Docs / Governing Intelligence in Renly
Governing Intelligence in Renly
How the controls in the Governing Intelligence framework map to what Renly does when your agents and automations run, which plan each needs, and which parts stay with your agency.
New to the framework? Read Governing Intelligence, or take the free AI governance assessment to see where your agency stands and download a draft strategy in Word.
Your agency's part, and Renly's
The framework's five domains are Strategy, Operating Model, Architecture, Design and Governance. Strategy and the Operating Model are decisions people make: the ambition, the risk appetite, the accountable officials, the committee. No software can own them, and Renly does not try to.
Architecture, Design and Governance are where software matters, because they decide what an AI system is allowed to do and how you prove it did only that. That is the work Renly does: it is the governed runtime your agents and automations run on.
The crosswalk
Each row names a control from the framework, where the framework describes it, and what Renly provides. Approvals and the Evidence Passport are on every plan; the rows say where a control needs a particular plan.
| Framework control | What Renly provides |
|---|---|
| Autonomy gatesGo, Ask or Stop. Section 8 | Each of an agent's actions is set to Allowed, Ask first or Never, and Renly enforces the setting where tools run. An action set to Never is hidden from the agent and refused if it is attempted. Sending email, posting and publishing always need a person. |
| Human participation pointsWho decides, with what information. Section 8 | Approval cards in chat, an Approvals inbox with batch decisions, and signed approve and reject links in email. Every approval is single-use, expires (after an hour in chat, after 24 hours for an automation waiting on a person), is bound by a SHA-256 fingerprint to the recipients, amounts, records or command the person reviewed, and records who approved it.Mark an agent, or a use in the AI Use Register, as two-person: its approval-gated actions then need two different people. The person working with the agent approves first; another organisation owner or admin gives the second approval in the Approvals inbox or from a signed-in email link, and only then does exactly the approved action run. The Evidence Passport records both approvers. Register setting: Enterprise |
| Decision passportProvenance for every decision. Section 7 | Every response gets an Evidence Passport: which model actually answered, which tools ran and who approved each write, the policy verdict, and hashes of the prompt and the answer. Records are sealed into your organisation's own hash chain, so tampering is detectable, and each one separates what Renly recorded from what the model says about its own work.Capture is best effort by default. On Enterprise, mark a use in the AI Use Register as Evidence mandatory and it will not run without a record; admins can verify the chain and export evidence for a period. Enterprise |
| Runtime policy enforcementGuardrails as systems act. Section 9 | With the Governed Runtime switched on, an independent policy service checks each chat request before it reaches a model, and each automation step and agent hand-off before it runs. If the policy service cannot be reached, requests are blocked by default; admins can choose to warn or allow instead. Enterprise |
| Kill switchRehearsed, not assumed. Section 9 | Admins can halt every automation and agent invocation in one click: nothing starts, resumes after an approval, or retries until they resume. A run already under way is cancelled from its run page. On Renly Desktop, Ctrl+Shift+Backspace pauses computer control, and any paired device can be revoked remotely. |
| RollbackUndo within the boundary. Section 9 | Code sessions keep checkpoints of the files an agent edits, so you can revert them; a chat can be rewound to an earlier message; and agents keep versions you can restore.Changes an aborted run has already made in other systems are listed for clean-up, not undone automatically. |
| Use-case registerIncluding discovered AI. Section 9 | The AI Use Register lists every active agent, automation and connector, scores inherent and residual risk across eight factors, tracks each use through a seven-stage lifecycle, names its accountable owner, and records the go or no-go decision at each of five gates, with who decided and when. EnterpriseSharing the register with the DTA stays with your agency. Renly exports a disclosure statement for the uses that need one. |
