Customer 360 Assistant
RAG vs. Tool Use
Question
Scenario
A sales leader asks for an “AI assistant that knows everything about every customer.”
The relevant data includes live CRM and product data, support tickets, call notes, and account plans. The assistant may also draft follow-ups or update an opportunity stage. The leader proposes putting everything into a vector database and building a chatbot.
Your task
Design the first version you would actually ship. Explain:
- what you would clarify before choosing an architecture;
- which requests should use direct queries, retrieval, tools, or deterministic code;
- where an LLM adds value and where it should not be trusted;
- how you would handle freshness, citations, permissions, and write actions;
- what you would deliberately leave out of v1.
Do not assume that every part of the system needs RAG or an agent.
Answer
Design the architecture
Explain the decision in writing, draw the system, or use both.
Write your answer before opening the hints or solution.
0 words
Interviewer nudges3 prompts · Open
- 1.Start from the user decisions and workflows, not the requested technology.
- 2.Separate structured facts, unstructured context, generation, and side effects.
- 3.A strong v1 may be a routed system rather than one universal agent.
Strong answerReview after your attempt · Open
Strong answer outline
Clarify the highest-value jobs, freshness, permission boundaries, latency, and the cost of an incorrect answer or write. Use direct APIs or SQL for current structured facts, retrieval for unstructured notes with citations and timestamps, and the LLM for intent, synthesis, and drafting.
Use an intent router with explicit structured-read, retrieval, draft, and approved-write paths. Put writes behind narrow typed tools, server-side validation, idempotency keys, and confirmation or approval. Enforce authorization at each source before data reaches the model.
Ship one or two frequent workflows first. Measure factual accuracy, evidence quality, task completion, latency, and adoption; leave broad autonomous execution out of v1.
Reference diagram
A simpler architecture than RAG
User request
Natural language
Intent + parameters
LLM or rules
Product API
Live structured data
Validated response
Grounded in tool output