Build an Agentic AI Workflow:
From Single Answer to Autonomous Action

!pip install anthropic --quiet

import anthropic
import json

# Paste your API key here — keep this notebook private
API_KEY = "your-api-key-here"
client = anthropic.Anthropic(api_key=API_KEY)

print("✓ Anthropic client ready")

# The insurance denial we need to appeal
denial = """
INSURANCE DENIAL NOTICE
Patient: Maria Chen, DOB 03/14/1958
Claim #: HC-2024-88421
Service: Continuous Glucose Monitor (CGM) device + supplies
Denial reason: Not medically necessary. Patient does not meet
  criteria for CGM coverage under policy section 4.2.1.
  Criteria requires: Type 1 diabetes OR insulin-dependent
  Type 2 diabetes with documented hypoglycemic episodes.
"""

# Patient clinical context (in a real system, Penny retrieves this from the EHR)
patient_context = """
Patient: Maria Chen
Diagnosis: Type 2 Diabetes Mellitus (E11.9), diagnosed 2019
Current medications: Metformin 1000mg BID, Glipizide 10mg daily
  (Glipizide is a sulfonylurea — a class known to cause hypoglycemia)
Recent labs: HbA1c 8.4% (elevated), last 3 months
Recent notes: Patient reported two episodes of dizziness and
  confusion in past month; blood glucose 58 mg/dL on one occasion.
  Physician ordered CGM to monitor for hypoglycemic patterns.
"""

# VERSION A: Single prompt — just ask the AI to write the appeal
response_a = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1000,
    messages=[{
        "role": "user",
        "content": f"""Write an insurance appeal letter for this denial.

Denial: {denial}

Patient context: {patient_context}

Write a professional appeal letter."""
    }]
)

print("=== VERSION A: CHATBOT APPROACH ===")
print(response_a.content[0].text)

def call_claude(prompt, system=None):
    """Helper: send a prompt, return the text response."""
    messages = [{"role": "user", "content": prompt}]
    kwargs = {"model": "claude-sonnet-4-20250514", "max_tokens": 1000, "messages": messages}
    if system:
        kwargs["system"] = system
    response = client.messages.create(**kwargs)
    return response.content[0].text

print("=== VERSION B: AGENTIC APPROACH ===")
print("Running 3-step agent pipeline...\n")

# STEP 1: Analyze the denial — identify the specific criteria gap
print("Step 1: Analyzing the denial...")
denial_analysis = call_claude(
    f"""You are an expert in insurance policy and medical billing.

Analyze this insurance denial carefully:
{denial}

Identify:
1. The exact policy criteria that triggered the denial
2. The specific gap between what the insurer requires and what they know about the patient
3. What clinical evidence, if present, would override this denial

Be specific and cite the policy section mentioned.""",
    system="You are a medical billing specialist with deep expertise in insurance policy language."
)
print("✓ Denial analysis complete\n")
print(denial_analysis)
print("\n" + "="*50 + "\n")

# STEP 2: Match clinical evidence to policy requirements
print("Step 2: Matching clinical evidence to policy...")
evidence_match = call_claude(
    f"""You are a medical billing specialist.

The denial analysis identified these requirements:
{denial_analysis}

The patient's clinical record contains:
{patient_context}

Identify:
1. Which specific clinical facts in the record satisfy the insurer's criteria
2. Which policy language the clinical facts speak to directly
3. Any clinical details that strengthen the medical necessity argument

Be precise — quote the relevant clinical facts and map them to specific criteria.""",
    system="You are a medical billing specialist who builds airtight insurance appeals."
)
print("✓ Evidence mapping complete\n")
print(evidence_match)
print("\n" + "="*50 + "\n")

# STEP 3: Draft the appeal using the structured analysis
print("Step 3: Drafting the appeal letter...")
appeal_letter = call_claude(
    f"""You are drafting a formal insurance appeal letter.

Use this analysis of the denial:
{denial_analysis}

And this mapping of clinical evidence to policy criteria:
{evidence_match}

Write a formal, professional appeal letter that:
- Cites the specific policy section and explains why the denial is incorrect
- References the clinical evidence that satisfies the insurer's criteria
- Is factual, specific, and concise — no generic language
- Ends with a clear request for reconsideration

Format it as a real letter ready for submission.""",
    system="You are a medical billing specialist drafting a formal insurance appeal. Be precise, factual, and professional."
)
print("✓ Appeal letter drafted\n")
print("=== FINAL APPEAL LETTER ===")
print(appeal_letter)

# STEP 4: Self-check — the agent reviews its own output
print("Step 4: Agent self-check...")
self_check = call_claude(
    f"""You are a senior medical billing reviewer.

An AI agent has drafted this insurance appeal letter:
{appeal_letter}

The original denial was:
{denial}

The patient's clinical record contains:
{patient_context}

Review the appeal critically:
1. Does every factual claim in the letter match the clinical record exactly?
2. Does the letter address ALL the specific criteria in the denial, or does it miss any?
3. Is there any language that could weaken the appeal or give the insurer grounds to re-deny?
4. Rate the appeal: READY TO SUBMIT / NEEDS MINOR REVISION / NEEDS MAJOR REVISION

Be specific about any issues found.""",
    system="You are a senior medical billing reviewer. Be critical and precise."
)

print("=== AGENT SELF-CHECK ===")
print(self_check)

# In a real system, the agent would route the letter differently
# based on the self-check result before presenting to a human
if "READY TO SUBMIT" in self_check.upper():
    print("\n→ Agent confidence: HIGH — presenting to administrator for final review")
elif "MINOR" in self_check.upper():
    print("\n→ Agent confidence: MODERATE — flagging issues for administrator before review")
else:
    print("\n→ Agent confidence: LOW — escalating to senior billing specialist")