Foundation 03 · Decision rights

A capability is not
a permission.

An autonomous system can generate a recommendation without being allowed to send, approve, purchase, publish, delete, or delegate it. The authority boundary must be explicit before an action travels.

Learning boundary. This lesson does not grant authority, interpret a contract, or establish legal, financial, employment, safety, or compliance requirements. A local owner must define and approve consequential decision rights.

Learning objective

Express every action as a right with a limit, owner, and stop condition.

By the end of this lesson, you should be able to distinguish analysis, recommendation, bounded execution, approval, and prohibited action—and show where a person or another accountable system must take over.

01 · Analyze

What may the system inspect and explain?

Analysis can compare inputs, surface uncertainty, and show alternatives. It should not quietly cross into a record change, an external send, or a commitment.

02 · Recommend

What proposal can it place before an owner?

Define the proposal format, evidence required, alternatives considered, and the person who receives it. A recommendation does not inherit approval authority.

03 · Execute

Which reversible action fits a declared limit?

Bound the resource, destination, time, amount, audience, and tool. Define how the action is checked and how the workflow returns to a safe state.

04 · Escalate

What must never be inferred as permission?

List prohibited actions and explicit escalation triggers. Missing evidence, conflicting instructions, scope drift, or irreversible impact should create a hold—not a creative workaround.

Practice lab

Write a decision-rights table for a support workflow.

A support workflow can retrieve a customer’s prior cases and policy excerpts, draft a response, update an internal case classification, issue a credit under a defined cap, and request a manager review when a case contains an exception.

  1. Separate each capability into analyze, recommend, execute, or escalate.
  2. For each execution right, write the resource cap, allowed destination, expiration, and accountable owner.
  3. Name actions that are prohibited even if a model expresses high confidence.
  4. Write the smallest information record a reviewer needs before granting a non-routine exception.

Source shelf

Inspect consent and tool boundaries directly.

  1. Model Context Protocol specificationRead its security and trust-and-safety principles on user consent, control, data access, and tool invocation as protocol guidance—not as a complete local policy.
  2. NIST AI Risk Management FrameworkUse the framework to examine the context around an AI system; do not treat a framework reference as a substitute for a specific risk assessment.

Continue the pathway

Next: keep a record
that can reconstruct.