What the roles are
One skill set, three destinations
Generative AI produces text, code, summaries or SQL from a prompt. Agentic AI wraps a model in a loop that plans, calls tools such as a database or a BI layer, checks the result and acts, with a person approving the consequential steps. For an analyst the practical difference is scope: a GenAI feature answers a question; an agent runs the multi-step workflow around it. Job postings sort the people who work with these systems into three families.
AI-augmented analyst
Data or BI analyst using copilots and natural-language query tools (Databricks Genie, Snowflake Cortex Analyst, Power BI Copilot). The work moves upstream: curating the semantic layer, defining metrics, validating what the model returns.
GenAI / applied-AI analyst
An analytics seat that also builds GenAI solutions on business data: chatbots, personalisation, RAG over documents, agentic workflows. Citi, JPMorgan and HCLTech post these in Bengaluru and Noida.
AI or agent engineer
A software role that builds and operates production LLM and agent systems: Python services, vector databases, orchestration, deployment, monitoring. Where an analyst with strong Python can go next.
Grounding is the analyst's leverage
Vendor benchmarks put a general-purpose LLM at 51% accuracy on business questions over raw data, and at 90% or more when grounded in a semantic layer the analyst defined. That gap is why analytics teams hire people who understand the data, not only the model.
Snowflake's Cortex Analyst benchmark, via Sundas Khalid. Vendor figures; AtScale reports a similar gap (over 80% wrong on raw data, 92–100% with a semantic layer).
"Apply GenAI to chatbot performance and personalisation, and build solutions using prompt engineering, RAG and agentic AI… Hands-on experience in Python, SQL and machine learning is a must."
Citi, Business Analytics Analyst, Generative AI, BengaluruWho it is for
Built for analysts who already pull their own data
Every GenAI analyst posting we read asks for SQL and Python first and the model layer second. The role is an extension of analytics work, not a switch into software engineering, and the entry bars in India run from fresher-friendly to three years.
✓A fit if you
- Write your own SQL on large datasets and have working Python (pandas, APIs); PySpark is a plus at Citi
- Are a data, BI, business, product or marketing analyst, or a decision scientist who already runs experiments
- Can define a business problem and present a structured solution to management
- Have built at least one LLM project, academic or self-initiated (HCLTech and JPMorgan both say these count)
- Are comfortable owning requirements, testing and adoption for an AI programme
- Treat data governance and PII handling as part of the job
…Not yet if you
- Work in Excel and dashboards only, with no SQL or Python: every posting found asks for both
- Expect a prompt-only job: LinkedIn lists prompt engineer among India's fastest-growing titles, but analytics postings pair prompting with SQL, retrieval and evaluation
- Want to train or fine-tune foundation models: that is a data-science or ML-engineering track, and none of these seats ask for it
- Have only a chat wrapper over an API, with no data behind it and no evaluation
Analytics engineers sit closest to the shift: 57% of their postings ask for modern AI, against 24% for analysts. AI Analyst Lab
Experience bars in current India postings
| Employer, role | Location | Years required | Also asks for |
|---|---|---|---|
| HCLTech, Analyst, Generative AI | Noida, Hyderabad, Chennai, more | 0–3 | Python for GenAI, LangGraph and LangChain, LLM concepts, data preprocessing, model evaluation, a hands-on LLM project |
| JPMorgan, Applied AI ML Analyst | Bengaluru | 2+ | Production Python with tests; RAG components; agent workflows; evaluation automation and telemetry; PII handling |
| Citi, Business Analytics Analyst, GenAI | Bengaluru | 3+ (2 with Master's) | Python, SQL, ML as a must; prompt engineering, RAG, LangChain, LangGraph, agentic AI, GenAI governance; A/B testing |
| Alstom, AI Business Analyst | Bengaluru | not stated | UAT, change management, training and adoption with business teams on AI programmes |
| Deloitte, Analyst, Generative AI | Bengaluru | not stated | Agentic workflows and LLM orchestration (role filled at time of reading) |
| GenAI / agentic engineer seats | Bengaluru | 5–8+ | Production LLM delivery, FastAPI, vector databases, multi-agent design: the engineer track, for contrast |
Postings read between 30 Oct and 5 Nov 2026 via company sites and job boards. The JPMorgan posting closed on 25 Sep 2026 and the Citi posting on 29 Oct 2026; both are quoted for their requirements, which recur across similar roles.
Skills to learn
The analytics foundation stays; four layers go on top
Analyst postings still run on SQL and BI, named in 91% of them. The GenAI skills asked of analysts are a small share today, but they are the same list employers ask of data scientists at four to ten times the rate, which is where the pay premium sits.
Share of data analyst postings asking for each skill
340 data analyst postings in a 1,294-posting US census, re-read 7 Sep 2026. AI Analyst Lab
The same skills in data scientist postings
385 data scientist postings, same census. The gap between the two charts is the skill gap an analyst closes to move into a GenAI seat.
