Autonomous Governance Agents
Governance work is mostly watching: watching for new AI tools, stale evidence, supplier gaps, and framework drift. Autonomous Governance Agents do that watching continuously and bring you decisions instead of alerts. Agents propose; people decide; nothing changes silently.
The proposal model
An agent never edits your governance data on a hunch. It writes a proposal: what it wants to change, which record it affects, how confident it is, and the reasoning behind it. You approve or reject. Once acted on, the proposal also records who decided and when — so an assessor sees an automated recommendation with a named human accountable for the outcome, which is exactly what every AI governance framework asks for.
The agents
- Shadow AI Triage Agent
- Runs daily. Clusters unreviewed Shadow AI detections, matches them against the AI catalog, and drafts an AI System or Use Case intake for the tools that warrant one — so discovered tools become governed tools instead of a growing backlog. Microsoft agents found by Microsoft 365 discovery that are not yet in the register each get a proposal too; approving it drafts an AI System intake, and RAIC links the agent to that system once the intake is approved.
- Evidence Freshness Agent
- Runs weekly. Finds controls whose evidence is stale or missing, drafts requests to the owners, and proposes status changes — before an auditor finds the same gaps for you.
- Supplier Document Intelligence
- Runs weekly. With "Read supplier documents with AI" turned on (in a supplier's Files panel; off by default), it reads each uploaded PDF when it is uploaded, page range by page range, and records the report type, dates, audit opinion, certifications, subprocessors and retention terms. It then adds certifications the documents show and flags gaps for review: missing data processing agreements, missing or out-of-date assurance reports, audit exceptions, documents about another company, subprocessors not on the supplier's record, and retention terms that differ from it. With AI reading off, it goes by file names and descriptions only. Reading sends the PDF to Google's Gemini API, which does not train on it and keeps requests for up to 55 days for abuse monitoring.
- Compliance Drift Watcher
- Runs nightly. Compares your current state against every framework you have enabled and files roadmap tasks for gaps that opened since the last check.
- Microsoft Agent Enforcement
- Runs daily, with the Agent 365 registry connector on. When the AI system a Microsoft agent is linked to loses its approval (rejected, retired, expired, or past its approval expiry date), it proposes blocking the agent in Microsoft Agent 365, and once the system is approved again it proposes unblocking an agent you blocked.
Microsoft Agent Enforcement makes no change in Microsoft
RAIC never blocks, unblocks or reassigns a Microsoft agent itself. Approving a proposal records the decision and who made it; you then make the change in the Microsoft 365 admin center under Agents → All agents → Registry (the proposal links there). If the approval lapsed because the agent's owner left, reassign the agent there instead of blocking it. Microsoft's API for these changes is still in preview; RAIC will offer to make the change for you once it is generally available.
Autonomy levels
Each agent is configured independently, so you can let a low-risk agent act on its own while keeping a high-impact one on a short leash.
- Off
- The agent does not run.
- Draft only
- Every proposal waits for a human approval, regardless of confidence.
- Auto-apply above confidence threshold
- High-confidence proposals apply automatically; everything below the threshold queues for review.
- Fully autonomous
- Proposals apply automatically. Every write is audited and reversible.
The safety rule that always applies
When an agent has no automatic writer for a given proposal type, that proposal queues for human approval no matter which autonomy level you selected. Automation is never inferred; it exists only where it was explicitly built and reviewed.
The Proposal Inbox
Proposals collect in a reviewable inbox on the Automation Agents screen. Columns are sortable, so you can work by confidence, by agent, by age, or by target record. Approving applies the change and records the approval. Rejecting is terminal: the proposal is closed and recorded, and no governance data is modified. A rejected proposal is a documented decision, not a deleted one.
Runs and metrics
A separate Runs view shows every scheduled and manual execution: when the agent ran, what it examined, how many proposals it produced, and whether the run completed cleanly. Agent metrics summarize acceptance rates over time, which tells you whether an agent has earned more autonomy or needs its threshold tightened.
How this differs from Aslan AI
Aslan AI drafts on request: you ask for a risk assessment, a gap analysis, an executive summary, or a policy, and it produces a draft you edit and accept. Agents work the other way around — nobody asks them anything. They watch continuously and interrupt you only when there is a decision to make. Aslan is the analyst you brief; agents are the monitoring that never sleeps. Most organizations use both.
What gets recorded
Every configuration change, run, proposal, approval, rejection, and automatic write is recorded in your organization's activity log and mirrored to the platform log. There is no unlogged automated action anywhere in the module.
Availability and access
Autonomous Governance Agents are available to Enterprise organizations. Configuring agents, setting autonomy levels, and acting on proposals require Organization Administrator or Workflow Administrator permissions; other roles can view proposals and run history.
