Preventing AI Hallucinations

You ask: "What's the prior authorization process for MRI scans?"

Gurubase searches for:
- "What's the prior authorization process for MRI scans?" (your exact words)
- "MRI pre-authorization requirements" (alternative phrasing)
- "imaging prior auth workflow" (domain-specific rephrasing)

This finds more relevant information than a single search would.

Question: "What's the deductible for out-of-network specialists?"

Document A: "Out-of-network benefits include a $500 deductible..."
Score: 95% - Directly answers the deductible question

Document B: "In-network specialist visits require..."
Score: 40% - Related to specialists but focuses on in-network

Document C: "General plan overview..."
Score: 10% - Barely relevant, only mentions deductibles in passing

Answer about "diabetes medication coverage":
- Found 8 relevant sources
- 6 scored above 80% (excellent)
- 2 scored around 60% (good)
- Overall Trust Score: 78%

Answer about "coverage for experimental treatment XYZ":
- Found 3 relevant sources
- 1 scored 55% (okay)
- 2 scored 35% (weak)
- Overall Trust Score: 41% (warning)

Question: "What's the best restaurant near the hospital?"
(For a healthcare policy Guru)

Gurubase response:
"This question appears to be outside my area of expertise. I'm specialized
in healthcare policies and procedures. I don't have information about
restaurants. Could you rephrase your question to relate to healthcare?"

Instead of:
Making up restaurant recommendations or trying to answer anyway (hallucination)

✓ DO: "According to the policy document, the annual maximum benefit
      for orthodontics is $2,000 per member..."

✗ DON'T: "The orthodontic coverage is probably around $1,500..."
         (inventing coverage amounts)

✓ DO: "The documentation covers standard claims processing but doesn't
      mention expedited appeals. I can help with standard claims."

✗ DON'T: "For expedited appeals, you would typically submit..."
         (hallucinating about missing information)

Situation: Policy document says Drug X requires prior authorization, but
there's a recently approved exception for emergency situations that was
manually documented by a compliance officer.

Without manual override:
AI might say: "Drug X requires prior authorization" (technically correct
but misses critical exception)

With manual override (edited answer):
AI says: "Drug X requires prior authorization. Exception: In emergency
situations, a 72-hour retroactive authorization is permitted." (accurate
and complete)

Question: "What's the claims appeal deadline for denied procedures?"

Trust Score: 85%

Sources Used:
📄 Member Handbook - Appeals Process (relevance: 95%)
   Section 7.2: Filing an Appeal

📄 Compliance Guide - Deadline Requirements (relevance: 88%)
   Chapter 4: Regulatory Timelines

📄 FAQ - Common Claims Questions (relevance: 82%)
   Appeals and Grievances section

Sources Filtered Out:
📄 General Plan Overview (relevance: 35% - too general)
📄 Provider Network Directory (relevance: 28% - different topic)

Question: "Is Wegovy covered under my plan for weight loss?"

AI Response: "Yes, Wegovy is typically covered for weight loss treatment
with a prior authorization. Most plans cover it with a $50 copay."

Problem: The AI hallucinated coverage details based on "typical" plans,
but your specific plan might exclude weight loss medications entirely.

Question: "Is Wegovy covered under my plan for weight loss?"

Gurubase searches policy documents...
- Finds: "Prescription drug formulary, Tier 3 medications, exclusions list"
- Doesn't find: Wegovy listed in covered medications

Response: "According to the current formulary, Wegovy is not listed as a
covered medication. Weight loss medications appear in the exclusions section.
I recommend contacting member services for alternative options or to request
an exception review."

Trust Score: 91% (based on clear policy documentation)
Sources: [Links to formulary and exclusions list]

Question: "What's the best investment strategy for my retirement?"
(Asked to a health insurance Guru)

Response: "This question is outside my area of expertise. I'm specialized
in health insurance policies and benefits. For investment advice, you
might want to consult with a licensed financial advisor."

Question: "Is Dr. Smith in network?"
(Provider not found in documentation)

Response: "I couldn't find information about Dr. Smith in the current
provider directory. This might be:
- A recently added provider not yet in the system
- A provider using a different name variation
- An out-of-network provider

I recommend:
1. Calling member services with the provider's NPI number
2. Checking the online provider search tool
3. Asking the provider's office to verify network status"

Question: "What's the annual deductible for my plan?"

Found: Two documents with different values ($500 vs $1,000)

Response: "I found conflicting information about the annual deductible:
- Summary of Benefits states: $500
- Member Handbook states: $1,000

This discrepancy should be clarified. I recommend:
1. Checking your most recent plan documents
2. Calling member services for confirmation
3. Reviewing your Explanation of Benefits (EOB)"

Trust Score: 45% (Low - due to conflicting sources)

Question: "What's the coverage for telehealth mental health visits?"
Reason: NOT_ENOUGH_CONTEXT
Missing: No documentation about telehealth mental health benefits found
User Intent: Benefits inquiry
Date: 2024-01-15

Week 1: 15 questions about "telehealth coverage" couldn't be answered
Action: Team adds comprehensive telehealth benefits documentation
Week 2: Telehealth questions now answered with 89% trust score
Result: Members get answers, fewer calls to member services

Initial: "What's covered for physical therapy?"
Follow-up: "Can you provide more details about the visit limits and
prior authorization requirements specifically?"