The Business Times' Column | What does AI mean for us? A venture capital deal-sourcing case study

This column was first published on The Business Times.

Access the print version here.

AS I prepared for a panel discussion titled “AI: What does it mean for us?”, I quickly realised that this topic can be explored from countless angles. The meaning of artificial intelligence (AI) varies depending on personal, professional, and societal perspectives. Investors, lawyers, or accountants each approach it with different expectations of its potential, while others focus on AI’s actual capabilities.

Sci-fi writer Ted Chiang once referred to ChatGPT as “a blurry JPEG of the Web”, a description rooted in the idea that AI’s understanding is far from perfect. Yet, at the same time, many marvel at the so-called “emergent capabilities” of super large LLMs, or large language models – complex behaviours that arise unexpectedly when models are scaled up.

This wide range of perceptions highlights one key truth: what AI means for us depends on who we are, what we do, and what we expect it to do. Some may see it as an advanced tool for automation, while others fear it could signal a looming threat to humanity’s future.

To truly grasp what AI means, however, it’s not enough to just read about it or play around with ChatGPT. The real value comes when you integrate it into your daily work.

Casting a wider net for deal-sourcing

It started in one of our weekly meetings, where the team was brainstorming ways to expand our deal-sourcing capabilities. Traditionally, most of our leads come through our extensive network, but we wanted to cast a wider net. One idea was to have interns scan job boards for startup activity, filter the companies based on key metrics, and pass qualified leads on to senior team members for further analysis.

As we discussed this approach, it became clear that this was an ideal case to test the power of LLMs. Could AI handle the volume of information and perform a first-pass analysis, mimicking the work of our interns?

Eight months later, we had a system in place that is now embedded in our workflow. It scans thousands of companies from LinkedIn and other sources, analysing their business potential based on indicators such as LinkedIn followers, team size, Web traffic, and app downloads.

The LLM then leverages its general intelligence to perform a holistic evaluation of the startup’s potential, showing a much-desired early-stage investor’s instinct. It now churns out a curated list of leads for our team to evaluate on a weekly basis.

While we haven’t yet found the next unicorn through this system, the results are promising, and the team is excited about its potential. This hands-on experiment taught us valuable lessons about AI’s capabilities – and what it means for us in venture capital (VC).

If you can’t measure it, you can’t manage it

Before diving into AI, you must start with clear goals. Be realistic: we’re a long way from artificial general intelligence. My aim wasn’t to create a system that would replace human analysts, but to build a tool that could handle the tedious, repetitive tasks. Think of it as a way to simulate the work of an entry-level intern – someone combing through vast amounts of data to surface interesting leads.

It’s also critical to define how you’ll measure success. As the saying goes: “If you can’t measure it, you can’t manage it.”

For us, that meant tracking the percentage of leads that our team was interested in following up with, compared to the total number of leads the system surfaced. Early feedback has been encouraging, and as a side benefit, the system now also serves as a centralised data repository that our team can use for quick research on deals.

The main design challenge in creating this AI-driven system was deciding between two approaches: modularised or end-to-end.

In a modular system, you break down the process into smaller steps or modules, refining each one individually to improve accuracy. For example, in a lead-generation system, a modularised approach might involve:

Assigning specific weights to each criterion (for example, company size: 0.3, industry relevance: 0.4, recent growth: 0.2, geographic location: 0.1)

Calculating a score for each potential lead based on these weighted criteria

Ranking leads based on their total scores

This approach allows for fine-tuning of individual components, but may miss complex interactions between criteria.

Alternatively, an end-to-end system treats everything as a black box, allowing the AI to handle the entire process from data ingestion to lead generation without explicitly defined steps.

The advent of LLMs and transformer architectures has made this approach increasingly feasible and effective for complex tasks, due to their versatility, context understanding, and transfer learning capabilities.

We started with an end-to-end system and found the results both promising and surprising. The current AI model exhibited what could be described as “intuitive reasoning”. The current behaviour wasn’t always perfectly logical. However, as the project matured, the team was tempted to explore a modularised approach. One reason for this shift was the realisation that LLMs can struggle with long context prompts and may lose track of precise details.

Interestingly, Tesla faced a similar challenge with its Full Self-Driving system. After years of using a modular approach, Tesla recently switched to an end-to-end design, which yielded breakthrough results. The lesson here is that end-to-end systems may uncover hidden connections in data that modularised designs overlook.

Pushing boundaries

While waiting for the next wave of AI advancements, our team has been pushing the boundaries of prompt engineering. We’ve experimented with techniques such as few-shot learning, where you insert a few examples into the prompt to guide the AI, and chain-of-thought reasoning, which helps the AI walk through a problem step by step. We’ve also explored systems that use multiple AI agents to cross-check results.

So far, none of these techniques has provided a game-changing advantage.

Part of the issue is that we’re asking the AI to tackle ambitious, complex tasks. Venture capitalist as a job is among the most complicated ones, and sometimes we ourselves find it difficult to describe to outsiders our decision process. Maybe a more advanced workflow will bring the system performance to the next level. At the time of writing, OpenAI just released its o1 model, which has built a hidden layer of chain of thought.

At one point, I considered fine-tuning an LLM to create a VC-specific model, especially after observing Tesla’s success with end-to-end training. It would transform the problem from one of system design to one of data collection.

However, we quickly realised that training such a model would require labelling over tens of thousands of leads, while our team collectively analyses only a few thousands of leads per year. For now, the data simply isn’t there.

Moreover, it’s crucial to recognise that VC investment is as much an art as it is a science. The ever-changing nature of startups and markets means that purely data-driven approaches have significant limitations.

Many successful companies, such as Slack (originally a gaming company) or Instagram (initially a location check-in app), had to pivot dramatically before finding their winning formula. These pivots, which often lead to success, are challenging for AI models to predict or evaluate.

A tool for productivity, not replacement

Building a VC deal-sourcing system with AI has deepened our understanding of its current capabilities. In some ways, everyone’s right: LLMs are, indeed, a compressed version of the Internet, capable of basic reasoning that feels more like intuition than critical thought. But they are also far more than just predictive text generators.

This hands-on experience has not only improved our operational capabilities but has also sharpened our perspective as investors. By directly engaging with AI technologies, we gained insights into their practical applications, limitations, and potential. This first-hand knowledge allows us to more accurately assess AI startups, understand the challenges they face, and identify truly innovative solutions.

In VC, deal sourcing is just the beginning. Other tasks – such as investment process management, post-investment value-add, and exit strategies – are even more complex.

While we don’t expect AI to replace humans in these roles anytime soon, it’s clear that AI will be a powerful productivity-enhancing tool in the years ahead. As we continue to explore and implement AI in our own processes, we are not just improving our operations – we’re also refining our investment thesis and strengthening our ability to add value to AI-focused portfolio companies.

*Edited by *Rahul Thayyalamkandy, Director, Vertex Ventures Southeast Asia & India.

For the latest news on Vertex Ventures SE Asia and India and our portfolio companies, follow us on Linkedin or subscribe to our monthly newsletter.