The use of real estate and infrastructure, building services systems, and real estate transactions generate an ever-increasing amount of data. On the basis of this data, new services can be created for real estate brokerage, optimization of energy use, and intelligently serving buildings and environments. Artificial intelligence (AI) solutions will be of great importance to users and owners of built property.
Blok – digital real estate brokerage
Blok is a digital real estate agent service operating in Finland and Sweden that combines a professional real estate agency and an online service.
Comprehensive online brokerage service
Blok serves its customers through its website. Blok provides an assessment of the apartment, manages the apartment's description, orders and goes through the apartment's documents, prepares housing advertisements, markets the apartment, and delivers brochures and signs to the seller for showing.
Blok responds to buyers’ inquiries and facilitates the bidding phase. Blok also prepares the deeds of sale and makes the necessary declarations to the housing association and the Tax Administration.
Utilizing AI in housing
Blok is a startup – its business idea arose from the personal experiences of its founders and their desire to reform the industry.
Creating AI solutions for a new way of doing things is often simpler than applying them to established policies. A pioneering company isn't burdened with ready-made policies or systems or processes that involve a large amount of manual work.
Blok uses machine learning solutions for the company's core services and products, housing pricing, and the “Etsivä” service, which seeks a suitable apartment for the buyer outside of the public market.
The pricing service is at the heart of Blok’s business. Pricing is based on comprehensive home sales data sets. Data sets related to housing are available from market data, completed housing transactions, and real estate register data.
As the housing market is constantly evolving, the training data for the model rapidly goes out of date. Of course, an AI solution using out-of-date data can't be reliable. The machine learning model must therefore be updated with the latest data – a so-called learning loop.
Learning loop: training machine learning models with new data
Training a machine learning solution with new data
Ready-made datasets, either free or purchased, allow the solution to be quickly developed and tested. However, it's equally important to ensure that the solution allows data to be collected even during use. Often, machine learning models need to be trained with regularly updated data for the solution to work reliably. This is called a "learning loop".
All data generated by GenAI models, including human interactions e.g. in a self-service agent, should be logged so that areas for improvement can be determined.
In the context of Generative AI (GenAI) development, it is critical to carefully evaluate and determine the strategies for continuous learning and maintenance.
This evaluation is essential for choosing between prompt engineering, fine-tuning, or a hybrid approach as the optimal method for model improvement.
Prompt engineering is a low-cost, flexible and very agile method of continuous improvement. Prompts can be improved continuously by analysing the logs of the data produced by the models, and adjusting the prompts accordingly to cover new issues discovered in the logs.
Fine-tuning, although an effective means of tailoring models to specific requirements, can be a financially demanding process, particularly if applied on an ongoing basis. Additionally, it is important to note that only select versions of models are eligible for fine-tuning. Most vendors providing API-based services typically restrict fine-tuning capabilities to older model versions, excluding the latest iterations from such modifications.
The hybrid approach combines elements of prompt engineering and fine-tuning to leverage the strengths of both methods for optimizing generative AI models. In this approach, prompt engineering can be used for quick adjustments and fine-tuning for more substantial, long-term improvements.
Data is the foundation of machine learning systems. It’s very common that when an organization develops AI systems, the initial stumbling blocks are data collection and data quality issues. When developing the first applications, an organization often has to create new practices for collecting, storing, processing, and managing data.
The management of unstructured data, in particular, has gained increased significance in the context of generative AI applications. Unstructured data, such as text, images, and videos, often contains the nuanced information that generative AI systems need to produce sophisticated and nuanced outputs. Therefore, the ability to effectively manage unstructured data not only addresses the immediate needs of generative AI use cases but also enhances the organization's capacity to leverage AI for innovation and competitive advantage.
Data is at the heart of the solution
Data availability often poses challenges and constraints. For Blok, data scarcity is customary to the field. Apartments may not necessarily change owners very often, resulting in minimal sales data.
Many AI solutions have the same problem. For example, a system for preventive maintenance of factories or machines may not have any data available on fault incidents. In this case, other ways of training the model must be found.
There is also another interesting, data-driven challenge to the home trade: everyone has their own view of the “right” price. It's possible that the price proposed by the Blok solution doesn't seem right, even if it's statistically correct.
If users begin to doubt the results produced by the solution, questions about the explainability of the model arise. The initial data for Blok’s machine learning model can be explained, but the reason why it ends up at a certain outcome is not self-evident. However, a mechanical method of calculating the average price per square meter of completed transactions and multiplying this by the property's number of square meters would produce weaker price-predicting results.
