Michael Aldridge (MA): The thirst for data has increased significantly. Say you're an asset owner in 50 or 100 funds. There was a time when performance and exposure data at the fund level was interesting. But now, firms want details on individual assets via, for example, Earnings Before Interest, Taxes, Depreciation and Amortizations (EBITDA) or revenue growth quarter-on-quarter to really understand the drivers of overall portfolio performance.

The way data arrives continues to be unstructured in most cases. Today, I'd love to tell you there are feeds and clean pipes of data traveling between the invested entity, the manager and the investor. But that's simply not the case. More and more PDFs are now in flight, and the problem really is twofold.

Number one is getting the content in the first place – whether that's via investor portals, email or the specific industry vaults from some of the major players. All of these things create an upstream series of frictions, which are essentially difficulties getting your hands on the documents. The second phase of the problem is extracting complex data from those documents and normalizing, structuring, cleaning and preparing it for use in downstream systems.

Susan Doyle (SD): The available data isn’t always consistent. We are an OCIO (Outsourced Chief Investment Officer), so we manage our clients’ private markets allocations. If we invest with a fund manager, we look through data on the underlying investments and provide transparency that is not necessarily disclosed in their quarterly reporting. But for certain types of investments, like fund of funds, you can't get look-through data. You won't get that data on direct deals. You may or may not get it on co-investments. You don't have the same data to work with across the portfolio. So, the biggest challenge really is to find independent sources of data and references outside of what the GP offers you.

Zeynep Meric-Smith (Z M-S): These transactions are bilateral by nature; they're called private for a reason. Two parties come together to determine the value of an underlying asset, which creates challenges while trying to create standardization of definition and understanding. Last year, we worked with a large global asset manager to help their private assets business define, develop and implement a common view of their data taxonomy and definitions across their infrastructure, venture capital, private equity and real estate businesses. Interestingly, even getting agreement of data definitions, such as assets under management (AUM), was challenging because one part of the business defined it based on committed capital, while another considered it based on fee earning AUM. So even getting that commonality and consensus within an organization, let alone across the industry – that's the kind of environment that you're working in.

Chris Liedtke (CL): There are certainly challenges around data heterogeneity and opacity. Collecting and engineering the right data sets for analysis is challenging because the vendor landscape is complex, and integrating public, vendor and internal data is not a trivial task. Smart data procurement and vendor relationships are critical. Data quality is a persistent issue, especially in private real estate, because much of the data is collected from individuals, and is potentially biased. We often discuss using different data sources to triangulate on-the-ground truth, by looking at the world through various data lenses. This allows us to gain perspective while reducing the risks associated with data quality. Furthermore, there is an incredible amount of publicly available data that helps us understand the market in which real estate assets exist, such as the Census in the United States and Eurostat in the European Union.

We are focused on the challenge and opportunity of building the technology, culture and processes around data to make it truly useful in solving our core business problems. This includes developing cloud infrastructure, assembling the right team and talent, and establishing governance in the form of data governance, project management and steering committees. It's about reimagining how we capture, store and structure internal data, then integrating it with external sources to drive efficiency and better serve our clients and the communities where we invest.

Frank Sannella (FS): Timeliness is also a big issue. We use a data service provider for exposure data that we garner from our look-through feeds. The data is provided on a quarter lag, which is common. However, we’ve experienced much longer lags where we've waited up to five or six months to get a quarter’s data.

And when we get the data, we have to do a lot of verification. Even though it's coming from a source that has vetted it, we have to do a secondary level of validation, primarily around corporate actions like spinoffs and name changes. The second level of verification is done by comparing previous quarter data to the current quarter, where we can uncover mismatches in many things, like company names, missing companies, key identifiers, Global Industry Classification Standards (GICS) codes, public/private status, etc., mostly due to corporate actions not being captured and updated in the data provider’s universe.

Someera Khokhar (SK): Looking at these issues of consistency and volume, the starting point for everyone is finding their documents – that in itself is a big challenge. But assuming the documents are stored in a single database, there is a general belief that those documents and the ability to search them enable the creation of data. That is the second challenge in this market, because raw text in a PDF or a word document of a private credit instrument is not data. In order for the words and sentences within private credit documents to become data, complex constructs of those words and sentences need to be organized, classified, and converted into logical structures, formulas and concepts. In the world of private credit, where I work, that encoding is ultimately the much more significant challenge.

If we use a simple example within credit agreements of a default; there are approximately nine ways that the market identifies the concept (defaults, terminations, relevant events, inchoate events, etc.) and that is perhaps one of the simplest of concepts. The disconnect between searchable documents and data is that to capture and codify the embedded data, you need systems and protocols to know what to look for (the “known unknowns”), including the semantic, linguistic and structural variations in what you're looking for, and how to find them. Documents to data is effectively a series of interconnected puzzles of concept, terminology and contextual connections all layered together.

