This prevents hallucinations and unreliable outputs.
You're deciding when Retrieval-Augmented Generation (RAG) is required and when a standard AI model is sufficient.
RAG is about grounding AI in truth, not adding complexity.
Step 1: Identify Knowledge Dependence
For each AI-eligible task, ask:
- Does the task depend on internal documents?
- Does it require up-to-date information?
- Does accuracy matter more than creativity?
If yes, RAG is likely required.
Step 2: Check Source of Truth
Determine:
- Where the correct information lives
- Whether it changes often
- Whether staff already search for it manually
Simple example: Answering HR policy questions using an employee handbook.
Complex example: Generating client responses based on contracts, tickets, and account history.
Step 3: Assess Hallucination Risk
Ask:
- Would a wrong answer cause harm?
- Would it mislead a customer or employee?
- Would people trust the output without checking?
Higher risk increases need for RAG.
Step 4: Evaluate Data Readiness
Confirm:
- Documents are accessible
- Formats are usable
- Permissions are manageable
If data is scattered or locked down, flag as RAG-ready later.
Step 5: Assign Requirement Label
Label each task as:
- RAG Required
- RAG Optional
- No RAG Needed
What You Should Have Now
✅ RAG Requirement List
✅ Notes explaining each decision
✅ Data readiness flags
Quality Check
- RAG is used only when necessary
- Risk drives decisions, not hype
- Data availability is considered
- Labels are easy to justify
Next Step: With RAG requirements set, you're ready to assess LLM risk for each opportunity.