Human approval is useful when a person receives enough information to make a real decision and software enforces the result. A button labelled “approve” does little if the action has already happened or the proposal changes afterwards.
Design the approval point around the consequences of the work. Identify who decides, what they see and which exact action their decision authorises.
Place review before the consequential action
Separate a draft or proposal from execution. The application can gather information and run checks before presenting a proposed change to the authorised reviewer.
For example, an illustrative finance workflow might prepare a classification with source references and validation results. A qualified person reviews that proposal before it is used in the controlled process. The example describes a design pattern, not an automated professional decision.
Some routine tasks can follow approved rules without a person checking every event. The appropriate boundary depends on the task, applicable requirements and the organisation’s policy. Do not infer it from the model’s confidence alone.
Bind approval to the proposal
Store the proposed action and its version. Record the named approver and the decision they made. If material details change, require the review appropriate to the new proposal.
The interface should show the relevant source, intended recipient or target record, changed fields and the results of software checks. Avoid asking someone to approve an unexplained summary while hiding the consequential parameters.
Approval should be enforced where the application authorises execution. A prompt telling the model to wait is not an equivalent access-control boundary.
Define who can approve
Use the organisation’s identity and role arrangements. A service credential can let software perform a permitted operation, but that credential is not a named person’s professional sign-off.
Check the approver’s authority at the time of the decision. Plan for revoked access, unavailable reviewers, rejected proposals and requests that need more information.
Keep the responsibility visible to the business. A technical team should not silently become the approver for decisions that belong to a qualified professional or process owner.
Do not turn confidence into permission
A model saying that it is highly confident does not establish the probability that an action is correct. Scores used for routing need a defined method and validation on representative work.
Software can check a fixed category list, required fields, allowed ranges and reproducible calculations. Those checks support review but do not eliminate interpretive errors. A syntactically valid proposal can still apply the wrong source or rule.
Use the failure cases to decide where a task should stop or escalate. An honest request for clarification is often more useful than a complete-looking answer based on missing evidence.
Design for the reviewer’s workload
A person can make mistakes or become overloaded. Test whether the review screen makes the important differences and uncertainties easy to inspect.
Track correction patterns and rejected proposals. A declining review rate is not automatically evidence of improvement; it might reflect workload, habit or a changed mix of tasks.
Avoid promises that an approval mechanism is impossible to bypass. Test the intended paths, enforce permissions at appropriate boundaries and assess remaining risks. OWASP’s excessive-agency guidance is a useful reference for high-impact tool actions.
Record the observable decision
Keep the proposal, relevant checks, approval and outcome linked under suitable access and retention controls. These records support investigation and review. They do not reveal a complete causal account of a model’s internal reasoning or automatically prove legal compliance.
The Ryzolv Harness Rules set out our approach to software validation, traceable calculations and named-human decisions. The implementation still needs to be matched to the work and tested in the actual system.
Written by Anshul Kaushik, Managing Director at Ryzolv.
Back to Insights