A few years ago, we wouldn't have trusted AI to draft an email. Today, two agents work alongside our team daily. Here's what changed, and what it bought us.
This is the first piece in a three-part series on how we've put AI agents to work inside AVV, and how we approach AI as a team. This one covers what the agents actually do for us. The next two get into where they break, and where we think all of this is heading. For more on how (and why) we built these agents, check out GP Binh Tran's Tech in Asia article.
About four years ago, Binh, a longtime technologist, flagged what the world was then beginning to call generative AI. We ran an internal future-visioning session, and a few months later shared what we were seeing with the broader startup ecosystem.
At the time, the technology underwhelmed us. It could draft a little and do some light research, but the quality was low, and we didn't lean on it for anything that mattered. "Just a few years ago," says GP Eddie Thai, "I didn't even trust it to draft emails for me."
To be clear, AI wasn’t new to us. Binh’s startup was using AI in 2009, and our first investment in 2015 was ELSA, an AI startup. We knew the tech would fundamentally change everything, but it was far from production-worthy then.
What's changed since is hard to overstate, and Eddie describes the technology entirely differently. "It's now basically serving me most as a kind of junior analyst role," he says. "Which is really saying something."
Three developments helped us decide AI agents were ready for production use, and none were about the underlying model.
The first was where the AI lives. Rather than have everyone sitting in a separate chat window, copying context in and out, we embedded our agents directly into Slack. As Eddie puts it, "embedding it in the space where we work already with humans makes it really feel like our team members. It lets others contribute and collaborate."
As a result, colleagues can see how someone is using the agents, copy the approach, or pick up the next turn of output themselves.
The second was building for specialization. We run two agents, both built in-house by Binh, designed to do different things.
Khoa is the generalist and the first point of contact. It's flexible, can search the web, write custom software, and orchestrates increasingly complex tasks: it takes a request, handles what it can, and routes the rest. It will even read its own setup and write new functionality when asked.
Ava is the specialist. Its job is to reliably pull from our systems of record, namely, cap table data in Carta, portfolio financials in Vestberry, documents in Google Drive, and email histories in Attio. A general-purpose agent can be impressively flexible and still get the details wrong; a narrow one, pointed at the right sources, is far more dependable for the work that has to be right.
There's a small, potentially uncomfortable side effect of working this way: some of us have started using "he" and "she" for the agents. Eddie caught himself doing it mid-conversation. "I'm using personal pronouns," he noted. "It's funny, right?" It feels odd to admit, but it's a signal of how the relationship has shifted from running a tool to working with something that feels like a colleague.
The third development was AI harnesses becoming less brittle and maintaining state, allowing the team’s knowledge to compound over time. This includes where information is stored, how we like things displayed, and internal team routines, all becoming shared knowledge that the agents store and remember from different sessions.
The clearest way to show what the agents do is to walk a single deal from intake to decision.
It starts before any agent is involved. New opportunities come in through our website, our channels, and our networks, and run through a single intake that scores each one against a few key parameters. The scoring doesn't decide anything, but it helps us point our attention at what looks most promising first.
Once we decide to look harder at a company, the agents take over the grunt work of profiling it by pulling together the materials a founder sent, plus what's publicly available, into a clear picture of what the product is, who's building it, and what stands out. This used to take anywhere from half an hour to a couple of hours of prep, something that now takes minutes. We'll often have an agent draft an initial question list to guide the first founder conversation, too.
After the call, the leverage compounds. Granola automatically stores the transcript and summary in Attio (our CRM), and the agents conduct a first assessment of what's strong and what's weak, drawn against the firm's own memory of how we evaluate deals, along with suggested follow-ups. A human still removes what we don't agree with and elevates what matters, but the consolidation and the first pass at sharing it with the team happen fast.
Getting from deal intake to a first substantive team conversation used to take a few days. Now, founder schedules permitting, it can happen in hours.
Finally, there's the memo. Drafting an investment committee (IC) memo, even after the research and conversations were done, used to take up to two days of focused effort. A first draft now takes a few minutes.
To ensure that this isn’t just us flattering ourselves, we compared notes with Francis Perelman, AI Program Manager at Capria Ventures, one of our LPs. Capria runs a fund-of-funds and direct investment model focused on applied AI opportunities in the Global South.
His read on the industry was frank: most funds assume they're using AI to transform how they work, but many are still doing things manually or leaning on the same old tools. Capria is one of the exceptions, and it has built a more elaborate version of AVV's arc above.
