Sign inStart your trial

Approvals

The approval queue, modes, confidence scoring, and how workflows learn from your decisions

When a workflow needs human judgment, it pauses and queues a proposal for you to review. You approve, reject, or edit. Approved work runs; rejections feed back into confidence so the system learns what good output looks like to you.

Approvals are always free. You can add as many as you want, and paused workflows cost nothing while they wait for you.

#The queue

Open Approvals in the sidebar. Each item shows:

  • Pending action: what will run if you approve, with parameters
  • Context: trigger data and the output of any earlier steps
  • Confidence score: a 0–100 quality assessment, explained below
  • Workflow name and timestamp

Three buttons per item:

  • Approve: run the action as proposed
  • Reject: stop the run and record the rejection
  • Edit: modify parameters before approving (change email text, swap a value, etc.)

Sort by priority, workflow, or creation time. Filter by workflow or assignee.

#Approval modes

Each Human Review node runs in one of three modes:

  • Always: every run pauses for review. Use for high-stakes actions: invoices, deletions, anything where the cost of getting it wrong is high.
  • Confidence: auto-approve when the run’s confidence score meets your threshold; otherwise route to your queue. The default threshold is 90. Lower it as you build trust.
  • Never: auto-execute without review. For low-risk steps where review isn’t worth the friction.

Modes are per-node. A single workflow can have one step on Always (the email send), another on Confidence (the routing decision), and another on Never (logging).

Changing the mode on a node doesn’t affect proposals already in your queue, those stay in whatever mode was active when they were created.

#Confidence scoring

Confidence is calculated per run, not stored as a workflow property. The same step can score 95 on one run and 62 on the next, the score reflects the input data and the system’s pattern history, not the step itself.

The score combines several signals:

  • Validation: does the output match the expected schema?
  • Grounding: do referenced entities (IDs, names, URLs) actually exist?
  • History: how often have similar proposals been approved before?
  • Anomaly: are values within the normal range for this step?
  • Model check: does an independent second-pass model agree with the proposal?

Each signal contributes to a final 0–100 score. If the score meets the node’s threshold, the run auto-approves (when the mode is Confidence). Below the threshold, it goes to your queue.

The system needs some history before all signals are useful. Brand-new steps lean on the validation, grounding, and model-check signals; once you’ve made a handful of decisions, the history signal kicks in and confidence becomes more accurate to your preferences.

#How workflows learn

Every approve/reject decision updates the system’s sense of your “good output.” Rejections, especially with feedback (free text or a category like Wrong Tone, Incorrect Data, or Missing Info): are particularly valuable. Workflows that start at low confidence and always escalate can graduate to high confidence on routine cases after a few weeks of feedback.

This is the core difference between Rills and rigid automation: workflows graduate from supervised to autonomous on their own based on your decisions, no flag flips required. Edge cases keep coming to you because they keep scoring lower; routine cases stop bothering you because they keep scoring higher.

You can also tune calibration directly via Insights , which surfaces signals with high error rates and suggests rebalancing.

#Sampling for quality monitoring

Even when proposals auto-approve, you might want to spot-check the system’s autonomous decisions. Set a sampling rate (0–100%) on a node, and that fraction of would-have-auto-approved proposals get routed to your queue anyway for review.

Sampling runs after the confidence check. Set it to 10% to occasionally verify drift, 0% when you fully trust the step, higher when you want more visibility.

#Routing

By default, proposals land in the queue for any workspace member with approval permissions. On Professional and higher plans, you can assign a node to a specific team member, only that person sees the proposal. Useful for separation of duties (only billing-team-member approves refund proposals, for example).

Priority levels (Low, Normal, High, Urgent) affect queue ordering and notification urgency. Urgent items surface at the top and trigger immediate notifications; low-priority items sort below everything else.

#Timeouts

Each Human Review node can have a timeout (default 24 hours, max 30 days). When a proposal expires, the workflow takes whatever timeout action you configured: Approve, Reject, Escalate, or Fail. By default, expired proposals cause the step to fail: stale proposals shouldn’t auto-execute after the moment has passed.

Set tighter timeouts for time-sensitive workflows (1 hour for social media reply, 15 minutes for fraud review) and longer ones for batch work that can wait.

#Mobile

The same queue is available in the Rills mobile app with a swipe-to-approve interface. Decisions sync immediately across web and mobile.

#Decisions are immutable

Once you approve, reject, or edit a proposal, that decision is recorded permanently, you can’t undo it. This creates a clear audit trail of every human decision on every workflow run.

#Pricing

Reviewing proposals is free. Approving, rejecting, editing, attaching feedback, all free. Workflows paused waiting for review consume no credits.

You only pay when an approved action runs (Workflow Credits) or an AI step calls a model (AI Credits). Rejections cost nothing.

#Related