Organization as intelligence

Replace hierarchical information routing with shared world models and an intelligence layer — humans at the edge for judgment and trust.

mine

written here

Concepthandbookproduct-buildingteamsorganization

Spin-out from AI-Native Product Building. Informed by Block — From hierarchy to intelligence.

Claim

Most companies bolt AI onto hierarchy — copilots in Slack, summaries in meetings, the same handoff graph with faster typing. Block's bet is different: build the firm as intelligence — a layer that routes context and composes capabilities the way middle management used to, with humans at the edge for judgment, trust, and ethics.

Span-of-control limits shaped org charts for centuries because nothing else moved information fast enough. When work artifacts and customer signals are machine-readable, coordination becomes a product problem — not a meeting problem.

The decision

Invest in org-as-intelligence when coordination cost (meetings, status, re-explaining context) measurably blocks shipping — and you can feed models legible work signals, not politics in prose.

Keep explicit hierarchy and owners when accountability must be auditable (regulated data, payments, safety) or when the team is small enough that one DRI and AGENTS.md already beat a world model.

Never confuse copilot rollout with org design — summarizing the standup does not remove the standup.

Roles at the edge

RoleOwns
ICDepth in a capability
DRIOutcome window — pulls context, not headcount
Player-coachCraft + people at the boundary

Agents inherit the same boundaries as humans: forbidden zones, review gates, incident updates.

Pressure-test

  • Is your "intelligence layer" portable — or does it require transaction telemetry only Block-scale firms have?
  • When composition fails, is that a roadmap item or a strategy hole?
  • Who is accountable when a DRI pulls the wrong capability from the model?
  • Does "no permanent middle management" survive regulatory surface area?
  • Is customer context in the world model consent-safe when inference errors?

Related