Set how far each agent may act on its own, the policy limits that decide straight-through processing, and the kill switch that stops everything. The evidence and the audit trail sit alongside.
When on, no agent may start a run or apply an action. Pending approvals cannot be executed.
An agent only moves up a level when its evaluation evidence supports it. Level 3 and above can act straight through, and only inside the policy limits below.
| Agent | Owner | Autonomy | Active | Runs 7d | Override 7d |
|---|---|---|---|---|---|
| Loading… | |||||
A run is auto-approved only if the kill switch is off, the agent is at level 3 or above, confidence and amount are inside these limits, the decision is valid and the primary evidence was found. Everything else waits for a person.
AML dispositions are a regulatory decision. The AML limit is locked at 0, so every alert disposition is approved by a named investigator.
Whatever the amount, a trade presentation refused under UCP 600 art. 16 needs an authorised checker, a request to return funds on a correspondent payment needs an operations approver, and no CECL allowance draft for a respondent bank is ever approved or posted without a person.
Computed from every logged run: how often reviewers agree with the agent at each confidence level, where they override it, and how often guardrails fire.
| Confidence | Runs | Auto | Approved | Rejected | Agreement |
|---|
| Agent | Level | Runs | Override rate | Failed | Avg tools |
|---|---|---|---|---|---|
| Loading… | |||||
Open any run for its step-by-step trail: plan, tool calls, observations, guardrail checks and who decided.
| # | Started | Agent | Case | Status | Conf. | Amount | Decided by | Summary |
|---|---|---|---|---|---|---|---|---|
| Loading… | ||||||||
We build the control framework with your risk, compliance and operations teams before any agent handles a live case.
Each agent is inventoried, documented and validated like a model, with an owner, intended use and limits aligned to your MRM policy.
Confidence, amount and case-type thresholds decide what a person must approve; regulated decisions such as AML always stay with people.
Agents propose, policy routes, named reviewers approve. The agent that prepares an action never approves it, and tool access is least-privilege.
Models run in-region; personal data is masked before prompts leave the data fabric, and only governed data products are exposed as tools.
Agents start at "suggest" and move up only on calibration and override evidence from shadow runs, reviewed with the business owner.
Every plan, tool call, observation, guardrail check and decision is logged and replayable for internal audit and regulators.