VC has long been criticised for relying too much on gut feel and warm intros to make bets.
But as the number of VC firms has exploded, more investors have turned to data to find an edge. And, most recently, that includes reaping the benefits of step changes in generative AI.
Take London-based VC firm Moonfire. Late last year, the early-stage firm was trying to improve one of the machine-learning-based tools it uses to identify promising companies. One engineer made the model 5% better after three months of work — a result partner Mike Arpaia called “awesome”.
Then large language model (LLM) GPT-4 was released in March this year. In a single afternoon, the same engineer improved the model by a further 20%.
“GPT-4 was really the step change in model performance…we’re entering an industrial revolution,” says Arpaia.
A number of investors tell Sifted that they’re beginning to explore sourcing and classifying deals using generative AI models, as more powerful models have become widely available and easily accessible.
How VCs are using generative AI
Investors say that there’s a key difference between the rise of generative AI and previous developments in data-driven deal sourcing. Generative AI refers to a class of deep learning models that are exceptionally good at generating and understanding content such as text and images.
“There’s been a shift in VCs engaging with AI, as access has been democratised,” says Arpaia. “Before you needed to be a linear algebra expert to do deep learning at that scale reliably, but now all you need is an OpenAI key.”
New models like GPT-4 are helping VCs cut down on hours of laborious work classifying startups in their internal databases.
Episode 1 has a proprietary data platform that sources 300-400 potential investments — from a number of places like Companies House, LinkedIn and GitHub — each week and has begun to use generative AI models to quickly classify them.
“We invest in B2B software companies, so we can ask a generative pre-trained transformer model [ChatGPT-like software] to filter out the ones that are in our scope,” Adam Shuaib, principal at Episode 1 tells Sifted. “Without that tech, I would have to spend five or six hours manually going through 400 descriptions…this wouldn’t have been possible a year ago.”
Blossom Capital has built a model that generates a one-pager about a company with details such as founder background and employee growth. It‘s planning on adding more generative AI features over the next 12 months, such as factoring in details like who owns shares, company job posts, customers and competitors.
Classifying startups using generative AI models “is definitely going to be a core part of how we source deals,” says Imran Ghory, partner at Blossom Capital and backer of Checkout.com and Pigment— and it’s hiring data scientists as it looks to double down on the tech.
Making analysts 10x faster
Pan-European investor Speedinvest, too, has recently onboarded an engineer — the first it’s hired specifically to work on the tech — to build generative AI tools for its investment team. The VC is working on a tool that can assess founders based on attributes like previous experience and education.
This is more challenging because of how much more nuanced the question of a founders’ experience is than the sector of a startup, says Florian Obst, principal at Speedinvest.
Hallucinations can happen even when asking generative models simple things, let alone the complex matter of what makes a top founder. “It’s a less forgiving exercise because we don’t want to miss out on good deals or focus on bad ones,” he adds.
Speedinvest hopes to roll out more advanced models to its investment teams towards the end of this year. “The idea is to make every analyst 10 times faster than before,” Obst tells Sifted.
While Obst says that extra efficiency will mean analysts can spend more time focusing on the “people side” of investing, others think automation means jobs could take a hit.
“You might find that the bigger funds at Series B and C might not need as many analysts or associates because a lot of that manual work can be done very easily and effectively by generative AI models,” says Episode 1’s Shuaib.
The future generative AI in VC
VCs using generative AI are careful about positioning the tech as something that will change the fundamentals of the business.
“VC is an industry where you can make good money in a number of ways, and it’s not absolutely necessary for VCs to have AI-driven company and founder evaluation,” says Arpaia.
While Moonfire has “pinned its efficiencies” to emerging tech like generative AI, a number of bigger funds might not be so quick to buy into it, he adds. “You don’t need to see all of the world to make 12 good investments each year.”
And some VCs will continue to focus on getting dealflow through their network, says Obst. While VCs like Speedinvest that cover multiple geographies will need to use tech like generative AI to source deals, “a lot of deals are still done between people where they don’t even hit the market”, he adds.
Even some of the most sophisticated data-driven, deal-sourcing VCs aren’t convinced by the power of generative AI models just yet.
Swedish VC EQT Ventures’ internal data and machine learning platform, Motherbrain, is run by a team of 25 people. It scrapes huge amounts of data online and suggests companies that investors should investigate.
“We’re not actually using generative AI — if you define generative AI as generating new content based on existing content,” says partner Lars Jörnow. While he doesn’t rule out integrating LLMs like GPT-4 into EQT’s models in the future, they aren’t as good at burrowing down into granular data as Motherbrain’s machine-learning models at the moment, Jörnow adds.
And there are worries that filtering startups based on AI models could further entrench the wildly uneven fundraising landscape in VC.
“A dumb AI model might look at the European unicorns, see that founders are typically white and went to Cambridge University and try to find more with those attributes,” says Blossom’s Ghory. There are a number of cases where generative AI tools have been found to double down on biases.
But, on the flip side, well-trained models could help to identify founders who normally wouldn’t be on a VC’s radar, he adds. “But you can also train AI models to focus on the things that really matter, like if a company is winning great customers or making great hires. It could mean that companies who might not be on your radar have an easier time finding funding.”