# Give AI a boundary. And an owner.

Decide what it may do, what it must explain, and when a person takes over.

Source: https://kimberlytegroup.com/perspectives/give-ai-a-boundary-and-an-owner/

A KimberLyte perspective.

## Useful autonomy has edges.

An AI system needs an explicit job. Which sources may it read? Which actions may it take? What requires approval? Where should it stop?

These are operating decisions. Make them before an impressive demonstration turns into a process that people depend on.

## Keep the evidence close to the answer.

For consequential work, an answer should carry enough context for someone to challenge it: the source, the definition, the relevant time period, and any material uncertainty.

Access should follow the person and the purpose. An assistant should not become a new route around the controls your organization already relies on.

## Design the handoff.

Name the person responsible for the outcome. Give them a way to review, stop, and correct the system. Test ordinary cases, edge cases, and failures.

Our standard is practical: can the people responsible understand what happened, decide whether it was acceptable, and intervene when it was not? That is the foundation for trust.

## Take it into the room

When access changes, what happens to the answers already in circulation?

Prepare a concise working note for an internal discussion. Use only information I am authorized to provide in this tool. Do not ask for confidential records or production access. If I have not provided enough context, ask for a sanitized description of the decision and pause; do not invent an organization, facts, results, or owners.

THE QUESTION
When access changes, what happens to the answers already in circulation?

CONTEXT TO ADD
[A sanitized description of our decision, current process, and known constraints.]

AUDIENCE
Write for the accountable sponsor. Lead with the decision to make, the consequence of waiting, and the evidence needed. Keep technical detail only where it changes that decision.

EXAMINE
1. Which sources, derived summaries, caches, and downstream actions participate in this workflow?
2. What policy governs reuse after access changes? Distinguish requirements we know from those that need a decision.
3. How does a permission change reach outputs already produced, and what remains unverified?
4. Who can pause the workflow, review affected outputs, and decide whether to resume?
5. What test would demonstrate the agreed behavior without exposing restricted information?

RETURN
A permission reversal test: source and derived outputs; agreed reuse policy; access-change event; expected behavior; evidence to capture; reviewer; resume conditions; unresolved policy decisions.
Use a compact table. Label each material statement as supplied fact, inference, or unknown. Link facts to the context or evidence I supplied; do not invent citations. For each unresolved item, name the role that should answer it and the smallest evidence request that could resolve it. Finish with one decision for the next meeting and what would change the recommendation. This is a working hypothesis for human review, not a finding about our organization.

Prompt adapted from KimberLyte Working tools: https://kimberlytegroup.com/working-tools/
The website's scenarios are fictional illustrations, not evidence about our organization.