How I built an AI analyst for my firm, no engineers needed

One of my partners recently asked a question on Slack that would have normally ruined somebody’s afternoon. They wanted to know if we had pro rata rights, which allow existing investors to participate in future funding rounds, in one of our portfolio companies.

The answer wasn’t in our customer relationship management (CRM) software or portfolio dashboard, but in a stack of PDF files in a Google Drive folder. For a human to find it, they’d have to open documents, read dense legal clauses, and hope nothing was superseded in the last funding round.

Image credit: Ulla

Instead, a bot named Ava handled the query, pulling the portfolio company’s Google Drive folder from our CRM. Ava then checked the firm’s legal domicile to determine which file covers that right, read the financing documents and the current charter, and answered while citing and linking to the document discussing the right.

Nobody on my team had to open or search a single PDF file, saving us a few hours of work.

I run a VC fund – I’m not a software engineer anymore. But I was able to build Ava over a weekend by describing what I wanted to Claude Code and being stubborn about the details. The toughest decisions weren’t about code but trust: what the bot was allowed to touch, which system won a disagreement, and what success with Ava looked like overall.

Many knowledge-based teams face the same multiple database mess, so consider this both a showcase and a dare.

The tab tax

Our tech stack includes four systems: our CRM Attio, equity management platform Carta, portfolio intelligence software Vestberry, and Google Drive. Yours may not be exactly the same, but it’s likely fragmented as well.

Finding the answer to any serious question involves most, if not all, of these platforms. We call it the browser tab tax: Every partner pays a 25-minute toll of clicking, searching, and rereading before each meeting.

See also: The global hunt for AI compute targets SEA

One approach to this problem was to build a chatbot to search documents. But what we wanted was a junior analyst who could read documents and data, then show us where the answer came from.

Enter Ava, built on Anthropic’s Claude Agent SDK. The architecture is best understood as a tiny organizational chart rather than a pipeline. Ava works with the help of AI agents we’ll call specialists, each of which has a specific function.

The orchestrator is the only part of the system that decides anything, with its routing logic written in English, not Python. A roughly 600-line prompt tells the orchestrator which questions to answer directly, which specialist to dispatch, and which actions it must never perform.

Meanwhile, the five other specialists exist to ensure that they each get a fresh start, so one task’s noise doesn’t bleed into another.

Don’t give the analyst the master keys

Why “wrong document” is the best answer

Paranoia: a product feature

The cost

The limits, and what I’d do differently

How to, if you want one