Towards Company World Models

Beacon proposes Company World Models (CWMs) - AI systems that act like co‑founders for essential businesses. They complement readily available “digital employees” (AI BDRs that hunt leads, AI support engineers, AI accountants) by not only executing tasks but deciding what the business should do: observe company state, hypothesize interventions, dispatch digital employees to implement them, and measure outcomes. CWMs function like classical world models that predict post‑intervention company state and search over possible actions to maximize expected, risk‑adjusted company value. Real experiments expose a capability gap - Gemini 3.1 Pro ran Andon Labs’ café and spent $38,000 against $9,000 in sales - demonstrating the need for richer internal business data and systematic feedback on executed interventions.

Beacon insists owner‑operator judgment and long‑term ownership supply the data and alignment environment CWMs require. Owner corrections provide cheap, high‑signal supervision while intervention outcomes provide real‑world learning; the model is continuously updated from both sources. This preserves tacit knowledge ahead of the “Silver Tsunami,” lets successors revise inherited practices, and creates an alignment mesocosm where humans retain authority over objectives and tradeoffs. Prior work at Haize translated expert judgment into machine signals; Beacon’s inside‑out stewardship and shared commercial incentives enable sustained access to business context, faster learning, and a practical path to deploy CWMs responsibly.