A multi-entity venture firm put AI in its inbox, meetings, and books without autopilot

The problem

A venture firm running multiple entities had a founder and a small ops team drowning in coordination work. Email triage was manual. Answers to questions like "how much did this LP invest" lived in scattered tools. The bookkeeper reconciled months of card charges across entities by hand, with weak ground truth on what each charge was for. The firm wanted AI in its daily operations, but only in a way the team would still trust in month six.

What we built

• The founder's email agent was built voice-first. Before it drafted a single reply, it studied a reference set of roughly one hundred of his sent emails. The ops team rated draft quality ten out of ten in the first month. Skipping that groundwork is why most email agents sound like a polite robot.
• A Slack agent answers live questions across Airtable, Drive, meeting notes, and Notion: who invested what, where a deal stands, what was decided and why. Institutional memory stopped being a person you had to interrupt.
• Every recorded meeting turns into contextual email drafts and action items with named owners, posted where the team already works.
• An entity-aware reconciliation engine took 224 raw card charges spanning four months and produced an evidence-ranked review queue for the bookkeeper: 65 fully evidence-backed, the rest sorted by what needed a question, with the questions already drafted. Hard rule: nothing posts to the accounting system without her sign-off.
• A day-long workshop trained an internal builder who now ships workflows in the firm's own stack. The stack survives the engagement because the client's team can extend it.

Stack