Review and Approve Actions

The approval queue is where you review workflow outputs before actions execute. This is the core of Rills' value: AI handles specific tasks like classification and drafting within your workflow, you approve before consequential actions fire, and the system learns which outputs to escalate and which to handle autonomously.

Why approvals matter

Most automation tools operate on rigid if/then rules or run fully autonomously. The first option is inflexible, the second creates anxiety about what the AI might do unsupervised.

Rills takes a different approach: AI steps within your workflows handle classification, drafting, and analysis. You review and approve before actions execute. Over time the system learns your preferences and handles routine cases automatically while escalating edge cases.

This eliminates "monitoring anxiety" — you're never wondering what your automation is doing or worrying that it made a bad decision. You have the final word without constant manual oversight.

The approval flow

Here's how a typical approval works from trigger to execution:

1. Trigger fires

Your workflow trigger activates — a new email arrives, a scheduled time hits, a webhook delivers data. The workflow starts running.

2. AI step runs

The AI step receives the trigger data and performs its configured task according to your prompt — for example, classifying the email or drafting a reply.

3. Execution pauses for approval

The pending approval appears in your queue with:

Context — What triggered the workflow and the AI step's output
Next action — What will execute if you approve (send this email, create this task, update this record)
AI output — The AI step's analysis or draft for your review
Confidence score — A quality assessment based on validation signals, pattern matching, and historical accuracy

4. You review and decide

From your approval queue (accessible via web dashboard or the mobile-optimized web interface), you review the proposal and choose:

Approve — Execute the action as proposed
Reject — Don't execute, and record that this type of proposal should be rejected
Edit — Modify the action before execution (change email text, adjust task details)

5. Approved actions execute

When you approve, the workflow continues and executes the action. The result is logged, and the workflow completes.

When you reject or edit, Rills records your decision and uses it to improve future confidence scoring.

Confidence scoring

Every execution that reaches an approval step has a confidence score from 0-100. This score is a quality assessment based on validation signals, pattern matching against past approvals, and the current input data.

How confidence works:

Low confidence (0-40): The execution has high uncertainty — edge case, ambiguous input, or not enough past data. These always require manual approval.
Medium confidence (40-70): The system has seen similar patterns and expects approval, but there's enough variation that it escalates for review.
High confidence (70-100): The system is confident based on your approval history. You can optionally set auto-approval thresholds for these.

Confidence improves over time:

As you approve and reject, the system learns your patterns. A workflow step that starts at 20% confidence (always escalates) can graduate to 80% confidence (rarely needs approval) after a few weeks of feedback.

This is how workflows learn — they start supervised, learn from your decisions, and gradually become autonomous for routine cases while still escalating unusual situations.

Auto-approval thresholds

Once a workflow has enough approval history, you can set an auto-approval threshold. For example:

Auto-approve actions with confidence ≥ 85%
Always require approval for confidence < 85%

This lets you gradually hand off routine decisions while maintaining control over edge cases. You can adjust confidence thresholds per AI node based on risk tolerance — different steps in the same workflow can have different thresholds.