niche or highly specialised fields. This can lead to more incomplete or incorrect responses, the more specialised the task, the greater the likelihood of this.
As with any dataset—not just those related to housing—LLMs excel at tasks such as text categorisation, summarisation, and data cleaning. These capabilities can unlock valuable insights, particularly when manual data processing isn’t feasible due to volume or complexity (e.g., word matching, punctuation correction).
For example, as part of their Home by Home plan, DG Cities successfully categorised repair descriptions for all properties within the Royal Borough of Greenwich for the past 3 years. This process enabled us to prioritise retrofit efforts and strategically allocate properties to different retrofit schemes by combining repair data with stock condition surveys. By cleaning and processing this historical data, we have been able to unlock a deeper understanding of housing assets, revealing new levels of actionable detail.
Notably, processing 600,000 repair instances incurred a cost of under £5 for LLM usage, with only moderate expenses associated with model development. This highlights the efficiency and cost-effectiveness of leveraging LLMs for large-scale data processing.
Once data is cleaned and processed, the following layer of analysis There are many mature frameworks for the use of prompted LLMs in text-to-SQL tasks, allowing for social housing providers to draw relevant data from large, complex datasets with natural language. This will impact providers both in terms of service delivery and cost savings.
By expanding the accessibility of data, councils can make more data-driven decisions on resource allocation, with the text-to-SQL acting as a data retrieval assistant. Furthermore, text-to-sql models can be combined with more powerful reasoning models such as Chat GPTs o1 in order to provide data-driven answers to complex and nuanced queries about the council’s housing stock, and in this case, the impact on service delivery can be transformative. With the barrier-to-entry for interacting with data essentially being eliminated with prompted models, the time-to-answer is significantly reduced, and it will ultimately reduce the need for outside experts to be brought in to perform analysis, meaning the cost savings can be significant.
Since the models are being used internally, even if more expensive models such as GPT’s o1 are being used to tackle more complex queries, the volume of requests to models will be far smaller than if it were to be a service provided to tenants, and therefore the cost burden for hosting this service is minimal. The model will require a simple interface for residents to interact with, and will incur the initial costs of developing the model and pipeline, which due to its slightly more complex nature will be larger than other more typical prompted models. Nevertheless, with no finetuning, or training procedures, as well as no need to research and design model architecture, the relative capital expenditures are quite modest.
These tools can provide immediate responses to routine inquiries, allowing human staff to focus on more complex issues, this will ultimately allow for tenants to receive quicker response times to queries, as well as reducing resource loads on housing teams. Furthermore, translation, which Large Language Models perform well in, especially in more general contexts, can be integrated into the chatbots to ensure communication and support can be provided for tenants from all backgrounds.
To ensure that responses are robust and adhere to council policies mean for more sophisticated and careful design of prompted large language models, with the need for multiple agents each with their own prompt to handle different types of requests, meaning that upfront costs for development can be significant. In this case, chatbots can only be used for a restricted set of possible queries, meaning it is likely to perform poorly for more nuanced queries. Since this will be a service provided to tenants, the overall traffic to the model will also increase, bringing up running costs of providing this service. Finally, with limited means to include examples for the model to follow, ensuring that answers are accurate, and adhering to pre-existing guidelines, cannot be guaranteed, and in applications where accuracy is key, a prompt-only based solution may be unsatisfactory.
Document summarisation using prompted LLMs offers an efficient solution for managing the vast array of legislative, historical, and planning documents—often held in non-digital formats—that underpin a council's vision for a locality. By swiftly condensing lengthy documents into key points, LLMs enable planning teams to access critical information without having to manually review each source. This accelerates decision-making processes and ensures that both overarching strategies and micro-level details are accounted for.
Fintuned LLMs for optical character recognition (OCR) which is able to digitise non-digital records, already populate the market. This can unlock valuable insights from archives, helping councils track precedents and maintain continuity in planning decisions. Since there are a finite number of these documents that need to be processed, this summarisation can be done once and then stored offline, minimising the upfront and running costs of the model. Although unlikely and suboptimal, if the resulting summarisations are concise enough, and can all be fit within the context window of the LLM, then all historic documents can be queried, furthering the level of support provided. Nevertheless Retrieval Augmented Generation (RAG), or AI technique 2 is better suited to this task and offers a more robust solution.
Prompted LLMs can efficiently categorise and summarise the wide range of ideas, concerns, and questions submitted during large-scale consultations. This enhances councils' ability to gauge public sentiment on planning initiatives, ensuring that decision-making is informed by a comprehensive analysis of resident perspectives. Although categorisation can be somewhat grounded through few-shot learning (examples), the effect of this can be limited on diverse data that can’t necessarily be captured in a small number of examples. To completely ground responses in truth, the only other alternative would be to finetune the model on the council’s own database, but is outside the scope of this technique and instead captured in AI Technique 3.
By making the interpretation of consultation feedback more accessible, councils can identify key themes and emerging issues more quickly and accurately. This significantly reduces the time required to process feedback and lowers the reliance on external analysts, leading to notable cost savings.