Aha!
We talked at the beginning about the term “black box”. This term implies we aren't able to tell exactly how a machine learning model ends up at a particular outcome, such as a prediction. Although it refers to the technical implementation and functionality of the model, the phenomenon is also emphasized in the implementation and use of AI solutions.
Users need to understand how and why a model ends up with a particular prediction. Otherwise, there is a risk that users won't trust the predictions provided by the model and the solution will be useless. This doesn't have to mean a technical understanding of machine learning models, but rather a process-specific theoretical understanding of what kind of data is entered to the model at any stage. The Granlund AI Energy Mapping image presented earlier in this chapter serves as a good example of an adequate high-level description.
In the domain of Generative AI (GenAI), the landscape is characterized by the availability of both black box models, whose internal workings are not disclosed by their creators, and open-source models, which offer transparency and accessibility in terms of their underlying algorithms and training data. Users engaging with these models have at their disposal various prompt engineering techniques to enhance interaction and output quality. Among these techniques, the Chain of Thought prompting method stands out as a powerful tool for elucidating the reasoning process of AI models. The Chain of Thought approach involves crafting prompts that guide the AI to provide a step-by-step explanation of its thought process leading to the final answer or output.
Sign up to solve exercises
Spacent – rental of business premises
Spacent is a startup that came out of Aalto University in 2019. Spacent provides a platform for hybrid work in the Nordic countries, based on more than 10 years of research on the flexible use of facilities and hybrid work models at Aalto University, MIT, and Tongji University in Shanghai.
A business idea to a real need
The university's premises' average occupancy rate was 24% during office hours, while additional space was under construction. The university wanted to increase the utilization rate of its facilities. This everyday problem gave rise to Spacent’s business idea of a cloud-based technology platform that brings together space providers and business premises in need.
As a platform operator, Spacent has two customer groups: premises owners and companies or individuals renting premises. Both parties have their own value propositions. For business premises owners and operators, the service offers an opportunity to increase the premises' utilization rate and make use of extra capacity. In turn, the user gets more flexible space solutions without overcommitting capital – customers can rent premises on an hourly basis in locations that suit their needs.
A neural network-based simulator
Spacent's technology platform has been in use by customers since the autumn of 2020. The usage data collected from the platform is used in a neural network-based simulation tool. Based on a customer's number of employees and locations, it can determine the effects of different business premises options on the organisation's space requirements, space costs, commuting, and CO2 emissions.
Aha!
Data processing and integrations always take more time than expected at the beginning of a project. This should be taken into account at the design stage, rather than rushing or trying to find shortcuts. This is especially the case when data needs to be obtained from multiple sources or when different systems need to be adapted to work together.
In GenAI projects, while data processing and integrations may still need additional consideration, a proportionately higher effort should be budgeted for QA, especially for stress testing, before launching the service to an external audience. The nascent nature of these technologies means they are frequently subjected to the discovery of new vulnerabilities or methods of "jailbreaking," which can exploit or circumvent the intended operational usage of the AI.
Such vulnerabilities not only pose a risk to the integrity and reliability of GenAI services but also have the potential to compromise user trust and the overall utility of the solution. Therefore, investing significantly in rigorous QA processes, including comprehensive stress testing, becomes crucial.
Future developments
So far, the amount of accumulated data has been fairly small, but the accuracy of the forecasts is constantly improving as the amount of data increases.
In addition to searches for suitable sites, Spacent intends to further develop the platform so that it can predict the use of the premises. The recommendation function of the sites and the commercialization of collected data for other uses are also being considered.
Sign up to solve exercises
Ramboll – proactive maintenance by monitoring the condition of structures
Earlier in this course, we introduced Ramboll’s Brutus solution, a traffic simulation model. The company's second AI solution is related to monitoring the condition of structures. In particular, it has been applied in forecasting the condition and lifecycle of wind power plant structures, but it can also be generalized for monitoring other structures.
Aha!
Most AI solutions are narrow, but there are some AI solutions are universal, so a solution developed for a specific use may be replicable for other tasks with only a small amount of work. Different utilization opportunities often emerge during a project, and the final use may differ from the original plan.
