Today I will share my personal experience about implementing Generative AI for a company I have worked for. The maintenance industry can be split up in different categories. In this blog I will focus on: Preventive, Reactive, and Predictive Maintenance. Preventive maintenance is the regular maintenance a building or house needs in order to prevent bigger issues. Preventive maintenance follows a yearly planning. For instance, restoring roofs or replacing central heating boilers. Most of the time this type of maintenance is done for multiple houses at once. Reactive maintenance happens when unexpected failures occur. These events are random and therefore the contractors are unable to plan this type of maintenance. Examples are a smashed window, leaking taps or a door lock that is not functioning anymore. A newer concept of maintenance is predictive maintenance and data-driven decision making. Predictive maintenance is data-driven decision making based on analyzing repairings in certain environmental conditions. The goal is to anticipate on certain issues before they become a problem.
Why is reactive maintenance inefficient?
Reactive maintenance is a significant challenge for social housing corporations, because of the randomness. Reactive maintenance cannot be scheduled or prepared for in advance, making it inefficient for contractors. In addition, the corporations struggled with taking in reactive maintenance repair requests. Tenants can request reactive maintenance by contacting the housing corporation. However, it is quite hard to get a complete technical problem description from the tenants, which resulted in unclear repair requests for the contractors. Moreover, the process of scheduling appointments with contractors was inefficient, since the corporations had no direct access to the contractors’ agenda. As a result, the tenants were called multiple times to resolve the problem, which had a negative impact on customer satisfaction.
Intake Module for Reactive Maintenance
In order to make the process more efficient, the company I worked for developed a software module for the intake of reactive maintenance requests. The Service & Repair Intake module uses a decision tree to help the employees with the technical description for the contractor. The employee is guided through a series of questions, ensuring a complete problem description, even without having a technical background. In addition, the service employees can directly schedule the appointment, because the agenda of the contractors are integrated in the intake module. The software solution led to an improvement of the One-Call-Does-It-All KPI: Renters only have to make one call to report the problem and schedule the repair appointment. The First-Time-Fix KPI also improved significantly, because the contractors could now resolve the issue during the first appointment due to the more complete repair description. In addition the customer satisfaction also improved significantly, because the planning of repairs was more organized than before.
Data-Driven Maintenance Predictions
Over time the company collected an extensive amount of reactive maintenance data. Reactive maintenance is unpredictable and therefore relatively more expensive compared to other types of maintenance. In an ideal world, reactive maintenance would be eliminated entirely and replaced by preventive or predictive maintenance. However, it is almost impossible to predict when a tenant breaks a window by accident, but there were some reactive maintenance repairs that were easier to predict. For instance sewer-related issues. Most of the time a street uses the same main sewer pipe, so problems can relate to multiple houses. By using machine learning algorithms and clustering techniques, the company I worked for analyzed historical and real-time data to identify patterns and peaks in specific areas. The prediction model uses the cluster analysis to identify maintenance patterns based on the historical data from the intake module. As a result, the algorithm can make a prediction about the development of the issue over time. The valuable insights are shared with the social corporations and they have to decide to stick with the reactive approach or opt for preventive maintenance for the entire area, because addressing an entire street’s issues at once is more efficient and cost-effective than dealing with individual complaints over several weeks.
The adoption of Generative AI
The process of sharing the signals and predictions to corporations was quite labor intensive and costly. The module is used to manage more than 150.000 social houses throughout the Netherlands. My idea to automate the process was to implement ChatGPT. I set up a prompt with a reporting template that is filled with the data-output from our model, including the signal, corporation details, maintenance prediction, and the costs and benefits. Whenever our model gave a signal, ChatGPT would generate the report instantly. The reports are automatically sent to the appropriate corporation through our Enterprise Service Bus. The integration of ChatGPT saved me 2 full working days a week and this was a significant cost-reduction for the company.
If I look back at my journey, I really enjoyed working with Generative AI especially when I saw the results. I got excited about implementing generative AI in other business areas of the company. I started straight with brainstorming about the adoption of Generative AI in other processes. As the technology will continue to advance, I expect the technology to have more possibilities for several inefficient business processes of social housing corporations.
Thank you for the interesting post! It’s a little unclear to me how the implemented algorithm predicts when maintenance is necessary? I’m particularly interested in the input data that’s at its disposal. I can’t imagine that the company has sensors in the sewer system or on the boilers of tenants, so what data could this model actually use to predict the need for maintenance? Or, does it simply deduce from the type of maintenance request (e.g., sewer-related), that surrounding houses will probably require maintentance as well? Also, since ChatGPT suffers from hallucination (making up stuff), what did you implement to ensure it’s output for the reports was correct?
The predictive model uses the following parameters to determine the development of a signaled cluster:
Type of building: house, apartment etc.
Location of the building: sub parameters are weather conditions, natural disasters like river floats storm etc.)
Building materials: the corporation can provide us with data of the used materials, like quality and expected material life etc.
Building history: when is it built, by who etc.
Maintenance history: what is the history of repairs
How does the tenant use the building: how many people live there, energy usage, water consumption etc.
Type of renter: yes this parameter is quite tricky to use (is it ethical to use this kind of information?) but we have examined different type of tenants, especially in poorer areas people are less likely to be careful with their house and this will lead to more repair request (=some kind of moral hazard effect, the repairs are paid in a service contract and therefore the renters are less careful in comparison with other tenants)
We do not use sensors in the sewers or anything like that, because it is too costly for relatively small repairs like reactive maintenance.
The model analyzes all this data and tries to match the cluster with other similar clusters. The model then will make a prediction based on how similar clusters developed based on the historical data. The social housing corporation gives feedback about the predictions and we use this to further improve the model, add new parameters and use machine learning (test datasets) to improve the accuracy of the predictions. I cannot dive too deep into the technical aspect of the model, because then I would leak the companies’ secrets 🙂
The signals from our model are quite straight forward. The prompt and reporting template I made for chatgpt only allows it to fill in the name of the corporation, the addresses of the relevant houses, the repair request and the financial data like costs. I prevent chatgpt from making things up, by only asking it to put all the data from the signal on the designated spot in the reporting template. Chatgpt will not generate new information and by doing so I ensure that the output is correct.
For now, we only signal the peak in repairs and show them our prediction. We do not advise the corporation in any way. We will only provide insights the corporation did not have before and that is the value our company creates. The corporation has to make the decision themselves based on the insights we provide. If our model improves over time, my boss is willing to take the risk to start advising (consultancy), but for now the risk of a potential backlash is too big if we give the wrong advice.
I hope this made things a little bit more clear to you.