Analytics foundation
Keep pulling your own data
- SQL on large datasets and warehouses; Python with pandas, PySpark where the stack has it
- Statistics and experimentation: A/B and multivariate testing, reading results honestly
- BI tools (Power BI, Tableau, Looker) and the semantic layer and metric definitions behind them
- Turning analysis into a recommendation a business partner can act on
LLM fundamentals
Know what the model is doing
- How models generate text; tokens, context limits and cost; what temperature changes
- Prompt, retrieve or fine-tune: choosing the simplest approach that works
- Why models hallucinate and how to limit it; prompt-injection defences
- Structured outputs so the model's answer fits your pipeline
Building with GenAI on business data
RAG, prompting and orchestration
- Prompt engineering with versioned prompts, not one-off chats
- RAG end to end: ingestion, text segmentation, embeddings, indexing, retrieval, citations (JPMorgan's wording)
- LangChain and LangGraph, named in the Citi and HCLTech postings; vector databases
- Natural-language-to-SQL and copilots (Databricks Genie, Snowflake Cortex, Power BI Copilot), and the grounding they need from you
Agentic workflows, evaluation and governance
Make it safe to run without you watching
- Planning and execution logic, tool integration, state, retries and error recovery
- Human approval for consequential actions; read-only credentials; capped tool calls
- Offline test sets, evaluation automation, telemetry dashboards, regression checks; retrieval and answer metrics (Ragas, DeepEval, LangSmith)
- GenAI governance and PII handling, asked for by both Citi and JPMorgan
Skill groups compiled from the Citi, JPMorgan, HCLTech and Deloitte postings and the AI Analyst Lab and InterviewStack posting analyses.
Projects to build
Put a model on data you already understand
The projects that match these postings sit on business data the candidate can explain, and ship with an evaluation set, guardrails and a live demo. Recruiters look for engineering depth and measurable outcomes, not a notebook. Pick the one closest to the seat you want.
AI-augmented analyst · agentic analytics
A natural-language data analyst agent
Question in, validated SQL and a chart out, over a real warehouse or a DuckDB/Postgres copy of one.
Must include: a semantic layer or metric definitions the agent is grounded in; read-only credentials; a test set of 50+ questions with known answers and an accuracy score; handling for ambiguous questions and empty results; a live Streamlit demo.
GenAI analyst · the Citi and JPMorgan asks
A RAG assistant over business documents
Policies, product manuals or past analyses, with source citations and a way to say "not in the documents".
Must include: ingestion, chunking, embeddings, retrieval; citations; retrieval and answer evaluation with Ragas, DeepEval or a hand-labelled set; a short write-up of chunking choices and what changed accuracy.
Agentic AI in an analytics team
An agentic reporting workflow
Pull data on a schedule, run checks, draft the weekly insight summary, route anomalies to a person for approval.
Must include: LangGraph or similar with explicit state; tool calls capped; human approval before anything is sent; retries and error recovery; telemetry on runs, cost and latency; a before/after on time saved.
The three artefacts to walk into an interview with
- A deployed demo reviewers can open without local setup: Streamlit, Hugging Face Spaces or Vercel.
- An evaluation set and its results: task success rate, retrieval accuracy, hallucination rate, cost per query, latency, plus the edge cases you tested (empty results, conflicting instructions, prompt injection).
- A written case study with an architecture diagram: the business question, the data, what the model decides versus what the code controls, what went wrong, what you measured.
A fourth option if you work in digital product analytics: a GenAI experimentation helper that drafts test variants or chatbot responses, which you then measure with A/B tests, with a log of prompt versions and a governance note on what the model may and may not generate.
What does not count
- A chat wrapper over an API with no data, no evaluation and no guardrails
- Prompt collections or "100 prompts for analysts" repositories
- Notebooks that never left the laptop
- Projects on toy datasets you cannot explain in business terms
- An agent with write access to anything and no approval step
Guardrails employers expect to see: read-only database access, sandboxed code execution, capped tool calls per turn, and human approval for consequential actions.
The interview loop
Your usual analytics rounds, with GenAI depth added
Analytics employers keep their SQL, Python and business-case rounds and add GenAI depth on top; Citi lists SQL, Python, ML and experimentation beside RAG and agents. Analyst and associate seats are tested on fundamentals and use cases; senior and engineering seats add live system design and a practical build.
Screen
Motivation, which GenAI projects you have built, what data they ran on, and what you measured.
SQL and Python
The usual analyst round, often with a twist: "Pull and shape the data you would need to build an evaluation set for this model."
LLM fundamentals
"How does generative AI differ from discriminative AI?" "Why does tokenization matter for cost and context limits?" "What does temperature control?" "Why do LLMs hallucinate, and how do you defend against prompt injection?"
RAG and prompting
"What problem does RAG solve?" "When should you use RAG instead of fine-tuning, and when both?" "What are the common ways a RAG system fails?" "When does chain-of-thought prompting actually help?"
Agents and evaluation
"What types of memory does an agent need?" "Which actions need approval, and how do you know the agent chose the right tool?" "How do you evaluate a RAG system on both retrieval and generation?" "Offline versus online evaluation: what is the difference?"
Business case
Take a workflow such as support triage, campaign analysis or weekly reporting and scope a first GenAI version: metrics, risks, governance, and how you would present it to a business partner.
Project deep dive
Your evaluation numbers and how you measured them, what broke, what the model decides versus what the code controls, and what you would change.
What strong answers do
- Start from the business question and the data, then choose prompt, RAG or agent, in that order of complexity
- Quote an evaluation number from your own project and say how it was measured
- Separate what the model decides from what the application controls, and name the guardrail
- Treat governance (PII, access, what the model may generate) as part of the design, not an afterthought
How to prepare
- Build one of the three projects above end to end before you apply; interviewers go deep on one project, not wide on many
- Rehearse the RAG-versus-fine-tuning and prompt-versus-agent trade-offs with examples from your own data
- Keep your SQL and experimentation sharp; the analytics rounds are not replaced by the GenAI ones
- Prepare one story where the model was confidently wrong and what you changed
Question sets from CloudVyn, Simplilearn, GeeksforGeeks and Learnomate; round structure from the Citi and JPMorgan postings.