Operations
Review and act on pending AI actions before they execute
Last updated 8 Oct 2026Monitor · article 2 of 2
Operations is where you manage the work that needs human attention. Actions that are blocked by a guardrail, held below your confidence threshold, or routed by an escalation rule land in the queue for review, alongside other operational signals like changes awaiting a decision, failed events, and active pauses.
Open Queue in the sidebar, under the Governance group. The Queue is the entry point; the same group also links to Changes. Pacing sits under Campaigns in the Operate group, and Tool health is linked from Integrations. The Queue row carries a badge with the count of pending approvals. Incidents and pauses are not shown in the sidebar: a status in the top bar reports them on every page (see Activity).
Where Status Lives
There is no separate status board. Both /workspace/operations and the old Overview address, which older anomaly and health-digest emails still link to, open the Queue. Each signal the Overview used to gather now lives on the surface that acts on it:
| You want | Go to |
|---|---|
| Approvals, changes, pauses, escalations, and failures waiting on a human | Queue |
| Whether incident mode or a pause is active | The status in the top bar, on every page. When incident mode is on, the Queue opens with a banner giving its reason and a Deactivate incident mode control |
| Campaign anomalies and the campaign lifecycle board | Campaigns — anomalies under Needs attention, and the Campaign lifecycle board groups every synced campaign into Preflight, In-flight, and Post-flight |
| Rules firing more often than usual | Policies — the sidebar row carries a dot while any rule is on watch |
| Actions whose actual outcome drifted from the prediction | Action results |
| Tool and connector reliability | Tool health, linked from Integrations |
Queue
The Queue (/workspace/operations/queue) lists everything awaiting a human in one list. It is broader than approvals alone — each row's Source says where it came from:
- Approval — an AI action held for review
- Change — a change set awaiting a decision
- Pause — an active pause (emergency stop)
- Escalation — a running execution that an escalation policy handed to a human
- Workflow failure — an automation run that failed
- Failed event — an action that started but never finished, shown as "Stuck: …"
- Connection error — a platform connection that is failing, or whose credentials have expired or are invalid
- Channel — a connected channel whose access failed, expired, or was revoked
Each row shows:
- A priority dot for its urgency — critical, high, medium, or low
- SLA — time remaining until the SLA deadline, or Breached
- Operation — what it is (action name, change set title, pause scope)
- Reason — why it's here (e.g. the guardrail that fired, or that it fell below the confidence threshold), left out when it would only repeat the source
- Source — which of the sources above it came from
- Review — opens the item's detail beside the list; on wide screens its status is shown next to it
Three tabs sit above the list: Waiting (everything that needs a person), Mine (items assigned to you or waiting on your approval), and Resolved (what was decided, lifted, or resolved in the last 24 hours; the tab says so, and ends with a line on where each kind of item is kept after that). On Waiting and Mine you can filter by source, urgency, and SLA-breached status (these are also reflected in the URL as ?sources=, ?urgencies=, and ?slaBreached=true), and search the loaded items.
A change that is already being sent is not waiting on anyone. It is listed under In progress, below the Waiting list, and is left out of the Waiting count, the Mine tab, and the filter counts. The filters and search apply to the rows that wait on a person; In progress always lists every change being sent, until it finishes.
The scope bar above the list orders and labels waiting work by platform and working scope. It never hides anything: the Waiting count, the tab counts, and the approvals count in the sidebar are the same with or without a scope. With a scope chosen, the Waiting tab has two sections, each with its own select-all. In your scope holds work in the scope, set above it, partly outside it, or with nothing recorded to place it by. Outside your scope holds work the scope provably leaves out, and each of its rows says so. Changing the scope clears any rows you had ticked, so a bulk action never carries rows you picked under a different scope, and the bulk bar tells you when any selected row is outside the scope. Mine and Resolved carry the same labels (a row outside the scope says so) and are never reordered or narrowed by the scope; rows of a kind that records no platform, such as pauses, escalations and workflow failures, read as not recorded. When the queue is longer than the page, a line under the list says how many are shown. Work outside the scope comes last, so it is the first not shown, and the Outside your scope heading says when none of it is on the page.
Reviewing an Action
Select Review on an approval to open its detail beside the list: the full action manifest (every field that will change, with before and after values), the guardrail evaluation results (which rules ran and their outcomes), and the confidence score (the AI's certainty, on a 0–1 scale).
Approving an Action
Approve authorizes the action for execution. The action leaves the queue, executes, and the decision is recorded in the Audit log.
Live execution against external ad platforms is connector-gated and not enabled by default. Until a platform connector is configured, approved actions mutate internal platform state; they do not push changes to a live ad account.
Rejecting an Action
Reject denies the action — it will not execute. You can enter a reason, which is recorded in the Audit log and feeds back into the AI's calibration.
Changes
Open Changes in the sidebar, under the Governance group, or go to /workspace/operations/changes. The page shows change sets — grouped changes filed for review or already applied — with their field-level diffs, when they ran, how they were approved, and their execution status.
Each change set has a stable URL (/workspace/operations/changes/[id]) you can share with teammates or link from an incident report.
Pacing
The Pacing page (/workspace/campaigns/pacing) shows the autonomous pacing engine: which campaigns it is managing, the adjustments it made, and how effective those adjustments have been. The engine is configured from chat.
Tool health
The Tool health page (/workspace/operations/tool-health) reports how dependable the agent's MCP tools have been over the last 30 days — success rates, latency, and the errors that recur most often. Use it to spot tools or connectors that need attention.
Notifications
Configure alerts for new pending approvals under Notifications in the Settings menu in the top bar, so you're notified when items arrive in the queue.
Related articles
Still stuck? Browse the FAQ or contact us.