Every BCG unit is operated by the same three-layer stack. No unit has full-time employees. Every workflow that used to demand an operations hire has been decomposed into three roles: brain, hands, eyes. Each role is owned by a specific AI system, because each system is disproportionately good at one part of the job.
Layer 01 — BRAIN. ChatGPT Projects holds strategy, memory, and context. Every unit has its own Project with the founding thesis, the compliance boundaries, the current roadmap, and the operating decisions made in every prior week. When a new question arrives — should we price this bundle at $79 or $99? should we list this lender as tier-1? — the Brain does not answer from scratch. It answers with the full context of what the unit has already decided, which keeps the unit consistent over time without a human memory in the loop.
Layer 02 — HANDS. Manus executes. When the Brain produces a plan — update the directory, draft the outreach email, compile the quarterly rate-band report — Manus is the layer that actually opens the tools, produces the artifacts, and follows through. This is the layer that used to be an operations hire. It is now an agent that can run a multi-hour job without supervision and produce the file, the draft, the row in the spreadsheet.
Layer 03 — EYES. Claude reviews. Nothing ships until the Eyes layer signs off. For code, Claude reads the diff and verifies correctness against the stated intent. For content, Claude flags factual claims that need sourcing. For compliance, Claude enforces the is/is-not boundaries declared on each unit's page — if the draft reads like regulated advice and the unit is not licensed to give it, the Eyes layer kills the draft before it ships.
The loop is tight. Brain strategizes. Hands executes. Eyes reviews. If Eyes rejects, Hands iterates or Brain re-plans. If Eyes approves, we ship. The whole thing runs as code — no human sits in the middle unless the Eyes layer kicks the decision upstream.
Why this works: a ten-company portfolio is too much surface area for a founder plus a handful of operators. It is not too much surface area for a founder plus a disciplined three-layer AI stack. The cap exists because humans cannot supervise more than ten units; the stack exists because ten units cannot be operated by humans alone.
This is not generative AI as feature. It is generative AI as the operating company. The difference is whether you have replaced workflows or replaced the chairs the workflows used to sit in. BCG has replaced the chairs.