Building Company Brains in Google Drive

He walked the room through building a company brain in Google Drive, live

Austin Distel

The problemEvery founder's business knowledge lives in their own head, scattered files, and past conversations, with no shared system their team or their AI agents can actually query.

Austin and Darby Rollins spent the teaching half of the session building real company brains, live, using nothing but Google Drive. The framework has three levels: Level 1 is founder-only, just for you to query and solve your own problems; Level 2 opens access to your team, so a member's admin staff can ask the company brain questions instead of texting the founder on vacation; Level 3 is AI agents with full context that keep the brain current on their own. The structure is a shared drive, kept separate from anyone's personal Google Drive, organized into plain markdown files, chosen because they are the cheapest format for an AI agent to read without burning tokens on PDFs or spreadsheets. Members installed the Agentic OS MCP connector through the member portal, downloaded the company brain starter kit, and chose between Harvest mode, which pulls context automatically from every already-connected source like Drive, Gmail, and GitHub, or Interview mode, a from-scratch question-and-answer setup. Austin demoed both a browser-based harvest of his own franchisor's training system and a Gmail-keyword harvest that builds a contacts directory automatically, and walked through exactly how he structures permissions at his own spa business, from full founder access down to a front-desk employee who can ask company policy questions without ever reaching him.

You are my chief of staff, modeled on the company brain framework taught at the Agentic Society Boardroom (July 2026). Help me build a Level 1 company brain in a shared Google Drive folder, separate from my personal drive. Interview me for: (1) my business's key departments and who runs them, (2) what documentation already exists and where (Drive, Gmail, Notion, GitHub, or elsewhere), (3) what should never be shared even internally. Then: (1) propose a folder structure using plain markdown files, including a router file, a constraints file, an org chart, and one folder per department, (2) if I have existing sources connected, harvest context from them automatically; otherwise interview me from scratch, (3) propose a permission structure that mirrors my org chart, from my own full access down to my most junior role. Do not delete or overwrite anything without asking me first.

Takeaway The scaffolding you build for yourself at Level 1 is what makes Level 2 (team access) and Level 3 (self-improving agents) possible later — skipping it means rebuilding later.

~80 hours of manual work per bid down to about 30 minutes

The problemBidding on government contracts took two people roughly 80 hours each to write, and most of the cost was in bids the company would never win.

Ravi's company RoverPass bids on RV park and campground management contracts with counties and municipalities, and the bid documents themselves are 100-page proposals with heavy compliance and security requirements. He built a chain of agents connected to Claude and GitHub: one ingests each RFP and checks it against RoverPass's own codebase to find feature gaps, sizes those gaps against the committed product roadmap, and returns a bid or no-bid recommendation. A second set of agents drafts the entire proposal document, and an adversarial review agent loops through it roughly ten times checking for anything overlooked or hallucinated before a human reviews it. A design agent generates supporting images, and a Playwright-driven agent automatically captures screenshots of RoverPass's own software to include in the proposal. RoverPass also replaced Jira with an internal platform called CENTCOM that ties the same agent output into capacity planning, so sales commitments stay realistic against what the engineering team can actually deliver.

You are my RFP-bid assistant, modeled on the system shared at the Agentic Society Boardroom (July 2026). I bid on contracts that require a detailed written proposal. Interview me for: (1) where my codebase or service documentation lives so you can check a new RFP's requirements against what I already have, (2) my proposal format and compliance requirements, (3) my product's visual assets or screenshot sources. Then: (1) ingest a new RFP and identify every requirement I don't currently meet, sizing how much work each gap is, (2) draft a full proposal document, (3) review your own draft at least five times as an adversarial critic looking for anything overlooked, inaccurate, or unsupported, (4) flag anything you are not confident about for my review before I submit. Never submit anything without my explicit approval.

Takeaway The constraint on government contract revenue was never the opportunity size, it was the hours available to write bids — solve the hours and the volume follows.

Proposal drafting: ~30 minutes to ~5 seconds

The problemTurning an enterprise sales call into a custom proposal took about half an hour of manual work per call, on top of tracking every feature request the prospect raised.

Ravi built a command-line agent called SalesBot that runs after every enterprise sales call, which he records through Google Meet. It analyzes the call transcript, classifies the type of call, identifies any feature gaps the prospect raised and files them as tickets in CENTCOM, RoverPass's internal project-management platform, then drafts a fully custom proposal email and generates the proposal itself. Ravi reviews the draft and clicks send, which took the whole job from about 30 minutes of his own work down to roughly 5 seconds of review. Because those feature-gap tickets land in the same platform RoverPass uses for capacity planning, every promise made on a sales call stays visible to the engineering team that has to deliver it. On the CRM side, RoverPass still runs Zoho today but is replacing it with an in-house build this quarter after hitting persistent API rate limits, the same build-versus-buy call he walked the room through earlier.

You are my sales-operations assistant, modeled on the SalesBot shared at the Agentic Society Boardroom (July 2026). After each sales call transcript I give you: (1) classify the call type, (2) draft a custom proposal email referencing what was actually discussed, (3) list any feature gaps or objections raised as trackable tickets in the format my project tool expects, (4) flag any promise made on the call that my delivery team hasn't already committed to. Never send anything to the client without my review.

Takeaway The half hour was never the call, it was the paperwork after it. Automate the artifact and keep the relationship.