For consequential actions, governance becomes executable
When a governed agent proposes an action, Cogna8 checks the current state, the controls and any approval required, then allows or blocks it before it runs. This is the part most governance tools stop short of.
Northwind Mutual ⌄Console › Authorisation › DecisionsSearch ⌘KProductionModelPolling
Decisions
Every action evaluated by the gate before it ran.
Evaluated3,9129 governed agents
Needs approval124 waiting over 1h
Blocked3721 state conflicts
Decisions per hour
AllowedNeeds approvalBlocked| Time | Agent | Action | Decision |
|---|---|---|---|
| 14:23:07 | Claims triage agent | Release settlement payment, $48,200 | Blocked |
| 14:22:41 | Claims triage agent | Request missing document | Allowed |
| 14:22:12 | Broker email drafter | Send renewal terms to broker | Approval |
| 14:21:58 | Customer reply assistant | Close complaint ticket | Allowed |
| 14:21:30 | Fraud signal model | Flag claim for investigation | Allowed |
| 14:20:12 | Claims triage agent | Update claim reserve, $120,000 | Approval |
| 14:19:47 | Underwriting risk summariser | Publish risk note to policy file | Blocked |
Decision rcpt_7Q2K9F
BlockedClaims triage agent, 14:23:07 AEST
- Action proposedRelease payment of $48,200 on claim 88214
- State checkedClaim amount conflicts: $48,200 and $41,750
- Controls evaluatedOR-15 Human approval above threshold, CPS 230
- Approval requiredClaims Ops Lead, requested 14:23:08
- DecisionBlocked until the amount is resolved and approved
- Policy
- payments.clean_state v3
- Replay
- Same inputs, same decision
Export receiptOpen approval
Decisions
Three outcomes, always explained
| Outcome | When | What happens next |
|---|---|---|
| Allowed | State is clean and policy is satisfied | The action runs; the receipt is kept as evidence |
| Needs approval | Policy requires a person, for example above a payment threshold | The request is routed to the named approver; the agent waits |
| Blocked | Required state is missing or in conflict, or policy forbids the action | The action does not run; the reasons are returned to the agent |
Deterministic by design
The final decision does not rest on a model judgement. The same inputs always give the same verdict, so any decision can be replayed for an auditor or a regulator.
- Gate policies derived from your controls, versioned and approved
- Conflicts between sources caught before they become actions
- Approval routing to named roles
- Append-only receipts: what was asked, what was known, which rule applied
Let's connect
Tell us where your AI programme is today. We start with a short, scoped pilot on your own AI estate, and every enquiry is handled in confidence.