RE-generative AI: How technology can transform commercial real estate
While real estate firms are looking for ways to leverage generative artificial intelligence, they should be mindful of the risks and adopt the approach that best suits their needs.
Broader AI adoption has laid a solid foundation to build upon—but there’s a long way to go
Leveraging artificial intelligence in business is not new. Using machine learning (ML) and language processing across a range of functional areas and skill sets has been in practice for decades.1 However, that traditional scope has had limitations, leaving more creative disciplines to human thinking. The emergence of generative AI has shifted our thinking—with use cases demonstrating that the automation of creativity and imagination could be a reality sooner than may have been anticipated.
Generative AI is a subset of traditional AI, in which a machine can create original images, videos, text, audio, code, or virtual environments based on a set of existing content that would have previously required human expertise to create. Both traditional AI and generative AI have the capacity to enhance how commercial real estate firms operate, but it is important to distinguish between the two and understand the true emerging capabilities of the “generative” aspects of AI.
There are indications that strategically leveraging broader AI tools can play a significant role in transforming the commercial real estate sector. According to Deloitte’s 2024 commercial real estate outlook survey, over 72% of participating real estate owners and investors around the globe are already committing, or plan to commit, hard dollars to some type of AI-enabled solutions within their organizations.2
Here’s what a Deloitte Center for Financial Services analysis of real estate firms’ investment into broader AI and ML companies found: Since 2017, there have been considerable levels of venture capital investment, totaling US$7.2 billion.3 AI and ML companies analyzed include those that develop unique large language models (LLMs), fine-tuning frameworks, front-end AI assistants, and digital infrastructure companies, among others. Since the emergence of generative AI in 2021, corporate investment volumes totaled over US$3.5 billion through October of 2023, outpacing the total from 2018 to 2020 by nearly 50%, and the total of the three years prior to the pandemic by 95%. The areas of greatest interest to these real estate investors (figure 1) include AI and ML services for transaction-focused function areas like property listings (42%), investment and valuation (20%), and real estate data analytics (8%).4
Despite stated commitments and the notable investment trajectory, more than 60% of respondents to the 2024 commercial real estate outlook survey also indicated that they are still reliant on some forms of legacy technology infrastructure and face difficulties in adopting emerging technologies like generative AI. This could potentially be a considerable challenge in integrating even rudimentary generative AI capabilities across their organizations.
More than meets the eye
The content created by generative AI applications that businesses and consumers see and use—text, image, video, and audio—is likely only the tip of the iceberg. These applications are supported by a generative AI technical stack of predictive models, data platforms, and infrastructure that together can enable generative outputs for more specific tasks or industries. Generative AI applications are only as good as the data that feed them.
Generative AI and its moving parts: High level architecture overview5
- Generative AI applications: The front-end interface that connects end-users to trained models
- Generative AI models:
- Fine-tuning and training frameworks: With traditional AI frameworks, the use of predictive models in generative AI applications create outputs based on assumptions about the inputs received. With generative AI, to bring a human-like level of communication, model fine-tuning, prompt engineering, and adversarial training are incorporated for better model understanding.
- Foundation models: LLMs process vast amounts of data to form “memories” that shape the model’s parameters.
- Cloud and data platforms: Platforms that store petabytes of big data that is fed into models
- AI infrastructure: The computational capacity needed to train and process these models requires leading-edge processing units on scalable infrastructure.
Generative AI in action: Prioritized sample use cases for real estate
There is a wide set of potential use cases (figure 2) for real estate companies across various generative AI modalities, such as text and data, image and video, and immersive experience, and across real estate functional areas, including property or facilities management, construction management and procurement, legal due diligence, enterprise-wide applications (risk and knowledge management), property leasing and marketing, property valuation, market analytics, architectural design and rendering, urban planning, and virtual environment generation.
These use cases fall under varying degrees of existing maturity, ease of adoption, and scalability within an organization. Cases such as summarization or auto generation of contracts come from cross-industry applications, and they have already been validated and can likely be more easily integrated by real estate firms through existing public applications or third-party application programming interfaces (APIs). More data intensive, privacy sensitive, or cost-intensive cases such as urban planning or scenario generation are still likely at a conceptual stage, at the time of this writing.
