# Can a VC fund run itself? Building a second brain for investing. > Source: > Published: 2026-09-17 08:02:44+00:00 Happy autumn, friends! **Operations Optimist** returns after a long summer! Last year I started out writing about operations, remote work, fundraising and hiring. Because that’s what I know. Then this year every facet of my work got touched by AI. *That’s why I’m reintroducing my newsletter.* *Here’s what to expect:* - *Experiments with AI in my day-to-day operations work at a tech startup* - *Case studies from other operators and founders who use AI* - *How to use it to advance our careers in operations* *Because I am increasingly seeing that a gym on every block doesn’t give everyone a six-pack. Meaning that access to the tools doesn’t mean we use them well.* So we’ll kick off with an insight from a venture capital firm that is redefining how to use AI in real time. Meet **[Bad Ideas Fund](https://badideas.fund/)**. We worked together a few years back: I was a part of their Fundraising School, an early member of the investment committee, and made a few angel investments alongside them. So I sat down with [Roberts Bernāns](https://www.linkedin.com/in/roberts-bernans-87264b17/) expecting to talk about agents, and how they had advanced the Notion-run fund I once knew. Turns out, it’s history already. In 2014, Roberts founded Nordigen, the Latvian fintech darling sold to GoCardless, which has now been acquired by Mollie. He is now a partner at Bad Ideas and is building Bad Brain, **a second brain for how the VC operates.** Venture capital is largely a pattern-recognition business. As an investor, you screen thousands of companies to find the few founders who turn sheer dedication and vision into something great. Founders submit applications and decks. The decks get scored. The scoring becomes an investment committee input that later gets discussed. The discussion becomes a call, then a follow-up. All of it lands somewhere as unrelated data. In a business of connecting dots, the opportunities for AI are limitless. And Bad Ideas has historically connected these dots via n8n automations, Notion dashboards and Airtable bases. But as I learned, that’s history too. So let’s dig in and try to answer an overarching question: **can a VC fund run itself?** ## **A garbage in, garbage out problem** When Roberts arrived, the data was scattered across Airtable and Notion. Everyone has their own pages, anyone can write anything, but there’s no mechanism for knowing what’s current and what is not. Notion upkeep is work that shouldn’t exist in the AI era. As he puts it: by the time you finish writing a sentence, it’s already out of date. Airtable had a second problem: it’s a bad source for agents. Slow to query and too loose to hold anything reliably. Frankly, it might also explain Airtable getting *spooned* in [that valuation drop](https://techcrunch.com/2026/08/04/bending-spoons-to-buy-airtable-for-1-28b/). Ouch. ## **Data has more than one dimension** Roberts gives an example of why this fails: Airtable- and Notion-like structures only see data in one dimension. Take someone on their investment committee who eventually decides to become a founder. Neither tool can query both of those at once, and the whole structure gets slightly less usable over time. He continues: anyone who’s worked at a rapidly growing startup knows the phase where the company grows faster than the codebase and you pay for it in blood, sweat and tears. Data is the same. Anything living freeform somewhere is tech debt. **His conclusion: it’s easier to start from nothing than from bad data.** Reworking years’ worth of data is an expensive decision, but proper foundations are worth it. ## **So how did they build the second brain for VC?** They moved everything into **[Supabase](https://supabase.com/)**. First, it plays well with agents: Claude, Cowork, Code and their own services all query it fast. Second, the migration forced them to think harder about structure instead of importing something that already didn’t work. Then comes the question of what goes into it. We’ve established that venture capital is a high-context business. To make that context usable, they don’t dump raw transcripts into agents. They codify every interaction into what they call **cards**. Every call, email, Slack or WhatsApp message gets processed, and the current facts get extracted into a card. Talk to a portfolio company, and that company’s card updates with what actually changed. The truth lives in one queryable place, not across hundreds of raw artifacts. Over time every interaction compounds into a library of cards you can parse. On top of that sits an application layer that each person can build for their own use case. Small VCs are known for having lean, generalist teams. So I can already imagine plenty of use cases for real-time data: deal flow scraping, marketing, portfolio reporting and more. As for more use cases, he mentions operations people plugging the fund’s internal MCP into Claude Cowork and asking it for portfolio reports. More tech-savvy team members go straight through Claude Code, or build dashboards in Replit that pull straight from Supabase. Each person’s work is magnified by having access to a gold-standard database that does not get stale. ## **Three stages of AI adoption in companies right now** That data layer is what enables companies to travel through the three stages of AI adoption, as Roberts calls them. 1. **Automation.** A year ago most VCs were doing n8n automations: startup emails a deck, automation parses it, CRM gets a row. Nice, but nothing gets done smarter, just faster — the data is still scattered. 2. **Copilots.** This is where Bad Ideas is now. Copilots are agents running on dedicated cron jobs, scheduled for a specific purpose. Like a marketing copilot that researches sources, drafts, and reviews what it posted, similar to a junior marketer who does their homework every night.Or their chief-of-staff copilot that preps the weekly. It writes follow-ups, checks progress against OKRs, and makes the meeting about blockers instead of status updates. The defining feature: a copilot makes the human faster; it doesn’t remove them. 3. **Autonomous.** Six to twelve months out is when he thinks they will see autonomous agents. Agents with so much structured, accessible context that they can start making judgment calls inside an organization. Think of it as a go-to-market agent that sits inside every expert session with portfolio companies and logs everything, so eventually it knows more about go-to-market than any single expert they could hire, because it sat in on all of them. **Bad Ideas is now building a fund on the assumption that every human interaction will be done by an agent later.** ## **The verdict** So is the fund run by AI now? Not yet. Is it getting closer every day? Yes. If you’re an operator setting out to build a second brain inside your organization, the real work starts with organizing your database. Try answering honestly: do you know where the truth lives in your company, and could a machine actually query it? #### **Suggested listen** As the last bit, I asked Roberts where he gets his AI inspiration. He mentions [Dwarkesh Patel’s podcast](https://www.dwarkesh.com/), his pick for understanding where AI is actually going. Recommended listening speed: 0.5x. ## **Here’s what we covered today:** - **Bad data is legacy debt** : easier to start from nothing than from a mess without clear ownership - **Intentional structure beats tooling** : the migration mattered less than the redesign it forced - **Three stages of AI adoption** : automation → copilot → autonomous agents. Most AI-native companies right now live somewhere between the first two. Thank you, Roberts, for sharing your insights! And as always, thank you, dear reader. This newsletter is free and runs on your support. Every like, comment or share puts it in front of someone new, and I appreciate it so much. See you in your inbox, Diana