AI Control Envelope

Problem addressed

A model metric alone does not control data, prompts, retrieval, permissions, human reliance, or operational drift.

Limitation of existing practice

Point-in-time model testing cannot establish continued fitness across changing context.

Proposed method: inputs and steps

1. Define approved users, decisions, inputs, and prohibited uses.
2. Set data and performance boundaries.
3. Specify deterministic guardrails and human authority.
4. Define monitoring thresholds and response actions.
5. Control model, prompt, retrieval, and tool changes.

Decision outputs

• Approved operating envelope
• Control and monitoring map
• Restriction and rollback criteria

Worked example

An AI complaint classifier may recommend a category but cannot set seriousness, submit the record, or operate below a subgroup recall threshold.

When to use—and when not to

Use when: Assuring probabilistic or AI-enabled functions.

Do not use as-is when: The function is entirely deterministic; use conventional boundary and control methods.

Limitations

Thresholds need representative data and ongoing review.

Inspection questions

• What use is prohibited?
• Which threshold triggers restriction?
• Can the human reviewer detect a plausible error?