My Gen AI Interview Experience at Accenture
Last week, I interviewed for a Generative AI Specialist role at Accenture, and the discussion turned into an in-depth conversation on “AI Agent vs Agentic AI.”
Since my resume mentions LangChain projects and building agentic workflows for an internal chatbot, the interviewer focused heavily on how well I understand autonomous agents vs rule-based AI systems.
Here’s exactly how the conversation unfolded:
Top 10 Real Interview Questions and My Detailed Answers
Q1: I see you’ve built a chatbot in your previous company. Was it an AI Agent or Agentic AI? How would you differentiate the two?
My Answer:
In my previous role, I designed a customer support chatbot using Rasa and later extended it with LangChain. Initially, it was a basic AI Agent:
- It followed pre-defined intents and actions.
- Could answer FAQs or create tickets but couldn’t handle anything outside its training set.
Later, we evolved it into an Agentic AI system:
- Added LangChain Toolchains for dynamic reasoning.
- Integrated a vector database (Pinecone) to store past conversations for context retention.
- Enabled it to use APIs for multi-step workflows like “check order status, update address, and escalate to human if needed.”
The key difference:
- AI Agent was reactive, responding only to specific inputs.
- Agentic AI became proactive, handling ambiguous queries and chaining tasks autonomously.
The interviewer probed: “Did you design the planning module for tool selection or use prebuilt LangChain agents?”
I explained:
- Initially, we used LangChain’s Zero-Shot Agent, but later built a custom planning module using Thought-Action-Observation loops for better control.
Q2: Explain the sense-think-act loop of AI Agents with an example from your project.
My Answer:
In my chatbot’s AI Agent phase:
- Sense: It parsed user queries using NLU models.
- Think: Used an intent-classifier to match pre-set intents.
- Act: Executed pre-configured actions like fetching from a database.
Example:
- User: “What’s my order status?”
- Agent: Detected intent “get_order_status” → Called API → Replied.
No reasoning beyond predefined flows.
The interviewer challenged: “How would Agentic AI improve this flow?”
I responded:
- Agentic AI would reason:
- “If order ID is missing, ask for it.”
- “If API fails, attempt retries or switch to alternative APIs.”
- “If user asks for an unrelated task, pause and clarify.”
- Added self-corrective steps using LangGraph to handle such multi-turn logic.
Q3: Can you share how your resume’s Agentic AI project enabled autonomy?
My Resume Project:
“Built an Agentic AI system for HR automation using GPT-4 + LangChain, enabling dynamic resume screening and interview scheduling.”
My Answer:
This system was agentic because:
- It planned tasks dynamically: Screening resumes, ranking candidates, sending interview invites.
- Used external tools: Called APIs for calendar invites, Slack notifications, and even rejection emails.
- Implemented a reflection step: If invite failed due to time conflict, it re-planned alternative slots without human intervention.
The interviewer asked: “Did it handle ethical concerns like biased screening?”
I explained:
- Added guardrails using OpenAI moderation API to flag sensitive or biased content.
Q4: Why are enterprises like Accenture moving from AI Agents to Agentic AI?
My Answer:
- Scalability: AI Agents can’t handle unpredictable scenarios; Agentic AI thrives there.
- Autonomy: Reduces need for humans in the loop.
- Complex Task Orchestration: Useful for multi-step tasks like legal document drafting, onboarding workflows, or IT troubleshooting.
I referenced a recent Gartner report that predicts “50% of enterprise AI systems will exhibit agentic capabilities by 2027.”
Q5: What challenges did you face while building an Agentic AI system?
My Answer:
- Tool Chaining Failures: Sometimes APIs returned unexpected data, breaking the chain.
- Solution: Added fallback strategies and test harnesses.
- Hallucinations in LLM reasoning:
- Solution: Combined RAG (Retrieval Augmented Generation) with fact-check steps before executing actions.
The interviewer was impressed with this answer and said: “We need candidates who can anticipate and solve such real-world problems.”
Q6: Which frameworks would you use today to build Agentic AI and why?
My Answer:
- LangGraph (by LangChain) → For multi-step workflows.
- Auto-GPT → For quick prototypes.
- CrewAI → For multi-agent collaboration.
- OpenAI’s Function Calling → To tie LLM reasoning with APIs.
Q7: How does Agentic AI handle memory differently from AI Agents?
My Answer:
- AI Agents: Stateless or store only session-level context.
- Agentic AI:
- Uses long-term memory (Vector DBs like Pinecone).
- Maintains episodic memory across sessions.
Q8: What is the role of Thought-Action-Reflection loops in Agentic AI?
My Answer:
- Thought: Plan what to do.
- Action: Execute API calls, tools, etc.
- Reflection: Analyze outcomes and decide next steps.
This loop allows self-correction and adaptive reasoning, making Agentic AI more human-like.
Q9: How would you prevent an Agentic AI system from taking unintended actions?
My Answer:
- Set strict permissions for API access.
- Add human-in-the-loop checkpoints for critical decisions.
- Use Guardrails AI or OpenAI Moderation API for safety layers.
Q10: If given a chance, how would you apply Agentic AI in Accenture’s enterprise ecosystem?
My Answer:
- Finance: Automate invoice reconciliation.
- HR: End-to-end onboarding workflows.
- IT: Autonomous troubleshooting bots for employee queries.
I also proposed integrating Accenture’s internal tools (like myWizard) with Agentic AI for enhanced automation.
Key Takeaways for Gen AI Candidates
Back up your answers with real projects from your resume.
Learn frameworks like LangChain and LangGraph.
Prepare challenges + solutions to show practical expertise.