Welcome to the first article of our new series, Products & Markets. As regular readers of Capital AI will know, we have written extensively about when it makes sense for investors to adopt AI, what the top use cases are and how to navigate the landscape. Many have already thought about where to implement AI within the investment cycle, and what kind of results they would like to see. Still, lots of fund managers don’t know where to start.

The product landscape in AI is vast, spanning everything from LLM providers to point solutions. The markets for these products are still being defined as solutions take off, others fail and customers provide feedback. We’ll share some of that feedback and customer insight, so that you can learn from their experience. In talking to dozens of readers involved in AI and investment decision-making over the past few weeks, the most salient questions and challenges we came across all started with:

“OK, we have a sense of what we want to do. But, we don’t have any AI expertise on the team...”

The two most common follow-up questions were:

1. “Should we go big and roll something out across the firm, or focus narrowly on addressing a single use case”. This might seem a bit unintuitive to managers who haven’t thought widely about gains from AI, but leaders ask this question because the benefits of AI often fall into two paradoxical categories: either, it’s a general technology that can up-level the whole firm’s capabilities, or its value comes from solving specific challenges in key bottlenecks.
2. “Should we build our own solution in-house, or should we buy something off-the-shelf?” Most often, managers are weighing the benefits of developing proprietary strategic assets (and the high cost thereof) versus moving quickly and trusting with more widely known solutions. Some are interested in a middle ground such as partnering with a solution provider, or working with experts to modify open source offerings.

Plotting these questions on a quadrant yields a helpful heuristic tool. AI is not going to be a one-shot decision for firms. Instead, we view it similarly to a general technology such as computing: an evolving, fundamental technology that requires expertise and repeated investment for decades to come. Thus, each use case and AI-related operations decision will require both technical and managerial knowhow. While we’re still in the early-stage fog of war, it can help to bucket strategic decisions into useful starting points.

Those starting points are what we’ll discuss in this article. No single recommendation is right for every firm, but by imparting our knowledge about what product developers and capital markets are up to, we hope to get you as close as possible. To address each question, we’re going to use AI products as a lens, since this is a product-oriented series. When most people here “AI” they either think chatbots or LLMs, so those make an apt place to start.

LLMs and Apps: Broad vs. Narrow Implementations

Again, some readers might wonder why “broad versus narrow” is an interesting question – shouldn’t investors simply adopt AI where it makes sense for their job? Once you dig in, however, you realize the question of where AI adds the most value to an organization is less clear. For certain investors, e.g. those at large firms, it might feel natural to say that adopting chatbots as research tools across the organization feels like a no-brainer to enhance broad-based productivity. For others, it’s much more intuitive to solve for a single use case.

The most advanced generalist AI models, like GPT-4o, Gemini 1.5, and Claude 3.5 Opus, are reshaping how PE professionals approach investing tasks with their broad capabilities. In some cases, products like LLMs, which feel “general” can actually perform best in specific tasks too. These models can process vast amounts of text, assist in creative idea generation, and offer sophisticated writing assistance, making them a potential co-pilot for anything from comprehensive due diligence and market analysis.

Consider BloombergGPT, a finance-specific LLM developed by leveraging Bloomberg’s extensive datasets. Despite its specialized training, GPT-4 outperformed BloombergGPT in nearly all financial tasks, illustrating a key trend: larger, more generalist models often surpass specialized models in their own territories.

This doesn’t negate the value of specialized models but highlights a strategy shift. For private equity firms, the utility of large-scale models cannot be ignored, even when sales calls from specific point applications beckon. LLMs provide a foundational toolset capable of adapting to diverse investment scenarios. However, while this concerns LLMs, most PE firms should consider whether training or buying an LLM into a custom workflow is the right approach at all.

Build In-House or Buy Off-The-Shelf?

While general models can be more powerful than specialized ones, LLMs are still only the building blocks of larger solutions that address specific use cases. In selecting those building blocks, some top firms are opting to build their own specialized models. In most cases, when we talk about these “in-house approaches” in finance or consulting, this means fine-tuning or re-training existing models like GPT-4 for more specific use cases. Take Morgan Stanley, which is perhaps the financial company that has most broadly deployed a proprietary model.

In many such cases, firms are grasping at immediate ways to implement adjacent solutions like chatbots and lead generators. These approaches see AI as a mere tool for cost-cutting rather than an enabler of transformative business processes. By prioritizing immediate efficiency gains, firms risk missing the broader potential of AI to redefine their operational and strategic horizons. This involves a strategic shift from using AI to simply 'do things faster' to 'doing things differently' and better. For example, AI can enhance decision-making with predictive analytics, uncover new investment opportunities through advanced data analysis, and create more personalized experiences across portfolio companies.

