Almost every VC uses AI. Very few can name a job an agent now does for them from start to finish.
That was the starting point for our latest virtual event, How VCs Are Putting AI Agents to Work, presented with our friends at Carta. We brought together three people who see agents from different sides of the table. Yohei Nakajima, GP at Untapped Capital and creator of BabyAGI, builds agents for his own firm. Lindel Eakman, founder of Eden Road Capital and formerly a partner at Foundry, has backed emerging managers for more than 20 years and sees how firms operate behind the pitch deck. Scarlett Pierce leads product marketing for Carta’s investor products, including the Carta plugins for Claude.
The conversation got specific fast, with live looks at the systems these folks run every day.
TL;DR
- Yohei’s fund playbook updates itself from his notes and emails, and he built most of it from the driver’s seat of an RV
- Lindel explains why the AI slide in a pitch deck earns a GP almost nothing with him
- A Carta customer took quarterly reporting from a week-plus across 28 spreadsheets down to one prompt
- There’s one step in Yohei’s process that he keeps entirely to himself
Here’s what we learned (or watch the full session below) 👇
This virtual event was presented by Carta
Carta is the financial infrastructure platform that founders, investors, and limited partners use to manage equity, fund administration, and investor relations. Carta’s intelligent platform connects the data, workflows, and people that drive private capital operations—bringing together cutting-edge AI and expert human service.
Yohei’s Playbook Writes Itself
Yohei runs his fund through a set of repos that any of his agents can read, whether he’s talking to Claude, ChatGPT, Codex or Grok. A chief of staff repo tells each agent how to behave. A personal memory repo pulls from his email, calendar and Granola notes, which he says capture 90% of his tasks. A shared work ledger keeps every agent in sync, so Claude knows what ChatGPT did an hour ago.
“I wake up in the morning and ask my agents what I’m doing today. And it’s reliable.”
It also reminds him of things he’s forgotten, including people he needs to nudge.
He had an agent read every email that mentions a portfolio company. Now he can see the acceptance rate of every investor intro he’s suggested, broken out by VC and by portfolio company. When a company starts raising a seed round, the agent pulls the most relevant investors he has relationships with, removes anyone he’s already recommended, and hands him a list to go over with the founder.
The centerpiece is what he calls the firm OS, home to the fund’s policies and processes. He describes it as a flipped playbook.
“People think of a playbook as something you write down and you follow. But that’s almost like the waterfall method of development.”
Yohei’s version reads internal notes and LP communications, infers the decisions the team has made, and proposes updates. If the firm starts talking to a new auditor, the OS logs when those conversations began and when they switched. Nobody types that in. Extraction and authority are kept separate, so every change stays a suggestion until Yohei confirms it.
He tried building a single “company brain” and abandoned it, because each person’s agent needs its own memory and sharing it created dedupe and leakage problems. The firm OS points to where the data already lives. The CRM is still the CRM, Carta is still the system of record, and the LPA is still a PDF. “That’s the truth,” he said. When the OS pulls a fact from the LPA, it notes the page it came from.
Most of this came together during a two-week family RV trip in August. With Starlink on the RV and remote access back to his iMac, he spent long stretches of driving talking to Codex, asking it to go through drafts like his DDQ and read back what belonged in the firm OS so he could say yes or no.
The LPA detail struck a nerve with Lindel, who said most GPs actually don’t know their own LPAs. They trust their attorneys to negotiate a good deal and move on to investing. Lindel recently did something over a weekend that tripped a provision in his own LPA. A team member who spent six years as a Cooley attorney caught it, and the fund needed approval from an underlying LP before making the investment the next morning. Running an agent against your LPA before you act is, in his words, a perfect use case.
Why Lindel Discounts the AI Slide
Every pitch deck now has an AI slide, usually with a cool name attached. Lindel treats AI as table stakes, and the variability he sees across managers is striking, especially among solo GPs. He recently met a Fund 1 manager with an interesting portfolio who isn’t getting leverage from AI. At this point, he called that a showstopper.
Compare that with a manager he recently backed. She left an existing firm and took time to build out her workflow system, with agents that give her the kind of leverage an analyst or a few interns used to provide. Lindel wants to know how a GP spends their day, because time is the real constraint in this job.
“I sort of discount it all immediately if they can’t show me rather than tell me.”
The analyst comparison is showing up in the data. In Venture5’s VC compensation survey, now in its ninth year, analyst comp dropped 20% last year, the first time we’ve seen that. John has also heard GPs say outright that they won’t make a planned hire because of the tooling they’ve built.
Lindel also sees an upside for the long tail of deal flow. VCs have always leaned on warm referrals because nobody can process endless inbound. With a filter that carries more context than one person’s memory, GPs tell him they’re spending more time on cold inbound than they ever would have. He sees the same split at the largest firms. He recently compared notes on two mega platforms, and the one using AI in investor relations is accelerating on capital raising and brand building while the other is not.
Yohei pushed the point further. Saving time on operations frees you to go deeper on a technology or an industry, and if that research happens with AI, the agent remembers it when a relevant company shows up later. The bigger lever for him is building in public. Open source projects like BabyAGI connect him with founders working on the same technology, and Lindel confirmed that’s exactly how he first learned Yohei’s name.
“Share a lot publicly, share more privately to LPs, and tell people that you’re doing that.”
