The hard boundary
Agents can execute work. They cannot own professional consequences.
If a recommendation is wrong, a customer message is misleading, a report invents evidence, a code change breaks production, or an automation damages trust, the operator owns the consequence.
"The agent did it" may explain the process. It does not transfer responsibility.
Assistance is not authority
The individual operator should make a bright-line distinction between assistance and authority.
Agents can help draft, compare, summarize, inspect, test, and propose. They should not silently make commitments, change permissions, publish externally, contact people, alter durable records, or take irreversible actions unless the workflow has explicit human-approved gates.
This is not anti-agent conservatism. It is professional accountability.
The accountability boundary map
For each workflow, classify actions into three zones:
Free
Low-risk, reversible work inside a bounded workspace.
Examples: draft a memo, summarize provided notes, propose an outline, inspect a local file, generate test cases, identify contradictions, create a local checklist.
Review-required
Work that influences other people, durable records, strategic decisions, or operational state.
Examples: edit canonical documentation, update trackers, merge code, change a source-of-truth note, prepare a customer response, publish a draft to a CMS, recommend a financial or personnel decision.
Forbidden without explicit approval
Destructive, external, privacy-sensitive, permission-changing, legally sensitive, or outside-the-brief actions.
Examples: send external messages, delete data, change access, execute financial commitments, contact customers, publish publicly, scrape private information, rewrite history, or act in a way that creates obligations for someone else.
The boundary should live in the brief, not in the operator's memory.
Accountability includes taste and context
Accountability is not only about preventing obvious harm. It also includes the quality of judgment.
The operator owns whether the sources were sufficient, whether the output fit the audience, whether the recommendation named the real tradeoff, whether the review was serious, and whether the work reflected the standards they claim to hold.
You cannot delegate taste and then complain that the work lacks taste.
Failure mode
The dangerous pattern is subtle: the agent produces a confident artifact, the operator feels emotionally relieved, and the boundary between "generated" and "endorsed" blurs.
But the outside world only sees the endorsed artifact.
The individual operator's signature is the moment they allow work to affect reality. That signature should mean something.
