When a workflow hits an approval step, it pauses and adds a proposal to your queue. You review, decide, and the workflow continues. Over time the system learns which proposals you keep approving and starts handling them on its own, edge cases still come to you.
#Why this exists
Most automation is either rigid (if/then rules) or fire-and-forget (you find out what your bot did after the fact). Rills sits in between: AI proposes, you approve. You’re never wondering what your automation just did.
#The flow
- Trigger fires.
- Workflow runs through any AI steps, conditions, and pre-approval logic.
- Workflow hits an approval step and pauses.
- Proposal lands in your queue with: what triggered the run, the AI’s output, the action that will run if approved, and a confidence score.
- You approve, reject, or edit the proposal.
- If approved, the workflow continues and runs the action. If rejected, it stops and records your decision.
Paused runs cost nothing, you can leave a proposal sitting for hours or days without burning credits.
#Where to review
- Web: Approvals page in the dashboard. Best for batch review at your desk.
- Mobile: swipe-to-approve in the Rills mobile app . Best for quick decisions on the go.
Both views show the same queue. Decisions sync immediately.
#What you see in a proposal
- The proposed action and its parameters
- Trigger data and any output from earlier steps
- A confidence score (0–100): a quality assessment combining schema validation, entity grounding, history of similar approvals, anomaly detection, and an optional second-pass model check
- The workflow name and when the proposal was created
You can attach feedback to any decision (free text + a category like Wrong Tone, Incorrect Data, or Missing Info). Feedback feeds back into confidence so similar proposals score appropriately next time.
#Approval modes
Each Human Review node runs in one of three modes:
- Always: every run pauses for review, regardless of confidence. Use this for high-stakes actions: sending invoices, deleting records, anything you want to personally verify.
- Confidence: auto-approve when confidence ≥ a threshold you set, otherwise route to your queue. Default threshold is 90%. Lower it as you build trust in a step’s output.
- Never: auto-execute. For low-risk steps where review isn’t worth the friction.
A workflow can have multiple Human Review nodes, each in a different mode, one Always before the email send, another on Confidence before a routing decision.
#How learning works
Confidence is calculated per run, not stored as a static workflow property. The same step can score 95 on one run and 62 on the next, depending on the input. As you approve and reject, the system updates its sense of what your “good” output looks like, and routine cases start scoring higher and auto-approving while edge cases keep coming to you.
This is the core difference between Rills and rigid automation: workflows graduate from supervised to autonomous on their own.
#Pricing
Approvals are free. Always. Reviewing, approving, rejecting, editing, none of it consumes credits. You only pay when an approved action runs (Workflow Credits) or an AI step calls a model (AI Credits).
#Next
Understand credits to see what you’re paying for.