THE ROLE DIVISION MATRIX

Human-machine workflows fail when roles are implied instead of assigned.

A workflow diagram may show a human, an AI system, a tool chain, a review step, and an output. But the diagram often leaves the important verbs unassigned: who directs, who plans, who executes, who supervises, who reviews, who refuses, who escalates, who verifies, who repairs, and who owns consequence.

The Role Division Matrix exposes the difference between a workflow that merely contains humans and machines and a workflow that assigns authority.

Field rule: No delegation without role split.

The broken loop begins when work is delegated without role clarity.

AI systems draft, rank, summarize, code, route, retrieve, recommend, edit, plan, and trigger downstream action. That creates a role problem.

A human may believe they are directing while the system is planning. A reviewer may believe they are supervising while the system has already executed. A manager may believe they own the workflow while a vendor system controls the decision path.

Icons are not roles.

A role requires a verb, authority, visibility, evidence, and consequence.

The Role Division Matrix asks:

Who is actually doing each part of the work?

Core verbs:

• Direct — define objective, boundary, and success.
• Plan — choose method, sequence, tools, and fallback.
• Execute — perform the action or change state.
• Supervise — monitor live work before consequence locks.
• Review — evaluate output against criteria.
• Refuse — stop, pause, reject, narrow, or reverse action.
• Escalate — move uncertainty or failure to higher authority.
• Verify — confirm result against independent evidence.
• Repair — fix consequence and prevent recurrence.
• Own — carry explanation, accountability, and closure.

The first audit question:

Which verbs are assigned, and which are only assumed?

Use the Role Division Matrix when:

• a workflow adds AI assistance;
• an agent is given tool access;
• a human review step is introduced;
• a human review step is removed;
• code, content, decisions, or records are AI-generated;
• a process involves vendors, platforms, or external models;
• a failure occurs and everyone can plausibly point elsewhere.

This matrix is especially useful before automation expands from drafting into action.

Drafting can tolerate looser boundaries. State change cannot.

5.1 Role collapse

One person is nominally responsible for too many functions without matching authority.

5.2 Role disappearance

A function disappears after automation is introduced.

5.3 Responsibility routing

Responsibility is assigned to a visible human while practical control lives elsewhere.

5.4 Role ambiguity

Multiple actors participate, but none has named ownership.

5.5 Role inversion

The human is treated as supervisor while the system directs the workflow.

Before delegating work to AI, answer:

1. Who defines the objective?
2. Who chooses the method?
3. Who performs the action?
4. Who watches live state?
5. Who reviews the result?
6. Who can refuse before consequence locks?
7. Who can escalate uncertainty?
8. Who verifies against independent criteria?
9. Who repairs harm or error?
10. Who owns recurrence prevention?

If the answer is “everyone,” the answer is probably no one.

Purpose

Create a structured record of who performs each control verb in an AI-assisted workflow.

Use when

• deploying an AI-assisted process;
• assigning agent permissions;
• reviewing human-in-the-loop claims;
• auditing workflow failures;
• reducing or removing a human checkpoint;
• documenting governance before implementation.

What it creates

A registry that maps each workflow action to a role, actor, authority level, evidence source, refusal path, repair owner, and recurrence owner.

Technical version

workflow_id: "customer-support-refund-triage"
workflow_owner: "Operations Lead"
risk_tier: "medium"
last_reviewed: "YYYY-MM-DD"

delegated_action:
  action_name: "refund eligibility recommendation"
  action_class: "recommendation"
  downstream_consequence: "customer refund approval path"

roles:
  direct:
    actor: "Refund Policy Owner"
    authority: "sets policy and eligibility criteria"
    evidence: "current refund policy"
  plan:
    actor: "AI workflow"
    authority: "selects triage sequence within approved policy"
    evidence: "workflow configuration"
  execute:
    actor: "AI workflow"
    authority: "generates recommendation only"
    evidence: "action log"
  supervise:
    actor: "Support Team Lead"
    authority: "monitors exception queue"
    evidence: "live dashboard + sampled logs"
  review:
    actor: "Support Agent"
    authority: "reviews recommendation before final action"
    evidence: "case record + model rationale + policy excerpt"
  refuse:
    actor: "Support Agent"
    authority: "rejects recommendation or escalates"
    evidence: "refusal reason field"
  escalate:
    actor: "Support Agent"
    authority: "routes disputed cases to Policy Owner"
    evidence: "escalation ticket"
  verify:
    actor: "Quality Analyst"
    authority: "samples closed cases"
    evidence: "QA checklist"
  repair:
    actor: "Customer Resolution Team"
    authority: "corrects outcome and issues remedy"
    evidence: "repair record"
  own:
    actor: "Operations Lead"
    authority: "owns recurrence prevention"
    evidence: "monthly failure review"

Manual / no-code alternative

Use a spreadsheet with columns:

Power-user alternative

Maintain a workflow registry in version control and require updates before permissions change, automation expands, review steps are removed, or new downstream systems are connected.

Output

A clear role map showing whether the workflow has actual control assignment or only implied oversight.

Failure prevented

Implicit delegation, role collapse, responsibility without authority, and post-failure blame routing.

“The human is still responsible.”

Maybe.

But responsibility without assigned verbs is not governance. It is exposure.

The better question:

Which human, doing which verb, with what authority, before what consequence?

This report uses DFEI diagnostic synthesis. It should be supported by source lanes on human oversight, accountability attribution, agentic delegation, auditability, and workflow governance.

Claims about specific legal responsibility should be avoided unless externally sourced and jurisdiction-specific.

Use “ownership,” “repair responsibility,” and “deployment power” as operational terms, not legal conclusions.

Objective: Assign human and machine roles before delegation. Relevant finding: AI workflows fail when role verbs are implied rather than assigned. Recommended execution output: role registry / workflow audit / permission review. Constraints: match authority to consequence; do not assign responsibility without refusal or repair power. Suggested first action: choose one AI-assisted workflow and fill out the Role Division Matrix before expanding automation.