Depending on the use case priorities and factors such as model customizability, data privacy, time to market, or cost implications, firms can consider appropriate approaches for generative AI integration. These approaches can include, but are not limited to, the following:
- Use existing generative AI applications: This is a faster way to integrate generative AI, but comes with risks of data leakage and unauthorized access to proprietary data.6 Firms can use widely available public applications online to summarize lease documents or modify existing images. However, these applications are trained on information from open sources and cannot respond to prompts regarding proprietary knowledge unless they are trained for a specific use case.7
- Integrate third-party APIs in your organization’s workflow: APIs offered by LLM service providers can be an effective way to integrate pretrained LLMs into existing systems, thereby reducing the need for in-house technical experts and the time for deployment.8 Some public providers allow API access to their models that are already pretrained for specific use cases, which allows firms to integrate those chat applications into the existing organizational workflows.9 Although these third-party APIs save time, they can potentially compromise data privacy.10 Firms should consider service providers’ responsible AI practices, safety principles, and adherence to upcoming or finalized AI-related regulatory requirements. Model version updates may require changes to APIs, which can increase model maintenance costs for the firm.
- Use an open-source model and train it on your specific proprietary datasets: Open-source models can be deployed using enterprise infrastructure, whether on-premises or in the cloud, preventing data leaks and helping to ensure more robust data privacy and security.11 Leveraging an open-source LLM offers cost savings, can enhance time to value, and requires fewer resources than building in-house LLMs.12 Training on proprietary data allows for visibility into algorithms and more control over the methodologies. However, the initial costs of implementation can be considerable—most notably, the costs of cloud or on-premises infrastructure.13 Training on proprietary datasets, fine-tuning, and adapting the models may involve working with a vendor, which can bring on additional costs.
- Create a private LLM (PLLM) model in-house: In-house PLLMs rely on proprietary data of the enterprise, give complete control over data, and model behavior that ensures the confidentiality of sensitive information14 such as bank account details, social security numbers, contact details of tenants, or other property information specific to the real estate sector. Concerns around data privacy and security have led to the adoption of in-house PLLMs in sectors like finance, health care, and life sciences.15 Since these PLLMs are trained on sector-specific data, the outcomes are highly relevant with reduced chances of statistical errors or hallucinations, but are likely the most expensive and resource-intensive of all approaches.
However, there is no one-size-fits-all implementation strategy across all use cases, and there may be tradeoffs for using certain approaches over others. For instance, if a real estate firm wants a highly customizable generative AI application that can conduct predictive analytics for leased asset performance (performance analysis use case), then, they should more likely consider an open-source model trained on proprietary data or a PLLM (figure 3), which may come with high costs to implement. That said, if they want to prioritize cost savings for a similar solution, using an existing application or third-party API may be a more efficient approach, with the tradeoff being additional potential data privacy risk.
Keeping a human-centric approach to generative AI
Generative AI should help augment or enhance the human experience, rather than outright replace it.16 As real estate firms continue to research the art of the possible with generative AI, and potentially prepare their data and technology to be able to incorporate it into their organization, our research found that some firms have already started recruiting talent that can help steer their digital transformation (see sidebar, “About our real estate generative AI hiring data research”).
Research findings indicate that job postings by real estate firms requiring generative AI skillsets increased by 64% in 2022 and by another 58% through August of 2023.17
About our real estate generative AI hiring data research
The Deloitte Center for Financial Services analyzed over 500,000 job postings for AI- and ML-enabled positions by real estate firms across the United States since 2021. Job postings were then evaluated for required applicant skillsets specific to generative AI, including capabilities with large language modeling or prompt engineering, among others.
A deeper dive into the types of real estate functions with the largest share of hiring activity offers a peek into which disciplines of the industry are early movers into leveraging generative AI talent.18 About 34% of these jobs are in architectural design, followed by construction/project management at 18%, legal due diligence at 17%, and human resources and talent management at 13% (figure 4). These functions are likely targets for early-stage generative AI capabilities, supporting creative functions like design and construction, and text summarization and synthesis in the document-heavy legal and human resource fields.
Not included in these statistics are those within the property management segment, of which a majority of jobs were exclusively tied to developing or integrating generative AI–powered chatbot applications. This segment proved to be an outlier in the data set due to rapid emergence exclusively in 2023. Nonetheless, hiring activity to support this function was far and away a leader for generative AI job openings at real estate companies. About one in every five jobs that required AI/ML capabilities posted by real estate companies in 2023 was tied to property management.19
Recent advances in chatbot capabilities are an evolution of virtual assistants that not only help with basic tenant relationship management, but also enhance responsiveness and data interconnectivity across real estate business functions such as accounting, maintinance, and sales.20 By utilizing natural language processing technology to better recognize user intent, slang, or grammatical errors, this next generation of chatbots is able to combine that more personalized communication with access to real-time balances for rent payments and maintenance request statuses, among other pertinent client data for inquiries.21 This automation can enhance responsiveness, potentially adding to tenant satisfaction and streamlining the request process.
