Docs/Concepts/Human-in-the-Loop
Core Concepts

Human-in-the-Loop

Understanding how Decisio balances AI automation with human oversight, and how to configure approval workflows that match your risk tolerance.

Philosophy

AI should augment human decision-making, not replace it. Decisio is designed with human oversight as a core principle:

  • Transparency: Every decision includes full reasoning
  • Control: You decide what auto-executes vs requires review
  • Learning: Your feedback improves the AI over time
  • Accountability: Clear audit trail for every decision

Approval Levels

Decisions are routed based on configurable policies that consider confidence, impact, and category:

Auto-Approve

Decision executes automatically without human review

TYPICAL CRITERIA

Confidence > 90%
Impact < ₹10,000
Routine category

Use case: Small price adjustments, routine reorders

Standard Review

Decision appears in approval queue for team review

TYPICAL CRITERIA

Confidence 70-90%
Impact ₹10K-₹1L
Normal priority

Use case: Most pricing and inventory decisions

Manager Approval

Escalated to manager with additional context

TYPICAL CRITERIA

Confidence < 70%
Impact > ₹1L
Strategic category

Use case: Large orders, major price changes

Blocked

Decision flagged as high-risk, requires investigation

TYPICAL CRITERIA

Anomaly detected
Policy violation
Manual override

Use case: Unusual patterns, policy conflicts

Configuring Approval Policies

Approval policies are configured in Settings → Approval Workflows. You can create rules based on:

Impact Thresholds

Set monetary thresholds for automatic approval:

  • Decisions under ₹X auto-approve
  • Decisions between ₹X and ₹Y require standard review
  • Decisions over ₹Y require manager approval

Confidence Thresholds

Route decisions based on AI confidence:

  • Above 90%: Auto-approve (high confidence)
  • 70-90%: Standard review
  • Below 70%: Escalate to manager

Category Rules

Apply different policies to different product or decision categories:

  • Strategic products always require review
  • Routine categories can auto-approve
  • New products require extra scrutiny

The Approval Queue

Decisions requiring review appear in the Approval Queue with:

  • Decision summary: What the AI recommends
  • Reasoning: Why this recommendation was made
  • Impact forecast: Expected revenue/margin change
  • Confidence: How certain the AI is
  • Alternatives: Other options considered
  • Context: Relevant data and history

Taking Action

For each decision in the queue, you can:

  • Approve: Execute as recommended
  • Modify & Approve: Adjust values before executing
  • Reject: Decline with optional reason
  • Defer: Hold for later review
  • Escalate: Send to another team member

Learning from Feedback

Every action you take teaches the AI:

  • Approvals: Reinforce similar recommendations
  • Modifications: Learn your preferred adjustments
  • Rejections: Avoid similar patterns in future

Over time, the AI learns your preferences and the approval queue becomes more accurate and requires less manual intervention.

Audit Trail

Every decision is logged with:

  • Who approved/rejected
  • When the action was taken
  • What modifications were made
  • The reasoning provided

This audit trail is essential for compliance, performance review, and understanding decision patterns over time.