They started in 2025 with comfortable classification work like scoring the flood of applications that arrives when they open a role in a market like India, and a chatbot to field HR policy questions.
As model context windows and reasoning improved, they moved to more venture-specific work: first analyzing pitch decks against their thesis, then whole data rooms (including ones with download restrictions), and eventually a full end-to-end framework with distinct agents for sourcing, diligence, and building the IC documentation behind an investment decision.
While Capria has more agents than us and has been using this tech as a firm for longer, the decision-making flow is the same: at every step, human approval pushes the process forward. "The agent is the driver in each of the steps," as Francis describes it. A person always decides when to advance.
The investment workflow described earlier is where the time savings are biggest, but it's far from the only place the agents have an impact.
On legal issues, where our legal counsel Natalie Hoya is a team of one, the agents have become a backstop. "Here it's only me," she says, "so I have to have something else that can help me to be the second eye." Finding a specific clause across a stack of documents, or building a portfolio-wide summary of key terms, now takes about half the time it used to. But speed isn't the main benefit Natalie points to. “More than that, it helps me certify the accuracy. That's most important." When there's no second lawyer down the hall, confirming that the right document is the one being read matters more than the time saved.
Critically, the agents do this by referencing actual language in the legal documents they scan in our database, not by searching for outside information.
Hau Ly, a partner who is heavily involved in our reporting and hiring work, uses the agents to pull fund- and portfolio-level data that would otherwise require combing through several platforms. Capital called versus deployed for an impact report, for instance, turns from multiple manual lookups across systems to a couple of minutes interacting with the agent.
The most human example came from portfolio support. Eddie had a coaching call with a founder near the end of a hard startup journey, walking him through a clean wind-down and the investor note he'd need to send. Normally, we would’ve asked the founder to take what was discussed and draft the email on their own. Instead, the founder opted into a faster path. We fed the call transcript to our agents, they produced a clean draft based on the company’s current and historical financials, and the founder sent it nearly untouched. Something that might have taken days of back-and-forth, or an hour or two of a stressed founder's time, took ten minutes.
There's one more example worth telling, because it shows where this is going. We'd been using Binh's agentic AI to help founders iterate on their fundraising decks, which involves taking sprawling information and a founder's story and shaping it into something a prospective investor can understand quickly. This custom deck generation tool was better than any professional service or tool we’ve seen on the market, but it sat on Binh's machine, which made him a bottleneck.
So we asked the obvious question: how do we make this available to the whole team without routing everything through one person? Binh asked Khoa to build his own setup, and within about ten minutes, the agent had written a solution we could all use. That now involves ingesting a draft deck, a company's website, and whatever other guidance we had, and generating a new turn on the deck. What might once have been hours of working with Claude Code to create custom software can now be accomplished by anyone interacting with Khoa.
To be clear, the output still needs work, and none of this is push-button. But the floor for what a small team can produce, and how quickly, has risen substantially.
Add it up, and the through-line isn't that AI replaces thinking: It's that AI hands time back to it.
The most useful frame we've landed on is the one Eddie keeps returning to: the agents play the role of a junior analyst or associate. They speed up research, summarize calls, handle follow-ups, get the team up to speed on an opportunity, and surface the issues worth a closer look. That frees up his time, as he describes it, "to focus on the qualitative gut factors, X factors of whether an opportunity makes sense." These are the parts of the job that can’t be delegated.
That's a meaningful shift for a firm of our size, as work that once required either more hours or more headcount now requires neither. The leverage shows up in how quickly we can move on a deal and how much of our attention we can reserve for the calls only people should make.
It's also a shift you can measure, and Capria has taken significant steps to do just this. They run an internal dashboard that tracks every agent and tallies the hours saved against a benchmark of what each task would cost done manually.
By their count, the agents have already run thousands of tasks and saved over 1,000 hours of team time across every level, from managing partner to junior associate. They share the running totals on their Slack to keep people engaged and make the impact visible, aligning with our reason for putting the agents in our Slack in the first place: when the value is in plain sight, more people utilize it.
But we don't want to oversell this, because the story has a second half. The trust we now place in these agents was earned slowly, and it remains conditional. Every memo draft still gets vetted, every number headed for an LP still gets checked, and the agents have a habit of failing in ways that are instructive for all involved.
That's the subject of the next piece in this series: the second eye, and what it can't see.