Matthew Clark (MC): I would make an argument that private credit is the hardest, for several reasons. We talk about private credit as a segment, but the reality is that it is made up of a series of different types of credit products from broadly syndicated, to direct lending, to collateralized loan obligations, etc. As an example, a direct lending transaction term sheet can stretch up to 200-300 pages. The bespoke bilateral nature of these credit transactions takes it one step further than infrastructure, equity or real estate, and makes it incredibly challenging to execute and maintain throughout the transaction lifecycle.

MA: It's not just the asset classes that have their own complexities. If you're looking at LP secondaries, what tends to happen is you receive a set of content for funds that you may never have seen before. So, it's not like you have reference data for this fund and all the assets – you're potentially seeing this cold. Maybe it's a 100-fund secondary opportunity, and what you're receiving is maybe two quarters of performance, the previous 12 months, calls and distributions and statements associated with that.

MA: Historically, this has been an almost entirely manual process. So, you, as the manager or the investor, do it yourself with a middle- or back-office team that's logging into portals, pulling down documents, having a spreadsheet open in one window and a PDF in the next, and just transposing numbers from that document into a spreadsheet that you then feed into a system. Or you outsourced it to someone who's essentially doing the same thing on your behalf. The manual approach has really been the one that has persisted and is still here today.

SK: This has been an ecosystem of experienced individuals – who are domain specialists. An experienced professional within project finance, for example, working in power projects has a very different market and practice to another equally qualified project finance professional, in a mining and metals market and practice. Vastly different domain expertise is required in each, albeit based on overlapping frameworks and concepts of rights, obligations and defaults. So, it’s now not just one market of private credit, there are subsets of deep specialism, which do need codification into a technological framework.

CL: Many data challenges have been solved historically with large teams doing manual data analysis in Excel and storing the results in shared drives. Now, we find ourselves in a new era where business leaders are focused on driving superior business outcomes and efficiencies by investing more in technology. This changing landscape has us investing more in cloud infrastructure, software engineering and data science, and moving away from manual analysis and on-premises data solutions.

Further, cloud infrastructure enables entirely new patterns where analyses are dynamic, automated in Python (a programming language) and available for consumption at scale via dashboards and chatbots. Large amounts of graphics processing unit (GPU) computing resources are available on-demand without the need to buy costly hardware. This migration to cloud-based solutions has allowed for greater scalability and flexibility in data processing, storage and engineering, while strengthening security. It speeds up our decision-making cycle and enables our investment teams to understand the world in real time and respond strategically on behalf of our investors.

MC: Asset managers have typically thrown bodies at the problem by hiring very smart, capable people who work incredible hours to solve these complex problems. But over time, as the market continues to attract more LPs and GPs, you see more margin compression and a push for efficiency. A primary area for efficiency is data and having a common data taxonomy and definition, supported by an increasingly consolidated technology ecosystem. Early in my career, I supported a lot of swaps and at the time derivatives were extremely customized bilateral transactions, like private markets at present. This is back in the Ice Age, of course, but now derivatives are very common, and commoditized with an efficient data and technology ecosystem. I think we’ll see the same happen with private markets over a period of time, as the market starts to agree on commonality.

SD: Private markets have become more transparent over the years. If you go back to the start of the REIT (Real Estate Investment Trust) and the CMBS (Commercial Mortgage-Backed Securities) markets, that’s where a lot of research into the real estate markets was made publicly available for the first time. So, over many years they became more transparent. Now, people want this data to be as transparent as daily, in terms of valuation.

MC: I do believe there is a role for GenAI and machine learning, absolutely, but the timeline for having a material impact on private markets is unclear. I don't think it's tomorrow but may be in the next five years and the resulting impact will likely be quick with a lot of upside. I don't think anybody can honestly look at what's happening in AI, and the capabilities of AI, and say it's not going to impact private markets in a material and meaningful way.

CL: BGO was early in deploying machine learning tools to inform our investment decision-making process. That initial success spurred a flurry of investment in talent, tools and technology to further develop our investment capabilities. It also had the unintended benefit of priming us to take advantage of the explosive levels of AI progress over the last few years. We see AI touching virtually every corner of our operations. GenAI allows our engineers to code faster, increasing productivity and reducing the time it takes us to build solutions. It enhances our investment and research teams' ability to analyze vast amounts of unstructured data in a fraction of the time it would normally take.

Paradoxically, while it has increased productivity, the advances in AI are coming so quickly that my team spends more time in research mode to ensure we can drive the best outcomes using the latest technology, while reducing the risk of technology obsolescence. While I’m biased, I see GenAI having an outsized impact within private markets because so much of our data is unstructured – it lives in PowerPoint, Excel and text documents – precisely where GenAI is the strongest. GenAI enables us to mine our existing internal data – which wasn't necessarily searchable, or was at least much more difficult to extract insights from without a GenAI system.

MA: They have created real opportunity in terms of the problem set we're talking about here. They are, however, just a tool. If you're a large asset owner that may be in hundreds or thousands of funds, or even more, people aren't looking to chat with that many documents. What they want is to be able to push that volume of content through a secure, robust, scalable platform that can guarantee you a high quality of structured data. Now, inside of those platforms, GenAI and LLMs can be incredibly powerful as tools for data science-led solutions. But these things need a front end. There has to be a user experience that wraps around these things.

