LEARNING LAYER

AUTHORITY & HUMAN RESERVE

OLD QUESTION

Who owns the task?

NEW QUESTION

Who may commit the organization?

CORE CONCEPT

Capability is not authority.

CORE

Technical capability does not automatically create social permission. A system may draft, speak, negotiate, or bind. Whether it may do so is an organizational decision, and sometimes a legal and cultural one.

The useful scale is not adoption. It is permission:

RESEARCH ONLY — gather and analyze. DRAFT & RECOMMEND — prepare; a human still releases. CUSTOMER INTERACTION — speak as the firm. NEGOTIATION — alter terms inside a boundary. COMMIT — finalize a commercially binding act.

Human reserve names the last of these as a political fact: some decisions remain human because the surrounding world still requires a person, not because the model is weak. Enterprise pricing, customer termination, and legal commitment can be technically automatable and still reserved.

Accountability is a separate axis. A person can be accountable for a communication they did not write. A team can own a model whose speech lands on an account manager's name. Shared accountability can look adult until a disputed clause needs a single signer.

Override is the test of the design. If only a rule engine can halt a commitment, interruption is no longer a human act. If an account manager may halt it, that manager becomes the last fuse.

The general manager in this case does not hold CEO authority. Persuasion, delegation, and boundary-setting are the work. Decision is not action. Action still depends on support, capacity, customers, employees, the board, regulation, and the system's own behavior.

IN THE SIMULATOR

The first decision is the spine of the case: how much authority AI sales agents receive. Later decisions do not undo it; they qualify it.

Approval rules, override rights, disclosure, and the final human-reserved list are attempts to place permission after capability has already been extended. Q4 asks whether you still maintain the boundary you set in Q1.

ORACLE-3 has analysis and no formal authority. That is intentional. A probability is not a permission.

REFLECTION

Who, by name or by role, may still say no—and will that no still work when the file is already committed?

← ALL LAYERS