Since these models are used internally to assist planning teams, the volume of requests is far smaller compared to a public-facing service. Consequently, the cost burden of running more sophisticated models remains minimal. Implementing such a system would require an interface for planners to interact with the analysis and some upfront development costs to establish the model pipeline. However, given the absence of fine-tuning, bespoke training, or model architecture design, the capital expenditure remains relatively modest.
The cross-cutting impacts of the use of prompted LLMs as a data prepping tool apply to Planning teams as well. For example, data from planning applications, land use surveys, or zoning records often come in unstructured formats. With LLMs, this data can be standardised, categorised and structured into tables, providing a cohesive overview of land and development activity across the council’s jurisdiction. Coupled with advanced querying capabilities, LLMs enable planning teams to extract data to identify trends, assess compliance with local plans, and make data-driven decisions efficiently—reducing reliance on external consultants and improving overall planning outcomes.
As discussed in social housing, prompted LLMs can be powerful data cleaning tools, and can transform the usability of existing data. This is particularly pertinent in the case of manually entered data, for example, surveys on road and transport infrastructure. By readying all of their data, councils can then have a clear picture of all transport infrastructure under their management, and pairing this with the data querying system described in the analysis of Social Housing, the decision making process can be supported, reducing the need for external consultations.
Sentiment analysis conducted on top of data collected from feedback portals can help the council understand the pain points in their transport network, much like with Social Housing. The benefits & limitations they will share will be the same.
In health consultations and social care visits, ensuring that patient concerns and problems are accurately recorded can be difficult, and often key information from visits can go missing. Maintaining consistent detail from consultations and visits can be integral to optimising and tailoring care to the needs of each individual. Most API providers already offer speech-to-text services, meaning development costs for an MVP are minimal, and running costs will only need to be considered.
Once transcription is performed, the resulting text can be passed into an LLM agent that has been prompted to take notes from the transcription, and can even be used to fill out report templates. It’s again worth noting that prompted LLMs are limited in their accuracy, especially when applied to a specialised downstream task such as health and social care. There are fine-tuned speech-to-text models which have been trained in medical scenarios, however, since the specific nature of this data cannot be determined, special care will need to be taken to guide LLMs as much as possible, and there will need to be an evaluation period where humans are kept in the loop, to assess whether accuracy is sufficient to fully integrate into existing workflows.
Since this is a model that will be handling private medical information, and recordings of voice, careful consideration and costs will need to be allocated to understand the data lifecycle and ensure that GDPR is met, and that data is stored securely. Furthermore, since audio files and resulting text files will likely be quite long, running costs of such a system may be quite high, and detailed cost-benefit analysis will need to be conducted in order to assess the feasibility of implementation.
Simple and immediate in its potential is the use of LLMs for translation between patients/residents and care providers. For example, during visits, speech-to-text can first record voice and then a separate LLM agent for translation can be used to translate the text to the target language, ensuring seamless communication. The entire pipeline, minus perhaps the interface, are being provided as out-of-the-box solutions from AI providers, therefore development costs are minimal and can be implemented rapidly.
Unsurprisingly, cleaning of data is cross-cutting and equally applies to public and social services. The impact of doing this can help elevate health data and enable more bespoke care. Combined with other social datasets such as income, fuel poverty etc. a clearer understanding of the drivers of health outcomes may support the mitigation of inequalities.
Chatbots can provide personalised advice at scale, acting as a first point of contact for patients seeking information about health concerns or social care services. They can support diagnostics by asking relevant follow-up questions and directing individuals to appropriate care pathways.
For remote monitoring, chatbots combined with sentiment analysis can engage patients in regular check-ins, collecting information on health status through carefully guided and curated questions. Sentiment Analysis systems can then be built on top of this collected data to flag potential issues based on mood or language patterns. It is stringently important that these systems are integrated into existing services, not just to cut costs and reduce human interaction with patients. Careful design can instead enhance and support the existing services, adding continuity to patient monitoring between visits and augmenting care by providing a clearer picture of patient health.
Despite this, the accuracy of these systems is a significant limitation, particularly when dealing with nuanced or sensitive medical information. They may struggle with the precision required for medical-grade applications. As discussed in other impact areas, governance on chatbot outputs can be difficult with purely prompted LLM systems. To address this, chatbot systems must be restricted to only answering only certain questions and outputs. This is achieved by assigning agents to certain queries, with their outputs restricted by specifying in the prompt.
Moreover, governance challenges must be addressed to ensure compliance with data privacy standards like GDPR. The handling of sensitive patient information necessitates robust data security and clear lifecycle management. These systems must undergo extensive testing and oversight before integration into workflows to guarantee they meet the accuracy, safety, and ethical standards expected of critical care provision. The opportunity for enhanced service delivery through chat bots is huge, however, solely prompted LLMs may not provide the sufficient robustness to achieve this potential.
© 2025 DG Cities. All rights reserved.