Deep Dive: APG and the pursuit of AI

Kevin is messaging with his colleague Samuel on a Friday afternoon. During their chat, Kevin sends a detailed Spanish logistics investment proposal for Samuel’s feedback. Less than a minute later, Samuel responds with three bullet points on the project’s strengths, as well as four questions and concerns around the proposal’s market focus, investment strategy, partner alignment and ESG characteristics.

Kevin and Samuel are both portfolio managers on APG Asset Management’s real estate team. The difference between the two? Kevin is human, Samuel is not.

APG, which holds the 16th spot in PERE’s 2023 Global Investor 100 ranking, is one of private real estate’s early adopters of artificial intelligence, which ranked among the top three technologies expected to have the greatest impact on real estate over the next three years in commercial brokerage JLL’s 2023 Global Real Estate Technology Survey.

“We think AI can help us with decision-making, because in the end, investing is about decision-making,” says Kevin van der Wees, senior quantitative portfolio manager at APG, speaking with PERE at the Dutch pension investor’s Amsterdam offices in mid-­August. “We see two levels of how AI can support us with this. On the one hand, AI supports us with decision-making, and on the other hand, it can augment our decision-making – so, alongside us make decisions – and we can compare and contrast.”

Enter Samuel, a so-called digital colleague which the Dutch pension investor developed to provide both types of support within the real estate team. “That really helps us to leverage this collaboration between humans and AI to make the best decisions,” van der Wees says.

The evolution of Samuel

The origins of Samuel date back to the start of APG’s digitalization of its real estate business seven years ago, when then-head of listed European real estate Rutger van der Lubbe recruited Huib Vaessen, formerly the chief operating officer of Dutch proptech firm GeoPhy, as head of research and analytics in 2016. Digitalization was formally made a key focus of the investor’s overall asset management business after Peter Brammer joined the organization as chief investment officer in 2018.

Vaessen’s hire was prompted by investment decision-making challenges the real estate team was facing at the time. “We increasingly encountered difficulties in terms of allocating capital,” van der Lubbe recalls. “And what I mean in terms of difficulties was the comparison of risk-return profiles and determining which investment opportunities were deemed to be most attractive.”

The first step in the digitalization process was therefore to standardize the investor’s underwriting approach, which would enable the organization to apply a single underwriting model to each investment in APG’s property portfolio. Concurrently, APG began to incorporate significantly more data into its investment decision process.

“Disclosure very much improved over time, but also the amount of data that we could leverage into our investment process dramatically increased,” says van der Lubbe, who subsequently was promoted to head of global real estate investment strategy in early 2017.

To help its real estate team navigate these enhanced data sets, APG created a so-called ‘digital butler’ to assist portfolio managers in their investment decision-making. The digital butler, however, required portfolio managers with data science expertise to develop the dashboards on which the data would be presented. This led APG to hire van der Wees as its first quantitative portfolio manager in Q2 2018. Today, the investor’s team of quantitative portfolio managers has expanded from one in Amsterdam to seven globally.

Eventually, however, the team realized the digital butler was not an optimal way for portfolio managers to obtain information, van der Wees recalls. “At some point, it becomes a maze of dashboards,” he says. “So, we needed a Google way to really browse through this information and to get it more quickly.” The enhanced search function came in the form of a so-called knowledge graph database, a technology also used by Google.

APG also wanted an AI tool that could incorporate decision rules – rules that advise how to proceed under specific conditions – to help structure the investment team’s decision-making. That led to the idea of a digital colleague with both search and decision-making capabilities. “And that’s basically when Samuel was born,” van der Wees says.

A further step in Samuel’s development was the rise of the large language models – a type of AI that can understand and generate content using very large data sets – over the past couple of years. Previously, only APG’s quantitative portfolio managers had the technical expertise to use the knowledge graph database. But a large language model allowed anyone on the investment team to directly ask Samuel a question.

