Every decision you make in the queue feeds back into the system, both by adjusting confidence over time and by surfacing suggestions for how to tune it faster.
#How workflows learn
Each approve or reject updates the system’s sense of your “good output.” Rejections are particularly valuable, especially with feedback (free text or a category like Wrong Tone, Incorrect Data, or Missing Info). Workflows that start at low confidence and always escalate can reach high confidence on routine cases after a few weeks of feedback.
Your decisions are what move a step toward autonomy: Rills suggests the switch, you accept it, and confidence is still scored on every run afterward. 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.
#Insights
Insights analyzes your runs and suggests changes that reduce manual review or improve confidence calibration. Available on Professional and higher plans.
#Where to find it
Insights aren’t a separate page. They appear inside your review queue, alongside the approvals waiting for you on the Approvals page. When Rills spots an optimization, it drops a suggestion card into the same queue you already work.
#What it suggests
Insights surfaces a few categories of suggestion:
- Lower the confidence bar: a step that’s been auto-approving cleanly for a long streak, where the confidence bar can safely come down so fewer routine cards reach you.
- Start auto-approving: an Always ask step with a clean approval streak that’s ready to switch to Confidence gated mode and start scoring and auto-approving routine runs.
- Refine a prompt: proposed prompt edits for AI nodes whose output you keep editing in the same way. You see the suggested change and can accept or dismiss it.
- Recalibrate confidence: when confidence scoring is diverging from your actual approve/reject decisions, Insights suggests adjusting how inputs are weighted so the score tracks your judgment.
#Acting on a suggestion
Each suggestion shows the diagnosis and the proposed change. Suggestions with a concrete change to apply have Accept and Dismiss; Accept applies the change as a new workflow version, so old runs continue on the prior version and nothing in flight is disrupted. Advisory suggestions with nothing to auto-apply have Mark read instead. Dismissing (or marking read) clears the card without changing anything.
#How it works
Insights uses your run history and approval decisions as input. The more decisions you’ve recorded, the more targeted the suggestions. Brand-new workflows show fewer suggestions until they have data to analyze.
#Related
- Approvals : the review queue and how to approve, reject, or edit
- Modes & Confidence : approval modes and how confidence scoring works
- Conversations : Delta is the agent that walks you through optimization questions interactively