Draper Builds AI Infrastructure, Now Manages Over 400 Portfolio Companies and 15,000 Contacts with an Automated Main Database
Draper Associates has announced the establishment of its internal AI infrastructure: two years ago, portfolio data was scattered across 8 Excel sheets, and now it is managed by an automated main database that handles over 400 portfolio companies, more than 15,000 contacts, and over 1,900 meeting minutes, all searchable in everyday English on mobile devices.
The main database operates on the principle of one record per company, with all other processes writing to or reading from it. The tech stack is built on existing tools: Airtable, Zapier, Power Automate, Google Drive, Slack, and Claude, with self-built services hosted on Railway, avoiding new vendor contracts. Meeting minutes are automatically attached to company records after being transcribed by Otter, and pitch links are converted to PDFs for archiving via MCP services.
The self-developed Draper Brain serves as the MCP server, connecting Airtable, cloud storage, and authorized external data to Claude, covering about 450 portfolios and over 4,900 co-investors. When uncertain, the system pauses to ask a person, avoiding guessing in fields that must be correct. Lead scoring is determined by the fund's own themes and historical negatives, rather than using a generic scoring system.
The Deal Engine is trained on 40 years of investment and rejection records, reading hundreds of incoming projects weekly and emailing them to the investment team. After submission by founders, Zapier triggers web searches and generates introduction emails to be sent; usage of this introduction feature has increased about tenfold since its launch. Schedule tasks read from a read-only inbox, CRM, and portfolio updates to generate daily briefs, while fund financials are synchronized via Power Automate.
The article is authored by Michelle Kwok, Andrew Hyun, and Patrick Blumenthal. The fund operates with approximately $2.3 billion in assets under management and plans to launch a $300 million Fund 8 in 2026. The portfolio includes publicly known paths such as Tesla, Skype, and later Carta, ICEYE, Polymarket, etc. Tim Draper mentions that he developed judgment from projects that were initially rejected in 1985, and now machines handle routine searches.
Who buys is determined by the investment team's time and judgment bandwidth, while who sells is the internal engineering that productizes routine searches. Funds are not directed towards new software purchases but rather towards cleaning proprietary data and maintaining prompts. Beneficiaries are partners who can use natural language to access files and founders who are matched with introductions more quickly, while those under pressure are project parties that still rely on multiple table exports and cannot be indexed by the main database.
The system clearly leaves "sitting down to meet founders and make decisions" to humans. Each tool has documentation written based on already launched products, requiring the next person to be able to get started in five minutes.
Source: Public Information
ABAB AI Insight
The Draper family has been in venture capital for three generations: Tim Draper started building judgment from projects that were rejected by others in 1985, and his descendants continue to manage the fund. The portfolio narrative has long been tied to early checks on companies like Tesla, Hotmail, Skype, and Baidu, as well as publicly failed cases like Yik Yak. This time, instead of discussing the next unicorn, the focus is on turning 40 years of "to invest or not to invest" into training samples, allowing internal models to first read hundreds of incoming projects weekly.
The funds have not gone to purchase a new venture capital operating system but are instead concentrated on a single main record for each company in Airtable, with Zapier and Claude integrating emails, minutes, pitch decks, and financial tables. Harmonic is used to supplement market profiles missing from internal files, while the Deal Engine uses its own rejection history for filtering. The motivation is that in the seed stage, human resources are thin, and after surpassing 400 companies, partners' time is consumed by searches; the strategy is to keep data ownership within their own database, where modifications equal changing prompts, rather than waiting for vendor releases.
Similar actions are seen in firms like a16z and Sequoia, which are productizing analyst workflows, and YC, which is modeling application scoring. Draper is in the "controlling internal information" phase rather than selling this system externally. Unlike external CRMs like Salesforce or Affinipay, it writes judgment criteria into a private model, avoiding generic lead scoring that would erase the fund's preferences for sectors like crypto, aerospace, and unconventional tracks.
Structural changes belong to the overlap of technological replacement and capital concentration. Once routine searches, minutes archiving, and introduction letter drafts are taken over by models, the same partner team can maintain a larger portfolio without linearly increasing headcount. The mechanism is that the scarce resource in venture capital is face-to-face judgment, not spreadsheet maintenance; whoever first turns 40 years of rejection records into callable filters can cover more project flows with the same number of people and redirect attention back to founder meetings.
ABAB News · Cognitive Laws
First determine where the data resides, then discuss what model to use
Judgment must be private; tools can be borrowed
Where the machine pauses, that is where humans should sign off