SD: I think that GenAI and LLMs can help with our providers. They send us all of this information and we check, and, when things don't look right, we circle back. Now the process of validation and verification checks are also automated for us, and our system puts it into an exception report, and that's how we validate this information. But can they do some of this themselves before they send the data? With these technologies, our providers could figure out that things have changed from the prior report they sent.

FS: I'll add to that. We shouldn't have to iteratively download a portfolio’s quarterly data set until we have a complete feed. This repetitive operating rhythm uses up our time and resources. We would prefer the data service provider publish the data feed once they've totally completed the portfolio before making it available to us. Currently, incomplete portfolio data is made available where the data service provider may not have all the look-throughs verified and completed for the portfolio. As a result, we're subjected to an iterative download process. This process needs improving, becoming cleaner and more timely. These new technologies could be a tool to achieve that.

Z M-S: One use case that my clients are leveraging AI for right now is going back to those credit agreements that are 200 to 300 pages long and feeding those unstructured data sets into an AI model to pull out the key data points. These are fed into a valuation mode or into your analysis, recognizing that those agreements may vary each time in terms of layout and structure and format.

SD: And oftentimes it's what's missing from the documents. There are thousands, perhaps hundreds of thousands of pages per investment. So yes, some of it, in fact a lot of it, is boilerplate. But some of it is not. So are there LLMs that could come in and see, not just what stands out, but what is missing?

MA: Gen AI can solve a number of things. It can help with data extraction from unstructured content. But can they then normalize that where, for example, you've got to rename “total capital committed,” which is called 57 different things across my 1,000 funds, into the same nomenclature? So, normalization and structuring are additional challenges. Driving insights from that data is an area where GenAI can really play a role.

SK: LLMs in their current forms are not experts in private credit constructs – they are not “anchored” to specific domain expertise, like the human specialists I talked about earlier, but they have transformative potential and significant opportunity for early adopters. The key is formulating and implementing the technical strategy to create domain-specific AI agents with domain-specific skills, knowledge and “reasoning.”

The first step is the algorithmic structures that unpack the underlying data constructs and codify the raw text. For example, the two words together of 'maturity date' in private credit have a very clear meaning. The two words separately, don't mean that at all. Connecting the concept of a maturity date to an underlying facility or terminology (which could be one of many in a single instrument or document) is a formula. Private credit is effectively a series of these semantic formulas, based upon unstructured data elements actually embedded in text.

Once documents are converted into a series of these data elements, which are then packaged into formulas, the elements and the formulas can be tokenized. This enables multiple layers of value to be created – for example, controlled, permissioned access into specific data elements within vast data sets, or the ability to create new data-driven instruments or algorithmic synthetics of a credit instrument or a component of it, all which can be monetized to create new value.

MC: One example that someone cited to me is to use AI in the investment analysis and decision process. So, in your investment committee, when you vote to make a decision on whether to green light an investment or not, you have AI be part of that process. Feed the investment thesis into an AI that should have all of the documents on the investments and ask questions like:

• "Does this look like a good investment for us?"
• “What risks are we not considering?”
• “What additional questions should we be asking?”

SK: An opinion-driven sentiment LLM, right? So, tell me: "Is this the best deal I've got?" We can tell the model the parameters of what “best” means. But your appetite for risk and your ability to make that judgment call comes down to a couple of things. One, how your institution judges investments, based on its strategy, and risk appetite, etc. And then, how do you take into account what's happening at the point in time that you're making the decision. If, for example, you've got a basket with one borrower, and that borrower's credit rating has suddenly dropped, you might want to rethink.

MA: LLMs are new and exciting toys, but how to really leverage and deploy them is the question. And, of course, there are issues around where the data is going. A publicly available LLM looks like you can do lots of cool things, but given the nature of private market investing, well, the clue is in the name! And people are absolutely right to have concerns around data security. “Is my data being used to train models that may benefit my competitors?”

SK: You've also got data that has historically lived in silos because of security. The ability to manage data privacy and confidentiality has, up until quite recently, been challenging, because the data lives in a document. Now, not every piece of the data in that document is going to be confidential. But the information cannot be detached from all the documents – not at speed and scale and with understanding of the data constructs I talked about before – so you've created these silos where you've just put arms around big chunks of documents, and big chunks of data, because there's no other way of siphoning between them.

Now you have technological developments where data can be detached from the underlying instruments and can be actually stored separately. So, if you think about blockchain technologies and tokenized debt instruments, you can take a data element, and you can attach it into a token, and within that token, you can have various levels of permissioning. So, for example, let's say it's my personal data: I can see my date of birth, my full name, my social security number; but perhaps you're only allowed to see my name. However, that bifurcation of permissioned views has, up until quite recently, not been possible. So, now you're starting to emerge from these silos and recognize that within these instruments there are certain components that have value to certain people, and you can actually permission it accordingly.