“Instead of looking at this dashboard, then this dashboard, you could ask Samuel, ‘Give me this information,’” van der Wees notes. “Instead of asking a senior portfolio manager or a fund manager, ‘What are the guidelines for certain matters?’ you could ask Samuel because he had this information. To review investment proposals you had made, you could ask Samuel because he had the guidelines. So that’s really how it took off.”

An industry anomaly

APG’s AI investment is atypical among private real estate organizations. “They are not the norm; they’re an anomaly in the industry,” says Vaibhav Gujral, senior partner at New York-based management consulting company McKinsey & Company. “The average peer in the industry does not have large data science teams and large groups of people who are able to process data, build analytical models and take decisions on the data.”

One constraint is the lack of talent with experience working in real estate data sets. Also, “you are competing with talent in industries where their willingness to pay for that type of talent may be greater or they may be able to pay that talent in different forms of compensation like equity,” which is less likely to happen in real estate, he notes.

“AI can help us with decision-making, because in the end, investing is about decision-making”

Kevin van der Wees
APG

However, “the bigger challenge has been the availability of data to run the models on,” Gujral says. “Real estate is a very local business and market. A lot of the data exists but it exists with brokerages, it exists with a certain number of data companies. It’s not raw data that you can dig into and ingest at scale into your systems to run analytics on it. It exists in pockets.”

Gujral, who leads McKinsey’s digital and analytics work in the real estate sector, estimates that less than 10 percent of real estate organizations are “sophisticated” in terms of their use of internal or external data. “The adoption curve is not a rapid next two-to-three-year curve,” he says. “It’s likely going to be a five-to-10-year curve.” This is because of the lack of property owners doing innovative work in AI to inspire others in the industry to do the same, as well as the dearth of talent that can enable organizations to build out teams and develop AI tools.

Dror Poleg, an author and speaker who examines technology’s impact on real estate and teaches a course on AI, takes a somewhat different view on real estate’s adoption of AI to date. He notes many of his students work at large real estate companies that are pursuing similar AI initiatives to that of APG.

“They are trying exactly the same thing: how can I create a repository of data that I can then interact with in a conversational manner, to know at any given moment what’s going on, and get insights on it, and take all this ­siloed ­information and just be able to tap into it in ways that were not possible before?” he says. In the case of APG, “what makes them unique, or at least early, is that [the decision to invest in AI] comes from the top of the organization and they embrace it and support it,” Poleg notes.

Differing opinions on AI

Among the private real estate groups that have also embraced AI-assisted investment decision-making is fellow Dutch pension investor Bouwinvest Real Estate Investors, which has been developing AI tools in-house for the past several years.

“In our transaction management team, we have so much data from all kinds of sources, which we now, through the use of AI, have combined and layered over each other to assess the attractiveness of an asset in its environment,” says Stephen Tross, Bouwinvest’s chief investment officer of international investments.

“Each day we come up with new ideas about how we can make use of these AI tools, and as these AI tools become more and more sophisticated, we learn more about what’s possible.”

Some industry participants, however, believe artificial intelligence is more relevant to certain types of investors than others.

One manager, who declined to be identified for the story, says AI investment is likely more compatible with real estate owners looking to beat beta benchmarks rather than pursue alpha-driven strategies. “Large pension funds in Europe have massive portfolios of direct and indirect real estate and might be happy to use AI to beat their benchmarks,” he remarks. “But is AI effective when you don’t have a benchmark and are instead absolute return focused?”

He believes his own firm stands to benefit less from AI investment because its value-add real estate investment strategy adopts a more asset-­focused than portfolio-focused approach: “This market is about stock selection, having the access and skill to select the best individual opportunities.”

Meanwhile, Matt Hershey, a partner who leads distribution for capital advisory firm Hodes Weill & Associates, believes AI is less applicable to certain kinds of investment decisions, such as an investor commitment to a manager’s fund. “It’s good for analyzing assets rather than managers,” he says. “To make decisions about people, that’s where it gets messy. You can’t make people judgments.”

