What I learned about venture capital from building AI tools

On paper, startup valuation looks like an algorithm. Evaluate the question statement. Measure the size of the market. Map your competitors. Check traction. Introducing the founder’s profile. Run the numbers. This is a structured, repeatable process that seems ideal for automation.

So when I was first tasked with building an artificial intelligence (AI)-powered deal sourcing tool for a fund, I expected the most difficult part to be the engineering.

I was wrong.

Through months of building and testing, I’ve discovered that venture capitalists resist automation not because investors are technophobes protecting their turf, but rather because the job is much more nuanced than it appears from the outside.

Here’s what I learned about the gap between what AI promises and what jobs actually require.

data problem

Let’s start with the input. In venture capital, data is sparse, private, fragmented, and often outdated. The last funding round recorded in our database may have been 18 months ago. You can see how the company fared at the end of the previous year by looking at its most recent Accounting and Corporate Regulatory Authority (Acra) filings.

These data points serve as a snapshot of where the company has been, rather than an indication of where the company is going. Real-time momentum is reflected in founder conversations and reference calls, not external databases.

This is a problem that quantitative hedge funds are familiar with. Many sophisticated companies have had to create their own leading indicators from all available public market data.

Ventures have less information to work with, more qualitative variables, and horribly slow feedback loops. In public markets, bad investment thesis can be punished in weeks, but in VC it ​​can take years.

Then there’s the problem of outliers. The investments that define a fund’s vintage are, almost by definition, investments that no model flags. The beginning of Bitcoin, the acceleration of Covid, the moment of GPT. Black Swans do not backtest.

A model cannot be trained to identify the next paradigm shift from the last paradigm shift in the data. And given that the industry primarily follows the power law, experts expect this handful of elusive investments to generate returns for the fund as a whole.

Perhaps most importantly, the most important signals in the early stages of investing are often not quantitative at all. It’s the story our founders tell about why this problem occurs, why we need this solution, why now, and why we do it.

A compelling story that reshapes the way you view the market is more valuable than any spreadsheet. Context trumps data. Stories move markets.

subjectivity problem

Investment decisions are deeply personal, a skill developed through years of pattern recognition and hard lessons, rather than the consistent application of a rubric. The same transaction can be a pass for one partner and a conviction for another, and both may be right.

When building the tool, we had to make fundamental decisions. It’s about how much you need to hard-code your company’s standards, or how much you trust the general intelligence of the underlying model.

If you hardcode it too much, your strict filters may miss things that don’t fit into familiar patterns. This may be exactly the kind of contrarian bet that defines the best VC returns. If a model leans too heavily toward black-box inference, explainability can be lost.

Another problem is that the dataset is small. How often does the fund actually trade? Probably dozens of times over several years. This is not a training set, just a sample. Old notes may also be locked into paradigms that no longer apply. For example, a paper built on assumptions from the Software-as-a-Service (SaaS) era may not be directly relevant to evaluating AI-native companies.

relationship moat

The data and subjectivity problems can be solved theoretically with good engineering, but the relational problems cannot.

There is a contradiction here. If your AI can see your transactions, so can everyone else’s. The information edge disappears the moment it becomes systemized.

The best deals in ventures don’t surface through data; they communicate through trust. The founders choose which investors to call before the round goes public. Co-investors secure a seat at the table for funds with truly profitable track records. Reference checking is only useful if you have a relationship that allows for honest feedback.

These are not inefficiencies waiting to be destroyed. These are characteristics of markets built on asymmetric information, where access is gained and reputation is currency.

Relationships give access. And in this industry, access is a big part of the job.

Where AI can actually win

None of this means that AI has no place in venture companies. In reality, some restructuring of workflows has already begun, and companies that pretend otherwise will be equally left behind.

AI can scan, flag, and process opportunities very efficiently. One obvious application is to qualify inbound leads. This reduces the chance of interesting companies being left out simply because no one had time to look them up.

Surrounding workflows are also being transformed, including a data room agent that summarizes and surfaces important information, deep research tools that compress days or hours of market analysis, and an investment memo agent built on the firm’s in-house style that helps automate Excel, PowerPoint, and first cuts, meeting transcription, and outreach.

These are not easy profits. These free up human bandwidth for the most important parts of the job.

So what happens to us?

Since I first started this project, AI has already moved beyond the co-pilot stage. Some of your workflows will not only be enhanced, they will also be automated. As frontier models continue to reach new heights, they may eventually creep into higher stages of judgment.

Maybe that’s okay. Maybe that’s a good thing.

As leaders at my company always remind us, investing is the first and easiest part of the job. Everything after that is the real work: board work, difficult conversations with founders, down-round rulings, long-term bets on people.

When AI ultimately handles sourcing, screening, and first-pass analysis, what remains is always the most important part anyway. It may not be something you can resist. Perhaps that’s the point.

The author is Vertex Ventures Southeast Asia and India Associate, Investment Representative