Yohei is pretty sure some of his LPs signed up because of his regular AI sessions.
Keep Your Data Where It Lives
Scarlett said Yohei’s diagram lines up with the bet Carta has made about where fund managers will want to work. Over the past six months she’s watched software companies do one of two things. Some release an MCP connector, which Carta sees as “a starting point, not an end point.” Others build agents inside their own app, which keeps users locked into that vendor’s workflows and roadmap.
Carta’s approach is the Carta plugins for Claude. A Carta Home artifact inside Claude replicates the web app dashboard, down to capital activity and reminders for LPs who still need to pay in, and lets you ask questions and run analyses on that data in conversation. The skill directory behind it grows daily. The newer Carta Work Hub moves from analysis into operations, so a fund CFO or GP can set up a capital call, a payment or an LP transfer from Claude and route it back to Carta’s fund admin team.
At Delta-v, quarterly reporting used to take a team more than a week, built by hand across 28 spreadsheets covering the firm’s funds and SPVs. Feeding those spreadsheets into Claude one at a time cut it to days. Then the firm connected the Carta plugin and used its Form ADV skill, which pulled the data straight from its Carta account into Claude.
“It was hard to imagine, but it was a one-shot exercise.”
Moving money raises the stakes. Often, capital call PDFs with bank details get passed around in ways that are neither efficient nor secure. Yohei said his banks are read-only for his agents today, and PayMan, a company in his portfolio, is working with banks on wires.
Scarlett’s advice is to start with permissions. When you connect any MCP to Claude, Codex or ChatGPT, go through which tool calls the agent can make on its own, which need express permission and which are one-offs. Carta also takes responsibility for protecting the edge cases for users who skip that step.
She’s most excited about connecting tools that never used to talk to each other. If your email, CRM and Carta account all have MCP connections, you can design a flow that doesn’t depend on any one vendor’s roadmap. Her message to fund managers is to “take it and build on top of it. Build that 20% that makes it the most valuable tool for the way that your fund specifically operates.”
What Yohei Won’t Automate
Scarlett turned the question around and asked Yohei what he refuses to outsource. His view is that AI can automate almost anything, so the things you choose to keep human end up defining your firm. For him that means decisions and relationships.
He’ll let AI filter out companies that are clearly not a fit. Anything that might be a fit, he looks at himself.
“I don’t actually even use AI to score companies.”
He worries it would bias him in a way he doesn’t want.
Relationships have a gray area. Yohei sees scaled relationship building with founders at large as part of a VC’s job, and he’s comfortable with some automation there as long as it’s clear it’s automation. Portfolio founders and LPs get a different level of care. Scarlett agreed on the LP side. “Don’t try and create a virtual equivalent of the coffee chat anytime soon.”
Even the automated touches can be more respectful of a founder’s time. Yohei’s metric requests now include the latest numbers the fund already has from that founder and the date they arrived, with a simple ask to reply if anything changed. “One of my biggest pet peeves… was asking the founder for metrics when you had them,” he said. Replies update his system automatically, and the response rate is significantly higher than an email without the metrics. Scarlett noted that Carta’s portfolio product includes data collection too.
Where to Start (and How to Keep It Running)
Yohei’s advice for most people is a lot simpler than his own setup.
“Pick one and get good at it, because there’s so much that all of them can do that I think a lot of people aren’t even taking advantage of.”
From there, connect the tools you already use, then add ones that could help. His next-level tip is to start every project by telling your agent to set up a local repo that stores your learnings and decisions. ChatGPT and Claude will keep memory for you, but a repo travels with you across agents. Scarlett gave the idea of organizing everything into files and folders a massive plus one.
Inside Carta, every team works from one rule. If you’re doing something more than once a week, find a way to automate part of it with Claude. Teams package those automations as skills and share them, and marketing runs on its own marketing OS holding knowledge that used to sit in the Google Drives of 75 different people. That created a new governance problem, since everyone building means a lot of skills doing nearly the same job. Scarlett described the past year as teams splitting into silos to get the most out of AI, and the next phase as learning to work together again with an AI platform as the shared work surface.
Keeping a system healthy takes its own care. Lindel hears from GPs that “your skills quietly fail.” Yohei handles that with a meta repo whose single job is a heartbeat. It tracks how the system evolves and tells his chief of staff agent when something breaks. The chief of staff mostly assigns work to other chats, so more runs in parallel.
John admitted his auto-drafted emails now pile up faster than he can send them. Yohei’s fix has two parts. When you fall behind, ask AI to prioritize and work through the list. If the system keeps surfacing things you don’t need, sit down with it, walk through a few examples, and have it update itself so it doesn’t happen again.
Lindel has watched a few technology transitions, and he thinks this one is big enough that the entire stack is being rebuilt on top of it. He’s blunt that a lot of older GPs aren’t moving fast enough. When he started in venture, the phrase was digital native.
“Now I think it’s AI native, and often that’s just a matter of your willingness to engage with the technology.”
There’s more in the full session, including Scarlett customizing a schedule of investments view live in Claude, John’s look at the batch workflows that draft Venture5’s CRM emails and build event invite lists, Yohei on when newer extraction models are worth swapping in, and Lindel on why portfolio companies built before and after this shift need different advice from their investors. Watch it above.
