liveBook · Manning

Tiered approval

As explained in AI Governance

Tiered approval is an authorization scheme that applies progressively stronger review requirements according to an action’s impact or risk.

An AI assistant that can update records or initiate transactions needs more than a generic confirmation dialog. Tiered approval matches the review required before an action to its impact, so consequential actions receive stronger scrutiny than routine ones. section 4.7.1

A tiered workflow can distinguish three levels of impact:

  • Low impact: For viewing data or minor updates, display the interpreted action and proceed after a brief timeout, allowing the user to intervene. section 4.7.1
  • Medium impact: For scheduling or standard updates, display key parameters and require a single confirmation. section 4.7.1
  • High impact: For financial transactions, bulk operations, or irreversible changes, require individual verification of each critical parameter. section 4.7.1

Confirmation is not a substitute for execution checks. Before approving a status change, verify required documentation and automated approval thresholds. Before issuing a refund, check that the original transaction exists and has not already been refunded. section 4.7.1

Claim updates with approval thresholds

ClaimAssist binds write actions to the authenticated user: customers can affect their own claims, while adjusters can affect assigned cases. For routine updates, it displays the intended action and waits for confirmation. Actions above automated approval thresholds require a human reviewer, regardless of the conversation. section 4.9

Some operations should be excluded entirely, such as changing contract terms or deleting records without recovery options. Where exclusion is impractical, a human can approve before execution or monitor the system in real time and intervene. Explicit boundaries and narrowly scoped credentials help limit what a manipulated assistant can do. section 4.7.1 section 4.9

How AI Governance covers it

How deep it goes
Details impact-based confirmation levels, pre-execution checks, human oversight, and operation exclusions.
Examples
Uses claim updates, appointment scheduling, financial transactions, and refunds to illustrate review decisions.
What you can do after
Evaluate and apply review requirements for AI actions based on impact and approval thresholds.
How it connects
Connects approval to prompt injection, authorization, human-in-the-loop oversight, and pre-execution validation.

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.