最終更新: 2026/06/22 19:15
commonid: rq-107-organizational-decision-roles-p…type: promptstatus: active

Research Question

In modern organizations, what is the standard decomposition of "roles" involved in making, recording, and re-using strategic/architectural decisions, and how does that decomposition compress when the executor pool shrinks to a single human plus AI agents?

[Note: This is a study of role taxonomy (who does what in a decision lifecycle), NOT a study of decision evaluation criteria (which axes to score on — already covered in our prior RQ on MCDA). The deliverable feeds the "Job executor" section of our JTBD for an internal decision pipeline.]

Context (≈400 words)

We operate a Decision Pipeline for Architecture Decision Records (ADRs) in a small-business accounting product. Monthly volume is 5–15 ADRs. The operator today is one founder-CEO plus several specialized AI agents (Claude Code clones scoped to drp / main / doc domains). A junior engineer joins in 2026-10. The pipeline runs 13 LLM-backed gates that absorb most of the "review" labor (triage, problem-worth screening A/B/C, cost gate, pre-gate, socratic critique, scoring, cross-validation, consistency, parallel review, policy alignment, body generation, slug, numbering), leaving only "final Accept / Reject" and "implementation" with the human.

We have already identified ~19 candidate roles internally (proposer, multiple flavors of reviewer, decision-maker, recorder, implementer, compliance checker, reader, retroactive updater). Most reviewer roles are filled by AI gates. The decision-maker role is filled by the human alone. We suspect three structural risks:

  1. Single-actor concentration on final Accept makes 4-eye principle / segregation-of-duties impossible.
  2. Several roles are unspoken (compliance check is done ad-hoc; no explicit "recaller" actor).
  3. We have no industry baseline to tell us which roles SHOULD exist before we redesign.

We need to learn how large/mid organizations decompose these roles, how the decomposition scales DOWN to small teams, and how multi-agent AI systems formally position AI as a role.

Questions (4)

  1. Standard role decomposition in large organizations: How do mature frameworks (DACI, RAPID, RACI, Bezos Type-1/Type-2, Holacracy, Sociocracy 3.0, Spotify model, J-SOX internal-control 4-step, OECD CG Principles) decompose the lifecycle of a strategic decision into named roles? Provide a normalized cross-framework role table (role name, responsibility, deliverable, separation requirement, typical handoff to next role). Identify roles that appear in 3+ frameworks (= robust) vs roles that appear in only 1 (= framework-specific).

  2. Variation by org size and industry: How do role count, specialization, and delegation patterns shift across startup (<50) / mid-cap (50–500) / large enterprise (500+) / multinational (5000+)? And across SaaS / SI / finance / manufacturing? Highlight where audit-driven role separation (SOX 404, J-SOX, ISO-27001) hardens specific separations (e.g., proposer ≠ approver). Where can a solo operator legally / safely collapse roles, and where can they NOT?

  3. Embedding AI agents as formal decision roles: In multi-agent systems (Anthropic, Cognition Devin, LangGraph multi-agent, AutoGen, CrewAI, OpenAI Swarm) and emerging governance literature, what patterns exist for assigning AI to proposer / reviewer / compliance-check roles while reserving final Accept for humans? What controls prevent AI from crossing into self-review (proposer = reviewer = a single agent), and what audit-trail conventions track AI-vs-human contribution?

  4. Compression strategy for n=1 + AI: Given the role table from Q1, which roles can a single human + a multi-clone AI safely cover, which roles need explicit compensating controls when collapsed, and which roles CANNOT be collapsed (must be staffed before audit / regulatory exposure)? Provide a decision matrix keyed on org-stage and audit-exposure.

Output

Structured report with:

  • Executive summary (3–5 key findings)
  • Per-question analysis with citations
  • Normalized cross-framework role table (Q1)
  • Org-size × industry variation matrix (Q2)
  • AI-as-role pattern catalog with anti-patterns (Q3)
  • n=1 + AI compression decision matrix (Q4) — must-staff / safe-to-collapse-with-controls / safe-to-collapse-without-controls
  • Priority ranking of recommended role additions / consolidations for our pipeline (must-have / should-have / nice-to-have)
  • References with URLs