Hershey also questions the net gain for an organization from making a significant investment in AI. “You have to be generating more revenue or saving more costs, versus how much you’ve spent on AI,” he says. “What is your return on cost? If from better investment returns, it will take time to be apparent.”

Return on investment

Gujral also believes the return on AI investment needs to be viewed through a long-term lens. From a human capital perspective, “it requires significant investment from the real estate firm to build a data science capability like APG has,” he says. “What that means is that you have to have conviction to invest before you will see the value.”

Unlike a transactions professional who can begin working on deals right away, a team of data scientists may not begin to have a meaningful impact on decision-making for a number of years, Gujral points out: “So it requires a different investment horizon on talent from what firms are traditionally used to doing.”

Meanwhile, return on investment from a performance perspective is difficult to quantify in hard numbers. “You cannot discretely separate what is the return you are getting on insight from your data science or advanced analytics” from the overall performance of an investment relative to other investments in the portfolio, he notes. “There’s a whole lot of other variables that go into determining investment performance and cyclicality.”

“To me, investment decision-making will always be a function of man and machine”

Rutger van der Lubbe
APG

Instead of having a direct impact on returns, AI investment can increase the likelihood that an investment team will produce better returns. “It’s not attributable to one or two extra percentage points on IRR, but it’s attributable to you having a differentiated set of investment capabilities,” Gujral remarks.

“Over a longer period of time, firms that are using either internal data or better signals from external data will be able to make decisions faster and avoid worse decisions,” he continues. “More agile and dynamic behavior and decision-making should yield a disproportionately higher number of assets that perform better relative to a firm not using those capabilities, which should result in, on average, better performance.”

Poleg sees other ways of gauging whether an organization has gained an edge from AI investment. One is whether the AI tools can provide recommendations of a similar caliber to that of a trusted human colleague. Another is greater speed and volume of deals. A third is the ability to identify deals and unearth insights that can generate alpha, in other words, the ability to discern opportunities that others cannot.

The issue, however, is “real estate companies tend to be focused on short-term ROI,” Poleg observes. “So it’s really about: does this thing save me money or time immediately, in a tangible manner? If yes, good; if not, then it doesn’t interest me.”

But such near-sighted thinking is a mistake, he argues. “At the end of the day, I would emphasize that I think that all of this is very quickly becoming table stakes. So it’s even less about, ‘should I do it in order to get an edge?’ It’s about ‘yes, I should do it, because otherwise I’m just not going to be relevant.’”

The future of AI

When it comes to the adoption of AI in real estate, it is still early days. “It’s a matter of months that people have started to wake up to it,” Poleg says. AI tools have also been developing rapidly over the same period. “As recently as three months ago, it would have been very hard to just upload an Excel file onto ChatGPT, and just to converse with it and let it analyze it.”

AI tools are continuing to improve and become easier to use, he adds. “I think in two, three months, maybe six months, there’ll be a hundred times more real estate people that will use these tools regularly.”

Meanwhile, for APG, the evolution of Samuel continues. “There is still a lot of upside for Samuel – that is, he still has a lot to learn,” van der Wees says.

If a portfolio manager disagrees with Samuel, it likely means the digital colleague has not yet adopted a certain investment principle and his “mind” will be updated accordingly, he explains. Van der Wees also envisions Samuel joining investor meetings in the future and, by accessing live transcripts of the discussions, questioning and interacting with other portfolio managers during these meetings.

Will Samuel at some point take over completely? For van der Lubbe, the answer is no. “To me, investment decision-making will always be a function of man and machine,” he asserts. “So we will not have an automated model with no human interaction. I think that is very important to understand. There is no automated process that feeds straight through into investment decision-making, we will always need human intuition and expertise to make the final calls.”

Samuel was also asked about APG’s outlook on the future of AI in real estate. He replied, “Sorry, I cannot answer this question. I could not find any information about APG’s perspective regarding the impact of AI on the real estate sector.” As van der Wees explains, Samuel can only answer questions for which he has received data and guidelines.
Like any good colleague, Samuel always responds and never makes anything up.