What to watch out for before going all out with your generative AI strategy
While the promise and transformative capabilities of generative AI use cases may be enticing, implementation of the technology can be a difficult balancing act, with firm-wide strategies around data, operations, and talent typically being the cornerstones for an integration strategy. Specifically, real estate firms should ensure they take into account the following factors:
Data strategy and model validation
“Location, location, location” is no longer the only determinant of strategic advantage in real estate; firms increasingly realize that “accurate, timely, and comprehensive data” holds the key to building a competitive edge. This is especially the case as emerging technologies, including generative AI, revolutionize the way we interact with data. Building an enterprise-owned, differentiated data set and making data-driven decisions can be the distinctive features that may set the firm apart from its competitors.
Foundation models like LLMs are trained on generic information that can be found online. However, real estate use cases will likely require training data to include market-specific, enterprise-specific, and asset-specific information to reduce the risk of hallucinations and bias in the models. However, the lack of publicly available information on leasing, tenant data, or operating performance of individual assets makes it potentially difficult to access timely and quality information at volumes sufficient to train these models.
Before venturing into the generative AI journey, real estate firms should assess the overall AI maturity of the organization’s technical infrastructure and consider whether it currently has access to the quality data required to fine-tune and train models. Those responsible for the transformation should assign leaders to defined roles, including data governance, quality, and ethics. Firms should choose between data governance frameworks to help ensure data trustworthiness, protection, and compliance.
Generative AI relies on foundation models that train on substantial amounts of both structured as well as unstructured and unlabeled data. Generative AI applications can leverage self-supervision techniques, reducing the need for annotation costs.22 The front-end applications and prompts are also more user-friendly and natural language can be used for interactions, making the technology more democratized and accessible across an organization.
Factually incorrect data or outdated information can lead to misleading outputs and result in reputational or financial risk, including legal exposure. Model outcomes can potentially degrade over time if real-world shifting dynamics and patterns are not incorporated into the model, or if the training data is not representative and diverse. Building explainability into models on why and how they come to specific conclusions, validating models on a regular basis, and providing avenues for human feedback into AI models are all important to reduce statistical errors and better understand model predictions.23
Organizational culture
Firms should consider a well-thought-out roadmap with clearly defined goals and milestones for effective generative AI adoption. Management should identify and prioritize high-impact business use cases, size the generative AI value opportunity, and bring employees along the value-creating exploration. Before making a significant investment in any solution or technology, it could be advantageous to first review proofs of concept to help ensure its feasibility and plausibility. Instead of sprinkling use cases across business units, embedding a strategy that weaves across enterprise-wide applications may offer some competitive differentiation.
Firms should also remember that financial KPIs are not the only indicators of success in generative AI technology adoption. Still, nonfinancial metrics such as increased new tenant acquisitions or reduced wait times in property maintenance and tenant satisfaction, cross-selling services, or time saved in payment fulfillment can also be critical indicators for success.
The human influence
Depending on the approach, whether relying on external partners or codeveloping AI solutions in-house, firms should assess the requirements for a skilled workforce, the emergence of new roles within the organization, such as prompt engineers or fine-tuning experts, and jobs made redundant with technology integration. Teams should work as copilots, wherein the humans work alongside the technology.
Risks associated with these models may also call for upskilling and reskilling of emerging roles and teams, including compliance, ethics, and data governance. For instance, a generative AI application deployed for project or construction management will likely require a specialist with domain experience in project management to first curate a reliable database with a diverse information set to help ensure compliance and job safety onsite and to avoid project delays. A generative AI model could also propose building designs that may sound feasible in a virtual space but impractical when considering real-world fabrication or zoning or regulatory compliance, which could have been prevented with the involvement of sector specialists. Keeping humans at the center of AI decision-making can help yield more realistic outcomes and reduce bias or hallucinations.
Disrupt or be disrupted
The industry is at a pivotal juncture when it comes to generative AI adoption. Some real estate firms are likely already beginning their generative AI journey, hiring talented individuals who can spearhead transformation and making investments in companies championing emerging capabilities. Transformation leaders should prioritize working with their vendors or data partners and identify use cases that are core to their businesses and be intentional about where they commit capital and resources. Commitment without a plan could lead to unmitigated business risk in a turbulent business environment for real estate.24 Organizations should stay informed with the latest developments of capabilities while also having clear and strategic goals for how to integrate and leverage the power of generative AI.