Control before execution
How it works
Each AI request follows one path. SentientROUTER checks the request against policy and human authority, gives it a class and controls what happens next.
The authority path
One request. One check. Three routes.
The check comes before the action. Without a permitted route, a request does not reach execution.
- SPermitted automationOnly inside the approved scope.
- DDraft for human reviewNo autonomous final action.
- HHuman approvalExecution waits for an authorised person.
Six steps
What happens to a request
The steps follow in a fixed order. Each organisation configures the rules, the roles and the integrations for its use case.
Capture the request
SentientROUTER identifies the proposed action, its source, the data that it needs and the context of the organisation.
Apply policy
It checks the request against the current version of the rules and against the delegated authority. The check runs before any autonomous action and before any external source is queried.
Choose a route
It returns class S, D or H, with the reason and the rule that applies. Uncertain or incomplete context moves the request to a stricter class.
Gate execution
In class S it permits an action inside the set limits. In class D it prepares a draft. In class H it waits for a person who holds the necessary authority.
Record the outcome
It keeps the class, the approvals and the execution context for review by the organisation. It passes them to a proof layer for later checks.
Review and adjust
The organisation assesses incidents, policy changes and behaviour before it widens any permitted autonomy.
Worked example
A credit request that waits for approval
A bank uses an AI assistant in agricultural credit work. The assistant receives a request to approve a loan.
The policy of the bank puts loan approval in class H. SentientROUTER returns route H. The system does not act alone, and the request waits for the person who holds the authority.
request: agricultural_credit_approval
policy: institution_policy_v1
route: H
autonomous_action_permitted: false
human_approval_required: true
evidence_state: pending_verificationHow to read the result
- request
- The proposed action: approval of an agricultural loan.
- policy
- The version of the rules that the check used.
- route
- The class of the request: H, human approval.
- autonomous_action_permitted
- False. The system does not act alone.
- human_approval_required
- True. An authorised person approves or declines.
- evidence_state
- The record of the event waits for verification.
Before it uses the workflow, the organisation defines the policy, the threshold, the authorised role and the evidence requirements.
Decision classes in detail