Skip to main content

Using AI for Transforming Customer Experience

Module: Module 3 — Strategy & Senior LeadershipCode: UAITCE (SKR)Faculty: Prof. Saurabh KumarSessions: 2Status: ✅ Drafted

Big idea​

Prof. Saurabh Kumar's frame on AI for transforming customer experience stacks five layers. At the base sit the three machine-learning paradigms: supervised (labelled examples — spam detection, default prediction), unsupervised (find patterns in unlabelled data — segmentation, anomaly detection), and reinforcement (learn by reward — game-playing, recommendation tuning). On top comes generative AI (LLMs that produce text, code, image), then agentic AI (LLMs that plan and act through tools), then prompt engineering (the practitioner skill of eliciting useful behaviour), and finally the responsibility layer (bias, hallucination, copyright, sustainability, regulation). The CX applications — Zomato's predictive ETAs, IndiGo's chatbot triage, Netflix recommendations, dynamic personalisation — are simply layered combinations of these primitives, deployed against the Gartner Hype Cycle that warns us where each technology actually sits in the trough or on the slope of enlightenment.

Key concepts​

  • Three ML paradigms. Supervised (labelled input→output pairs: classification, regression), Unsupervised (no labels: clustering, dimensionality reduction), Reinforcement (agent + environment + reward: sequential decisions). Each fits a different CX problem.
  • Generative AI and LLMs. Transformer architecture, next-token prediction trained on internet-scale text. ChatGPT (Nov 2022) democratised access. Key risks: hallucination, training-data bias, copyright provenance.
  • Agentic AI. LLMs that plan, call tools, and execute multi-step workflows (browse, query, write, retry). The path from chatbot → copilot → autonomous agent. Orchestration and guardrails become the binding constraint.
  • Prompt engineering essentials. Role + task + context + format + examples (few-shot); chain-of-thought for reasoning; retrieval-augmented generation (RAG) to ground answers in your data; structured output for downstream automation.
  • Gartner Hype Cycle. Innovation Trigger → Peak of Inflated Expectations → Trough of Disillusionment → Slope of Enlightenment → Plateau of Productivity. Locating a technology on the curve disciplines investment timing.
  • CX use cases. Zomato (delivery-time prediction, demand forecasting), IndiGo (chatbot triage, voice IVR), Netflix and Spotify (recommendation), e-commerce search and personalisation, sentiment-driven outreach, hyper-personalised retention offers.

Self-check​

A retail bank wants to build a customer-service chatbot that can answer policy questions accurately, never invent product features that don't exist, and quote the exact clause it relied on. Which architecture choice most directly addresses all three requirements?

  • A. Fine-tune a small LLM on bank product PDFs
  • B. Retrieval-Augmented Generation (RAG): the LLM retrieves the relevant policy passages from a vector index of authoritative bank documents at query time, generates the answer grounded in those passages, and cites the source clause — reducing hallucination and giving an audit trail
  • C. Use the largest available LLM with no grounding
  • D. Replace it with rule-based scripts
Supervised vs unsupervised vs reinforcement learning
Supervised: labelled (input, output) pairs — spam vs not. Unsupervised: no labels, find structure — customer clustering. Reinforcement: agent acts in environment, learns from reward signal — game play, recommendation tuning.

Click the card to flip

1 of 4

Continue learning​

🎓Coursera · CourseAI For Everyone (Andrew Ng)The most efficient executive primer on AI capabilities and limits — 4 hours, no math, the right intuition.📊Stanford HAI · ReferenceStanford AI Index ReportThe annual, authoritative dataset on AI capabilities, costs, adoption and policy — the source-of-truth for the field's trajectory.🧰OpenAI · ToolOpenAI Prompt Engineering GuideThe vendor-canonical reference on prompt patterns, RAG, function calling, and evaluation — the working practitioner guide.📊Gartner · ReferenceGartner Hype Cycle MethodologyThe five-phase curve and the underlying methodology — indispensable for disciplining when to pilot vs invest at scale.
🪞 Apply it — reflection prompts
  1. Pick one customer-facing process at your firm. Which of the three ML paradigms (supervised, unsupervised, reinforcement) fits the problem best — and is that what's actually being attempted?
  2. For one current AI initiative, draw the architecture: is the LLM grounded in your data via RAG, or relying on parametric memory? Where are the hallucination and citation risks?
  3. Locate each of your three biggest AI investments on the Gartner Hype Cycle. Are any sitting on the Peak of Inflated Expectations — and what would shift them to the Slope of Enlightenment?

📝 Going deeper. Andrew Ng's AI For Everyone (Coursera) is the most efficient executive primer. For the LLM/RAG architecture pattern, the OpenAI and Anthropic prompt-engineering and retrieval guides are the working references. On responsible AI, Stuart Russell's Human Compatible (2019) is the most readable serious treatment.