AI Operations
1. What it is
AI Operations is the admin surface for governing every autonomous AI agent in the platform. From here you pause or resume agents by domain, review actions that require human approval, monitor agent health, and configure trust levels that control how much autonomy each agent earns over time.
- Enablement: AI Operations is available to every tenant automatically — it is not a licensed app. It appears in the admin sidebar for any user with the admin role or above.
- What it is not: AI Operations is not where you configure what an individual app does — it governs the safety envelope around the AI (limits, approvals, pauses). Individual app behaviour is configured in each app’s own settings.
2. Standing it up
AI Operations itself requires no setup — it is live as soon as any autonomous agent domain is enabled for your business. The typical sequence:
- Enable an app that has an AI agent (e.g. Shifts Scheduling, Time & Attendance, Recruiting). This is done from the app’s own settings or the Apps Marketplace.
- Go to Admin → AI Operations → Agent Settings. Each enabled domain appears as an accordion panel. Confirm that Autonomous operations is turned on for the domains you want.
- Every new domain starts at the Probation trust level — all actions require your approval before they execute. Check Admin → AI Operations → Approvals regularly to process the queue.
- To verify it is working, go to Admin → AI Operations → Overview. You should see domain cards with status indicators and recent activity.
Silent-until-paired: enabling an app’s AI agent in the app’s own settings is not enough if Autonomous Operations is off for that domain in AI Operations. The agent will be technically enabled but will not execute any actions. Both switches must be on.
3. How it fits together
Autonomous Actions
An autonomous action is any operation an AI agent proposes or executes — a schedule change, a leave adjustment, an employee data update. Every action is logged with a domain, capability, confidence score, priority, and status.
The lifecycle is: Detected (pending) → Pending approval → Approved (awaiting run) → Completed (or Rejected / Failed at any step). Actions that do not require approval move straight from Detected to Completed. The entire stream is visible in Action History.
Domains
Each agent operates within a domain — a named area tied to a marketplace app. The 15 canonical domains are: Scheduling, Attendance, Leave Management, Timekeeping, Skills & Certifications, Compensation, Employee Data, Service Desk, Training, Onboarding, Performance Management, Procurement, Recruiting, Contracts, and Mango IQ. Pause/resume controls, trust levels, and daily limits are all set per domain.
Trust levels
Every domain starts at Probation and can advance through three trust levels:
- Probation — all actions require approval. Advances after 50 successful actions with a 95% success rate.
- Standard — routine actions execute automatically; high-impact, bulk, and critical actions still require approval. Advances after 200 successful actions with a 98% success rate.
- Trusted — only critical actions require approval.
Trust advancement is evaluated daily by a background job. An admin can also set the trust level manually from Agent Settings (a reason is required for audit).
Rejection escalation
If a domain accumulates 3 or more rejected actions within 30 days, every action for that domain is routed to the approval queue regardless of trust level. This is additive — it cannot relax an existing approval requirement.
Approval Workflows (multi-level)
Separate from the autonomous-action approval queue, Approval Workflows let you define multi-level approval chains for specific app processes. These are configured at Admin → General Settings → Approval Workflows and are scoped to a marketplace app. Currently, three apps support trigger-specific workflows: Inspections (submission, template creation, corrective action), Frontline Execution (campaign request, campaign publication), and Policy Hub (policy publication). A workflow can have up to 10 approval levels with optional time-based escalation.
Agent Health
The Agents page shows the operational status of every registered agent — enabled or unavailable, with a link to each agent’s management surface. Data is sourced from the shared agent registry and cached for 2 minutes.
4. Running it
Reviewing pending approvals
- Go to Admin → AI Operations → Approvals. The badge count in the sidebar shows outstanding items.
- Filter by domain or priority if needed. Actions are listed oldest-first.
- For each action, review the description, confidence score, and affected user. Approve (optionally with notes) or reject (a reason is required).
- Use the checkboxes and Bulk Approve / Bulk Reject for high-volume queues.
Approved actions execute immediately. If execution fails, the action moves to the failed state and can be retried from its detail page (up to 3 retries).
Pausing and resuming agents
- Per domain: from the Overview dashboard, use the pause/resume button on any domain card. A pause reason is recorded in the activity log.
- Pause All: the emergency stop at the top of the Overview page pauses every enabled domain (including standalone agent apps like Platform Admin Agent) in one action.
- Resume All: the counterpart. Domains paused globally by a System Admin (console-level) cannot be resumed from this page — only the System Admin can lift those.
Reviewing action history
Go to Admin → AI Operations → Action History. Filter by domain, status, priority, and date range (24h, 7d, 30d, or custom). A trend chart shows daily volumes by status. Export to CSV covers up to 10,000 rows.
Retrying a failed action
From an action’s detail page (Action History → click any row), use the Retry button. Retries are capped at 3 per action. Sample-data actions cannot be retried.
Managing approval workflows
- Go to Admin → General Settings → Approval Workflows.
- Click New Workflow, select the app it applies to, and name the workflow. If the app supports trigger processes (e.g. Inspections), select which process this workflow gates.
- Add approval levels (up to 10). Optionally enable time-based escalation.
- Toggle the workflow on or off with the enable/disable control. You can also duplicate an existing workflow as a starting point.
5. Settings
All per-domain agent settings are managed at Admin → AI Operations → Agent Settings. Each domain has an accordion panel.
| Setting | Default | What it changes |
|---|---|---|
| Autonomous operations (on/off) | Off | Whether the agent can propose and execute actions for this domain. Both this and the app’s own agent toggle must be on. |
| Trust level | Probation | Controls which actions need approval. Can be set manually; a reason is required and recorded in the audit log. |
| Max actions per day | Per trust level (Probation: 10, Standard: 50, Trusted: 200) | Daily cap on agent-executed actions. 0 means unlimited. |
| Confidence threshold | Per app config | Minimum confidence score (0–100) an action must reach to proceed. Actions below this are skipped. |
| Pause on error rate | Per app config | Error rate percentage (over 24 hours) that triggers an automatic pause. 0 disables auto-pause. Requires at least 5 actions before it activates. |
| Require approval for high-impact actions | Off | When on, actions affecting more than 5 people, or involving compensation or termination, always require approval — even at the Trusted trust level. |
Per-trust-level overrides for max actions/day, confidence threshold, and error rate can be set in the advanced section of each domain panel.
6. More help
- AI Operations FAQ — specific questions about limits, thresholds, and troubleshooting
- Agents page (Admin → AI Operations → Agents) — each agent links to its own app’s management surface
- Trust & Safety page (Admin → AI Operations → Trust & Safety) — Company Brain stats, rejection feedback, and guardrail configuration
- Ask AI — the assistant answers questions about AI Operations from these articles.