How Bain Capital Ventures Incubates AI

Editor’s note: This article is part of an ongoing series in which Crunchbase News interviews active investors in artificial intelligence. Read previous interviews with Felicis, Battery Ventures, General Catalyst, Bessemer Venture Partners, Accel, Insight Partners, Index Ventures, Sequoia Capital, Section 32, M12 and Sapphire Ventures, as well as highlights from these stories from 2023.

Bain Capital Ventures is investing out of its 10th flagship fund of $1.9 billion announced in February 2023 after the launch of ChatGPT, the technology that took generative AI mainstream.

Despite the firm’s many decades in the investment business and its sizable funds — it has $10 billion in assets under management — investing in AI companies comes with a new set of risks and a highly competitive funding environment.

Rak Garg, partner at Bain Capital Ventures

“This market is unlike any other. The field is moving so quickly,” Rak Garg, a partner at BCV who divides his time between AI infrastructure and cybersecurity investments, said in an interview.

To keep up, BCV set up a lab in Palo Alto, California, more than a year ago to support promising researchers and incubate companies in the artificial intelligence sector. The lab also hosts paper readings and debates, and invites prominent researchers in the field to present.

Research first

The only way to stay ahead of developments is to take a research-first approach, Garg said.

“Two years ago, we were doubting that large models could even exist and be served reliably,” he said.

Then, a year ago, vector databases were seen as the answer to enterprise AI.

And now the answer is “some combination of alignment, fine-tuning and vector databases.”

It takes small groups of people to figure out, “how do you get these models to be performant? how do you prompt them in the right ways?” he said.

BCV has made four seed investments from the lab since it launched less than a year ago.

The first company to publicly launch from the lab was Contextual AI, a startup that is building models for enterprises to access proprietary data to make employees more productive. The company raised a $20 million seed funding led by BCV in mid-2023.

Contextual co-founders Douwe Kiela and Amanpreet Singh met at Facebook AI Research (FAIR). There, Kiela helped develop RAG, or retrieval augmented generation, for large language models.

If you give “the model the right data at the right time and then loop back and look at the way the human interacts with that data, you end up with a sort of closed-loop system where it retrieves the right static information and then dynamically enriches it with interaction data,” said Garg.

Contextual has seen strong use cases in finance, healthcare, legal and other industries. It can also help employees answer questions like, “who should I prioritize this week?” or “what should the message be?” Garg said.

Compared to a technology like ChatGPT, Contextual models are smaller and require less compute.

Contextual built a joint system of “two small models with the right embeddings,” he said. “We can fine-tune them continuously on interaction data, then we can approach something that gets to be really, really accurate for that enterprise.”

“You would never use the Contextual system to generate a Shakespearean sonnet about employee insights,” he said.

The second company to come out of stealth from the lab is Prophet Security, co-founded by Kamal Shah and Vibhav Sreekanti. It raised $11 million in seed funding last month led by BCV. The company says it uses AI to analyze and triage security alerts in seconds, saving time for security teams.

Other investments from the lab still in stealth are in data science and analytics.

Beyond the lab

Garg said that competing with established software companies with an added layer on top for summarization will not cut it in this market. “The co-pilot for Atlassian is going to be built by Atlassian,” he said.

That’s why BCV is focused on investing in hyper-specialized models on the infrastructure side and inefficient processes for which there are no incumbents, said Garg.

BCV is an investor in Unstructured Technologies, a company the firm seed-funded before ChatGPT became mainstream. The company recently raised a Series B from Menlo Ventures.

Unstructured has a proprietary model that extracts textual information from PDFs, email or SaaS applications to get clean data for a database or for pre-training or fine-tuning data. It boasts 50,000 customer deployments, said Garg.

Other notable BCV portfolio companies in AI are Moveworks, which uses AI to support internal teams to address issues in a remote-work world; Poolside a foundation model company that aims to generate code with natural language prompts; and Cleanlab, a data cleaning company.

In software, more dollars are spent on services integrating technology solutions than on products, Garg said. Generative AI presents an opportunity to turn service revenue into product revenue.

“These services are taking data from one place, molding it in different ways, putting it in a different place,” he said citing Unstructured as an example. “Models are really, really good at that.”