This is especially true for GenAI, whose capabilities are thought to be emergent. The same underlying GenAI model can be used for a variety of use cases by appropriate prompting or fine-tuning. The model was however never trained to perform the specific task during the training phase. GenAI Language models were trained to predict the next word, given an input sequence of words, or to predict the missing word in a sentence.
The physics of structures are well known, and their behavior can be simulated in realistic environments by using precise plans of the structures. With comprehensive instrumentation, data can also be collected during use to provide feedback on the design and simulation. However, such extensive instrumentation is expensive, so when a wind farm consists of hundreds of identical structures, it's more cost effective to install comprehensive instrumentation in only some of the sites and then predict what will happen to the remaining structures. The forecast for all the structures can be produced using data from the turbine control room software.
Use of simulations is especially important in modeling the behavior of structures. Errors mustn't be allowed to occur in the real objects, so creating suitably accurate training data for an AI solution is challenging. Simulations can be used to model structural errors in a digital environment and then identify and predict errors using the resulting data.
Towards preventive maintenance
The operational reliability of wind turbines must be guaranteed even in demanding conditions. Even small disturbances have a large cost effect, because after a breakage maintenance personnel can't get to the site quickly due to long distances. For this reason, the goal is to move towards preventive maintenance by predicting when a machine or component is in danger of breaking down. In this case, it’s possible to carry out maintenance before the damage occurs.
Ramboll's solution serves an identified business need and generates clear financial savings for customers. In addition, preventive maintenance also has a positive impact on the environment as unnecessary on-site visits can be reduced.
Sign up to solve exercises
Lassila & Tikanoja – a real estate service business at the minute level
Lassila & Tikanoja (L&T) is a Finnish service company group focused on providing environmental maintenance and support services for real estate and facilities. L&T operates in Finland and Sweden and the company employs about 8,100 people.
Lassila & Tikanoja's Real Estate Services unit has data analytics solutions at three levels:
Production control
Big data and learning from it
Agile service development and customer interface
The utilization of the data was preceded by an extensive three-year project in which the division completely renewed its own enterprise resource planning (ERP) system and related processes.
L&T utilizes analytics for production control, which is the focus of its business. The company is a good example of how advanced analytics can often be enough without needing to take advantage of machine learning.
Advanced analytics is often enough to create value.
Real-time ERP
In connection with its ERP reform, L&T launched the development of the Dynamic Scheduling system. According to the company, the result was perhaps Finland's most advanced human resource optimization tool, which ensures the right resources are in the right place, at the right time.
In the system, the workday consists of six-minute periods in which tasks are processed in order of urgency. The employees have a mobile app where they record when they are starting a task on the job site.
The employee’s mobile app assigns tasks to the nearest six minutes
Situational awareness allows users to locate nearby employees. In this case, they can be called to a location quickly if a more urgent work task, such as a pipe break, comes up.
Analytics for customer management
Analytics also help in customer relationship development and short-term forecasting. In its contract portfolio, L&T monitors the development of invoicing, customer satisfaction, complaints, customer feedback, and the number and quality of customer service requests.
In its agile service business, the company invests in analyzing customer feedback to get a deeper understanding – for example, if a particular issue or customer feedback is repeated on a regular basis.
Challenges and opportunities for the use of AI
L&T’s challenges with implementing AI are quite common. Existing systems and their architecture may not support the use of AI. Another challenge relates to a balanced consideration of internal and external service needs.
In many cases, developers of AI solutions emphasize internal efficiency. However, it's important to corroborate the development need with the solution's users. The users must always be involved in the solution's design and development to ensure it generates value for all parties involved.
From idea to service
L&T has tried to solve user-oriented challenges with a company-wide “from idea to service” concept. In practice, this means there is a tool anyone can use to suggest new ideas. The agile development team picks ideas every other week and evaluates them using the following criteria:
Does the idea have commercial potential?
What is the technical implementation of the idea and what does it require?
What resources does the implementation require, and are the resources available?
The team will work on the idea with the customer before making a final investment decision. This ensures there's real demand for the solution.
Future development of AI
In the future, existing analytical expertise will be enriched even further by utilizing machine learning. To this end, L&T is building an information management infrastructure that will also produce data to support AI solutions.
Thanks to these actions, it's possible for the company to take a leap forward in the development of AI. The fact that L&T is focused on bringing value through analytics tends to create trust in the possibilities of AI more broadly.
As the use of AI in the industry is in its infancy, the company will be able to differentiate itself from competitors through AI without significant investment.