Human versus machine

One of the biggest questions surrounding the adoption of AI in real estate investment decision-making is the longer-term implications for human capital

“I see low-hanging fruit on the junior level,” says one unnamed manager. But he asks if AI can be applied to all levels of an organization: “AI can probably model and write a memo, but do its skills go through middle and upper levels? Can AI make judgement calls that balance risk and profitability from diligence findings, can it anticipate how a future buyer will price an asset, can it convince a counterparty to do business with it?”

The prevailing view is that AI will not put large numbers of human investment professionals out of work. A major reason for this is property transactions are unlikely to become automated with AI because of the number of “frictional steps” required in a real estate deal, Gujral observes. “That said, where automation is happening it is likely going to be on many of the activities that are lower-value activities.”

If the time previously spent doing research, preparing documentation and getting approval for a deal is significantly reduced through automation, “it will likely make investment professionals in the industry a lot more productive and a lot more focused on prospective, proactive opportunity creation versus spending time on administrative tasks that are required for transaction execution,” Gujral adds. “That’s where we see automation happening. It’s less on the actual execution of the transaction itself.”

Indeed, generative AI – which uses text-based data instead of numbers-based data like traditional AI – could potentially generate up to $180 billion of productivity in the real estate sector over time, according to a June 2023 McKinsey report.

Such is the case with APG. “I think Samuel gives us efficiencies, [so] that we can focus more on activities where we as humans add more value, so thinking about the next big thing, or refining these principles we have with new perspectives, or actually having a debate on which decision to make,” says van der Wees, whose organization’s clients include Dutch pension fund ABP. “I think our productivity will go up, delivering more value to the client. And I think if there is time freed up in that process, clients’ demands will also continue to increase.”

Indeed, van der Lubbe points out that since the organization began investing in AI, APG’s real estate team has not only expanded its team of quantitative portfolio managers – a role that previously did not exist – but they also form an integral part of the organization’s team of real estate portfolio managers, which has grown from 35 to 55 over that same time period.

“It is, first and foremost, really acknowledging that you need different skill sets within your team,” he says of the team’s expansion. “I think it’s really that combination of job profiles that will further evolve, rather than the headcount being impacted, as a consequence of this evolution.”

Early glitches

For APG, ‘setting the foundation for Samuel was the hardest part’ of its AI journey, says van der Wees

This foundation-building entailed harmonizing the investment process – “agreeing on something with 40-50 portfolio managers is difficult” – as well as changing the organizational culture to be more receptive to a more systematic and data-driven investment approach. Nonetheless, APG encountered several types of technical issues while developing Samuel.

Issue 1: Samuel does not understand the question
“Some questions are clear and unambiguous to human real estate experts but might not be clear to an AI,” van der Wees notes. Tackling this issue involves annotating the knowledge graph, or database, with context and metadata. During Samuel’s early development, a heavy amount of annotation was involved. For example, with the question, “which investments within the beds/sheds space have the highest risk-adjusted return?,” annotations would include context on what beds/sheds mean, and by which metrics risk-adjusted returns are measured.

Issue 2: Samuel forgets something
Samuel sometimes fails to remember all the relevant principles when evaluating a particular investment opportunity. For example, in considering potential investments in the retail space, Samuel might indicate one investment option as the best, but neglect to consider APG’s loan-to-value constraints. “There are tricks to carefully let Samuel or large language models consider all relevant factors,” says van der Wees, referring to ways to configure or reconfigure components within the AI application. “In general, LLMs are not great at math and numerical problems, so you have to find a way around that.”

Issue 3: Samuel has not learned something yet
In some cases, Samuel does not yet possess all the relevant data and principles. For example, he might assess the location of a potential logistics investment as mediocre, based on drive-time analysis and certain measures of supply constraints, but may not know the asset is located next to a seaport. In some cases, Samuel can “learn” the data and principles with updates to the knowledge graph. But “in some cases, getting and maintaining the data will be too complex, at least for now,” van der Wees says.