Human approval workflows in Agentic Ai - Model Pipeline Trace

Model Pipeline - Human approval workflows

This pipeline shows how an AI system works together with humans to make decisions. The AI suggests answers, and humans approve or correct them before final use.

Data Flow - 5 Stages

1Raw Input Data

1000 requests x 10 features→Collect user requests with details→1000 requests x 10 features

Request: 'Approve loan for $5000, credit score 700'

↓

2AI Model Prediction

1000 requests x 10 features→AI model predicts approval probability→1000 requests x 1 probability score

Prediction: 0.85 (85% chance to approve)

↓

3Human Review Queue

1000 requests x 1 probability score→Select uncertain cases for human review→200 requests x 1 probability score

Selected requests with scores between 0.4 and 0.6

↓

4Human Approval

200 requests x 1 probability score→Humans approve or reject requests→200 requests x 1 final decision

Human decision: Approve or Reject

↓

5Final Decision Output

1000 requests x 1 probability score + 200 human decisions→Combine AI and human decisions→1000 requests x 1 final decision

Final decision: 1 (approved) or 0 (rejected)

Training Trace - Epoch by Epoch

Loss
0.7 |****
0.6 |****
0.5 |****
0.4 |****
0.3 |****
    +----
    1 2 3 4 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Model starts learning, accuracy low
2	0.50	0.72	Loss decreases, accuracy improves
3	0.40	0.80	Model converging, better predictions
4	0.35	0.85	Stable improvement, ready for deployment

Prediction Trace - 4 Layers

Layer 1: Input Features

Layer 2: AI Model Prediction

Layer 3: Human Review Check

Layer 4: Final Decision

Model Quiz - 3 Questions

Why are some requests sent to humans for approval?

A Because AI always needs human approval

B Because AI is unsure about those cases

C Because humans want to approve all requests

D Because data is missing for those requests

Key Insight

Human approval workflows help AI systems handle uncertain cases by involving humans, improving overall decision quality and trust.