Human approval workflows in Agentic Ai

Introduction

Human approval workflows help make sure AI decisions are checked by people before final use. This keeps results safe and trustworthy.

When AI suggests important actions like approving loans or medical treatments.

When AI is unsure about a decision and needs a person to confirm.

When legal or ethical rules require human review before acting.

When AI handles sensitive data and mistakes could cause harm.

When building trust with users by showing humans are involved.

Syntax

Agentic_ai

workflow = HumanApprovalWorkflow(
    ai_model=model,
    approval_condition=condition_function,
    human_reviewer=reviewer_function
)

result = workflow.run(input_data)

HumanApprovalWorkflow connects AI output with human checks.

The approval_condition decides when human review is needed.

Examples

This example sends AI results to a human if confidence is below 80%.

Agentic_ai

workflow = HumanApprovalWorkflow(
    ai_model=my_model,
    approval_condition=lambda output: output['confidence'] < 0.8,
    human_reviewer=ask_human
)

result = workflow.run(data)

This example sends all AI outputs to human for approval.

Agentic_ai

def always_approve(output):
    return True

workflow = HumanApprovalWorkflow(
    ai_model=my_model,
    approval_condition=always_approve,
    human_reviewer=ask_human
)

result = workflow.run(data)

Sample Program

This program creates a simple AI model that outputs a prediction with confidence. If confidence is below 0.9, it asks a human to approve. The human approves only if confidence is at least 0.5. The program tests three cases with different confidence values.

Agentic_ai

class HumanApprovalWorkflow:
    def __init__(self, ai_model, approval_condition, human_reviewer):
        self.ai_model = ai_model
        self.approval_condition = approval_condition
        self.human_reviewer = human_reviewer

    def run(self, input_data):
        ai_output = self.ai_model(input_data)
        if self.approval_condition(ai_output):
            print('AI output needs human approval.')
            approved = self.human_reviewer(ai_output)
            if approved:
                print('Human approved the AI output.')
                return ai_output
            else:
                print('Human rejected the AI output.')
                return None
        else:
            print('AI output auto-approved.')
            return ai_output

def simple_ai_model(data):
    # Simulate AI output with confidence
    return {'prediction': 'Accept', 'confidence': data}

def human_review(output):
    # Simulate human approval if confidence >= 0.5
    print(f"Human reviewing output with confidence {output['confidence']}")
    return output['confidence'] >= 0.5

workflow = HumanApprovalWorkflow(
    ai_model=simple_ai_model,
    approval_condition=lambda out: out['confidence'] < 0.9,
    human_reviewer=human_review
)

print('Test with confidence 0.95:')
result1 = workflow.run(0.95)
print('Result:', result1)

print('\nTest with confidence 0.7:')
result2 = workflow.run(0.7)
print('Result:', result2)

print('\nTest with confidence 0.4:')
result3 = workflow.run(0.4)
print('Result:', result3)

Important Notes

Human approval adds time but improves safety and trust.

Design clear conditions to avoid too many or too few human checks.

Keep human reviewers informed and trained for consistent decisions.

Summary

Human approval workflows combine AI with people to check important decisions.

They help catch mistakes and meet rules for sensitive tasks.

Use simple conditions to decide when to ask humans for approval.