At Nesta, we aim to accelerate the adoption of heat pumps in the UK to help reduce household carbon emissions. Heat pump adoption in UK households remains low, partly due to a lack of installer availability and confidence among newly trained professionals.

Engagement with industry stakeholders and installers revealed high demand for technical support but limited capacity, causing long wait times. Technical documentation is often too concise for newcomers, while manufacturers worry that longer materials may go unused.

To address this, we are exploring how generative AI can make documentation more accessible, helping heat pump installers work more productively.

Our goal was to identify impactful use cases for deploying generative AI in the heat pump sector and to advance its adoption. Our aim was to empower installers by boosting their confidence, increasing productivity, and improving the quality of installations by using this technology.

In collaboration with Renbee, one of our ventures, we’ve developed and tested an early version of our generative AI chatbot Installer Pal. Now integrated onto Renbee's platform – which is designed to simplify installation for renewable energy installers – the chatbot aims to provide heat pump installers with instant access to essential support materials, saving them time whilst on the job.

Research with industry stakeholders and installers highlighted the potential of generative AI chatbots to tackle key challenges for heat pump installers. From user testing and research, three strong use cases emerged:

• Training
  Help newly-trained installers build confidence and test their understanding of heating and heat pump systems. The chatbot acts as a virtual mentor, reinforcing training and providing a resource for revisiting key concepts when needed.
• Assisting
  Support installers with situation-based questions, whether conducting surveys or addressing on-site technical issues. The chatbot provides quick, context-specific guidance, enabling installers to troubleshoot problems and make informed decisions without waiting for a senior engineer.
• Safeguarding
  Help installers quickly check the latest MCS guidelines or regulations, ensuring installations meet industry standards and remain compliant as rules evolve. The chatbot serves as an always available reference, reducing errors and non-compliance.
  

User testing revealed that the chatbot has the potential to significantly boost productivity by saving time on tasks such as retrieving heat pump dimensions and technical guidelines. Installers also found the chatbot to be a helpful co-pilot, enabling them to validate ideas and explore technical concepts, which is particularly valuable for those new to the field. Although some responses lacked specificity, users appreciated access to credible resources, noting that it could help them better adhere to industry standards and improve the quality of their work.

Technical development

Over the course of this project, Installer Pal has grown from a prototype with a single installation manual to a version integrated onto Renbee’s platform, now offering access to over thirty documents.

Evolving from concept to early version

The prototype combined a large language model (LLM) with a heat pump document database to ensure accurate, relevant answers. Early user testing showed it could support installers needing quick technical help. Based on feedback, we expanded its resources to include industry manuals, best practices, and other trusted materials, making it more versatile and aligned with installers' needs.

Continuous improvement

Early user testing revealed a need for concise responses, simpler language, and better alignment with UK-specific terminology to make Installer Pal feel like a helpful colleague. Updates to the model and improved information retrieval enhanced its ability to differentiate between topics including installation processes and compliance requirements.

Evaluating the chatbot

As the chatbot developed, it became clear that manually judging responses by whether they looked right was insufficient. We adopted a systematic evaluation approach with specific metrics to objectively track performance and identify areas for improvement. For instance, one metric assessed how effectively the chatbot retrieved and incorporated database information, highlighting gaps in its knowledge base.

We also implemented a monitoring platform to track metrics, gather user feedback through a smiley face rating system, and monitor operational costs to ensure the viability of scaling a tool like Installer Pal.

For those interested in exploring the technical details, the codebase for Installer Pal, including the systematic evaluation methods and database integration, is available here.

User testing

User testing evaluated how Installer Pal performed in real-world scenarios and its potential to support heat pump installers. Installers highlighted its ability to act as a co-pilot, helping them bounce ideas and validate technical concepts. The chatbot proved especially time-saving, such as quickly retrieving heat pump dimensions – a task often taking 30 minutes manually.

It also showed educational value, helping users test knowledge they had learned the previous week and accurately retrieving specific information, such as return flow temperatures from installation manuals.

However, areas for improvement emerged. Sometimes Installer Pal’s responses were generic, making them less actionable. Transparency about information sources was noted as critical, and formatting issues –including unreadable equations or lack of images –were flagged. Usability across devices and clarity on when and why to use the chatbot were also identified as areas needing improvement.

Installer Pal excelled at retrieving MCS documentation, saving installers time with instant access to detailed content. However, when information wasn’t in its database, it produced generic responses. Improvements that could be made include adding page references, links to documents, enhancing device usability, and improving formatting with images and clearer equations.

Following a technical handover from Nesta, Renbee plans to expand Installer Pal beyond information retrieval –envisioning it as a co-pilot for completing forms such as the distribution network operator (DNO) application. We’re exploring vision embedding models to interpret schematics and looking into ways manufacturers could provide more comprehensive, accessible manuals.

By making technical information more actionable, we bridge the gap between siloed expertise and on-the-job needs, supporting the shift to low-carbon technologies. We invite installers, manufacturers and other stakeholders to share their thoughts and try Installer Pal on our dedicated landing page.