Using AI in VC: Automating Market Size Research with LLM Agents
The Augmented Venture Capitalist.
Hi all! If you’re already a subscriber feel free to dive right into this week’s edition. 👇🏾
If you’re new, a warm welcome and thank you for subscribing! This newsletter shares the best VC wisdom and knowledge curated from seasoned investors, researchers, and technologists.
You’re part of 400+ industry insiders who’ve joined along on a journey to mastery & impact.🚀
I’m sending this newsletter off-schedule as an experiment. Thank you for reading!
Agents are the latest craze in AI. Powered by large language models, agents are programs that make autonomous decisions to complete specific objectives. You can use agents to automate a wide array of tasks, and in this post, I’ll share a test case I coded up over the weekend to automate a common task for VCs: researching market size and trends.
Before I get to the example, it’s worth highlighting that tech startups and indie developers aren’t the only ones experimenting with AI agents. Corporates are paying attention, too. The luxury hotel brand IHG is building agents to assist travellers with holiday planning. Mercedes-Benz has agentic chatbots that help you book test drives. At Bristol Myers Squibb, a global pharmaceuticals business, what took weeks for a first draft documentation process now takes minutes, thanks to employee agents.1
It’s also possible to realize gains with research work in venture capital. That’s what I set out to explore and below, I’ll share a brief look at what’s possible. Note that the idea here is to provide draft analyses, not conclusive work that determines an investment decision. But used carefully, agents are bound to augment, transform, and supercharge what VC teams can do.
The Benefits & Outputs
Here is a sample report my agents created in just under 2 minutes. I simply gave it the company website for Marshmallow, a fintech unicorn in the UK, and my agents provided a snappy market assessment with links to information sources.
Here’s a preview of the agents in action.
How it Works
Step 1: Choose an Agent Framework
Getting started is now easier than ever. You don’t have to write all the required code, either. You can use a framework that provides scaffolding for your agents. Options include CrewAI, LangGraph, and AutoGen. There are also no-code alternatives like RelevanceAI, but I’ve found that writing code is usually faster and better for quick experiments. For that reason, I went with CrewAI.
Step 2: Architect Your Agentic Workflow
My plan was simple: Create three research agents and have a fourth one summarise the findings of the other researchers. So I have one agent that reads a target company’s website, two that conduct a top-down and bottoms-up market sizing, and a final agent to collate the findings. Here’s a preview of the code.
Step 3: Run Your Agents. Review Performance. Iterate.
The first time I ran the agents the outputs were a mixed bag. Sometimes you get great results and at other times, the results are questionable. But all this depends on how good the underlying LLM models2 are, the tools you provide your agent, and the prompts you instruct these agents with. Here’s another example report. It’s for Robin AI, a legal copilot startup.
The Augmented Venture Capitalist
Agents aren’t yet good enough to be relied upon entirely and they probably never will be. However, if you put time and thought into building agents that are specific to your style of analysis and work, you’ll be able to generate first drafts rapidly. You can then drill down into the more curious areas of an investment opportunity, such as the magic of a founding team and the potential that can’t be quantified by AI or data.
In my experiment I mostly used GPT4 because it provides the best results for now. However, for more straightforward tasks, such as extracting and summarising text, I went for an open-weights model by Mistral.