| Assurance before deploymentGates G1 to G5. Section 10 | Write test cases for an agent and run them against real data in a read-only sandbox, with an AI judge. Mark an agent as production tier and every change waits for a different admin's approval after its tests pass; an override needs a written reason, which is audited.For gate G4, an admin can run a pending change in shadow mode: Renly replays a sample of real requests against it in the same sandbox (up to 50 a day by default), the judge compares its answers with the live ones, and the agreement rate appears in the register. Anything the pending version tries to change is listed, never carried out. |
| Third-party AI containmentNo uncontrolled egress. Section 7 | Tools on external MCP servers ask for approval by default, read-only hints count only from servers an admin marks as trusted, and each agent can be limited to named connectors, read only or read and write. MCP: Pro Plus and Enterprise |
| MonitoringDrift, overrides, failures. Section 9 | Stream governed agent actions to an OpenTelemetry endpoint as GenAI traces, carrying hashes rather than content, and audit and evidence events to your SIEM, such as Splunk or Microsoft Sentinel, as signed webhooks. Owners are told when a run fails or an approval expires, and each agent carries a runtime reliability score that flags bursts of failures. |
| Incident managementIncidents and near misses. Appendix K | The AI incident register keeps each incident and near miss with its severity, owner and timeline, linked to the agent, chat, automation run and evidence records involved. Renly raises an incident itself when an agent's reliability drops or someone reports a response, and Sev 1 and Sev 2 incidents notify owners and admins. Record whether the privacy regulator or the DTA must be told, close each incident with a resolution and lessons learned, and export the register to CSV. EnterpriseDeciding whether to notify, and notifying, stays with your agency. |
| Personal informationPrivacy obligations. Section 8 | Before a tool result reaches the model, Renly masks tax file numbers, Medicare numbers and IHIs (checked against their check digits), passport, driver licence and Centrelink numbers when they are labelled, card numbers, credentials and some US formats. ABNs and ACNs are public and stay readable. |
| RecordsDecisions kept as records. Section 7 | Set how long audit logs and chat transcripts are kept (chats are kept forever by default), and export evidence bundles (the records, a hash manifest and an attestation) for a chat or a period. Chats export to PDF. Place a legal hold on one chat or the whole organisation, with a reason: a held chat cannot be deleted by its owner, an admin or the retention schedule until an admin lifts the hold. Pro Plus and Enterprise |
| Staff AI literacyMandatory training. Section 6 | The free Renly Academy teaches staff to use AI and agents responsibly, issues a certificate anyone can verify online, and gives managers a team completion register to export. Use it alongside the training your agency mandates. |
| Identity and accessZero trust, least privilege. Section 7 | People sign in with Microsoft, Google or your own identity provider over SAML 2.0, such as Microsoft Entra ID or Okta, for email domains you prove you own with a DNS record. Once a test sign-in passes, require single sign-on for your organisation; SCIM 2.0 provisions and deprovisions people from your IdP. SAML: EnterpriseSign-in inside the Teams tab and the Azure DevOps extension does not yet apply a required sign-in provider. |
| Sovereign foundationWhere data and models run. Section 7 | Renly's databases and file storage run on Microsoft Azure in Australia East. Model calls go to the provider of the model you use, which may process them outside Australia. On Enterprise you can route model calls to your own provider (BYOLLM) and wrap stored credentials with your own Azure Key Vault key. BYOLLM and BYOK: Enterprise |
What stays with your agency
Renly records decisions; people make them. These parts of the framework are yours, and no platform can carry them for you:
Decide and appoint
- Endorse the strategy and the risk appetite (Section 5).
- Designate the accountable officials, the Chief AI Officer and each use-case owner, and tell the DTA (Section 6).
- Run the AI governance committee and make the gate decisions (Sections 6 and 10).
Assess and report
- Complete impact assessments with the DTA's tool, or an equivalent, before deployment.
- Publish the transparency statement and explain automated decisions in your privacy policy.
- Run the incident process and decide on notifications (Appendix K). Renly keeps the register, raises incidents from what it sees and supplies the evidence.
Not in Renly today
Every control in the crosswalk is built into Renly. Where one stops short, its row says so, as with clean-up after an aborted run, two-person approval of steps that never ask for approval, and sign-in inside Teams. We would rather you hear it from us than find it in a security review.
This page was last checked against the product on 9 October 2026. Questions from a security or procurement review? Email support@renly.ai.