While re-training and fine-tuning actually offers perhaps the largest potential for investing in long-term strategic advantages, we see many similar examples where investors are tempted to deploy efficiency improvements rather than using AI to develop new capabilities.

For companies that do want quick fixes, fast experimentation and point solutions, the vast landscape of off-the-shelf applications is perhaps a better place to start. In the next section, we’ll talk about where firms should begin depending on their strategic priorities.

Recommendations for Strategic Advantage

When you want to buy a firm-wide solution: Decentralize

Another unexpected argument for broad based adoption is that unlocking the full potential of AI within the private equity sector fundamentally requires placing this technology into the hands of those with deep domain expertise — in our case, PE Vice Presidents, Associates and Analysts. Their day-to-day experience provides the nuanced understanding necessary to critically evaluate and effectively deploy AI-driven solutions in ways that are both strategically advantageous and operationally sound.

Consider the scenario where an AI is used to evaluate complex financial models or to act as a copilot in financial analysis techniques. Investors on the actual deal team are uniquely qualified to evaluate the utility and precision of these AI tools, compared to managers who may be further removed. By integrating more structured GPTs, RAG-based systems or (at a minimum) prompt-engineered frontier models for strategic problem solving, deal teams on the front line can refine and customize AI outputs to better fit specific investment strategies or operational needs.

One example is a custom GPT instance, which is relatively inexpensive compared to the cost of using ChatGPT pro. One upper market PE firm we spoke with gave every investment professional on their team budget to access such a tool trained on finance-specific knowledge, and incentivized individuals to come up with value-adding uses by paying a bonus to investor who identified the most effective use case across the firm. They then scaled this practice - developing questions based on CIMs - across all deal teams.

Thinking of buying a point solution? Consider a Partnership

Starting your foray into AI with the purchase of an application designed to address a narrow use case might sound devoid of any strategic advantage. But one approach we’ve seen work well, especially for smaller firms, is to partner with a provider in a more collaborative way to influence early AI roadmaps or get a jump on implementation before competing investors do.

Startups within the AI landscape are often willing to customize implementations for investors looking to get into the game early, and the ability to shape a product to the needs of a specific investment strategy is often both more valuable and more cost effective than training a niche model which may not even be more effective than something off the shelf.

One example is the approach many of our customers at Keye take. By working directly with our founding team to understand how AI-powered due diligence can best fit into their firm’s existing workflows, lower middle market firms see minimal disruption to their existing practices, while leap-frogging some of their upmarket competitors in terms of deal analysis capabilities.

Building a point solution in house? Test & Iterate

Some investment firms we spoke with identified use cases for AI that either were not covered by the market of external providers, or offered some kind of strategic advantage by building it with firm-specific data and software. Because building a tool in house can be costly and outside the core capabilities of financial firms, we recommend an approach of rapid prototyping: building a minimum viable product with low overhead (such as by using existing technical resources), and iterating quickly based on the feedback of internal teams.

This approach simulates the “product-market fit” and feasibility analysis process that many external providers go through to bring their products to customers, and safeguards against a relative lack of AI expertise at financial firms. Once validation has been achieved among internal stakeholders, investment can always be scaled up to improve the tool’s capabilities.

One such use case is building a tool to improve deal sourcing by leveraging ML and GenAI tools to generate leads using past success cases. This type of solution lends itself highly to internal capabilities and knowledge, as successful firms likely want to build on past strategies, while stripping out the human error and messiness of sourcing at scale.

Building a more generalized tool? Be ready to invest more upfront

While this is perhaps the riskiest strategy, larger firms will likely be able to generate significant advantage by leveraging scale and capital reserves to go big on a proprietary, firm-wide solution. Examples include firm knowledge bases and co-pilots like McKinsey’s Lilli or EQT’s Motherbrain. These solutions offer a trifold advantage, because they generate the obvious broad based productivity enhancements promised by AI, while also addressing firm-specific use cases, and potentially acting as monetizeable white-labeled service offerings to outside customers.

While the news is dominated by firms who are ready to train and build custom solutions in-house, this only makes real sense if you have clearly defined your need for GenAI, and have capital and proprietary data to invest in its development.

The questions of broad vs. narrow and in-house vs. external providers do not address every aspect of getting started with AI, but they do present a useful quadrant to help financial firms interested in the technology to identify a starting point beyond LLMs. In our next Products & Markets article, we will address how this applies to other parts of the stack, and examine specific products making rapid AI adoption possible.

Have questions or want to get in touch? Reach out to us at founders@keye.co