GENERATIVE AI: ZERO TO JOB-READY › UNIT 4 › MODULE 2 OF 4
AI Agents and Automation Workflows
⏱ 14 min read📝 Assignment insideNo coding required
In Module 4.1 you learned how RAG gives an LLM access to your documents. This module covers the next step: what happens when an LLM does not just answer a question but takes actions — browsing the web, writing and running code, calling APIs, filling forms, and coordinating with other AI systems. That is what AI agents do.
Agentic AI is the fastest-moving area of enterprise AI in 2026. Understanding it at a conceptual level is now a baseline expectation for anyone entering an Indian IT organisation.
The scale of agentic AI adoption
40%
of enterprise applications will include task-specific AI agents by 2026, up from less than 5% in 2025 — Gartner
57%
of companies already have AI agents running in production — G2 Enterprise AI Agents Report 2025
$182B
projected global agentic AI market by 2033, from $7.6B in 2025 — Grand View Research
India’s AI journey reached a turning point in 2025. The national conversation has shifted from whether AI will change business to how it can be implemented responsibly and efficiently. About 47% of Indian organisations now operate multiple GenAI use cases, and as AI systems evolve from assistive to agentic, this steady shift reflects how technology is beginning to redefine work, governance, and innovation.
Chatbot vs AI agent — the critical difference
Most people’s mental model of AI is a chatbot: you ask a question, it answers. AI agents are fundamentally different.
Can take real actions: browse web, run code, call APIs, send emails
Uses tools to interact with external systems
Maintains memory across a task (and sometimes across sessions)
Operates with varying degrees of autonomy — from supervised to nearly independent
The one-line definition: An AI agent is software that can decide the next step, use tools, and complete tasks toward a defined goal — without human approval at every step. Unlike chatbots, agents operate in a continuous loop of plan, act, observe, and adapt until the task is complete.
The five core components of an AI agent
Core components of an AI agent are perception, reasoning, memory, planning, and tool-based actions such as web search, code execution, APIs, and file systems. Here is what each means in practice:
👁
1. Perception — reading the environment
The agent takes in information — your instruction, a web page, a document, a database query result, an API response, or the output of the previous step. Perception is what the agent knows before it decides what to do next.
Agent receives: "Find the last 5 support tickets assigned to Team A, summarise the top issue, and send a Slack message to the team lead."
🧠
2. Reasoning — deciding what to do next
The LLM at the core of the agent decides what action to take next based on the goal and the current state. This is the intelligence layer — the agent does not follow a fixed script; it reasons about the best next step.
Agent reasons: "I need to query the support ticket system first. I will use the ServiceNow API. After getting results, I will summarise them. Then I will compose a Slack message."
💾
3. Memory — retaining context across steps
Agents have both short-term memory (what happened in this task so far) and optionally long-term memory (information persisted across multiple tasks). This is what allows a multi-step task to be coherent — the agent remembers what it already did.
After retrieving 5 tickets, the agent remembers all 5 results while composing the summary — it does not re-query the system for each step.
📋
4. Planning — breaking goals into steps
For complex goals, the agent creates a plan — an ordered sequence of sub-tasks. More advanced agents revise this plan mid-task when they discover unexpected information. Planning is what separates a simple tool-call from a genuine multi-step workflow.
Plan: [1] Query ServiceNow API → [2] Parse 5 ticket results → [3] Identify top issue pattern → [4] Draft Slack message → [5] Send via Slack API
🔧
5. Tool use — taking actions in the world
Agents are connected to tools — APIs, web browsers, code interpreters, file systems, databases, email clients, Slack, Jira, GitHub, etc. When the agent decides to take an action, it calls the appropriate tool. The tool's output feeds back into the agent's perception for the next step.
Agent calls: ServiceNow API (get tickets) → Python interpreter (parse data) → Slack API (send message). Each tool result is fed back into the agent's context.
How the agent loop works — a complete example
Three real agent workflows — how they work in Indian IT
Workflow 1 — Bug triage and resolution agent
Used at Indian IT delivery teams for Level 1 incident response
1
New Jira ticket arrives: "Login API returning 500 errors — 23 occurrences in last hour" Perception
2
Agent queries application logs for the time window and error pattern Splunk API
3
Identifies root cause: database connection pool exhaustion during peak traffic Reasoning
4
Searches knowledge base for similar past incidents and their resolutions RAG lookup
5
Drafts resolution steps and creates a Jira comment with root cause, evidence, and suggested fix Jira API
6
Sends Slack notification to on-call engineer with summary and Jira link for human approval before any fix is applied Slack API + Human checkpoint
Workflow 2 — Candidate screening agent
Used by Indian IT HR teams and staffing firms for volume hiring
1
New job application received for "Python Developer — 2–3 years experience" Perception
2
Agent reads resume PDF and extracts structured data: skills, years of experience, last role, education PDF parser + LLM extraction
3
Compares extracted data against the job description requirements and scores the candidate Reasoning
4
If score above threshold: drafts personalised email inviting candidate for technical screening Email draft
5
HR manager reviews and approves the email before it is sent — no autonomous sending without human sign-off Human checkpoint
6
Updates ATS (Applicant Tracking System) with screening status and score ATS API
Routes high-risk flags to human compliance officer for final decision — no autonomous account blocking Human checkpoint
Where AI agents are deployed in Indian IT in 2026
💻
Software development
Agents write code, run tests, identify failures, propose fixes, and open pull requests — the full development cycle with human review gates. TCS has 150+ specialised agentic AI solutions in this space.
Claude Code, GitHub Copilot Workspace, Cursor — all agent-powered
🏫
Customer support
Agents handle Level 1 and Level 2 support: look up account information, process standard requests, update records, and escalate complex cases to humans. Always on, no attrition, consistent quality.
Widely deployed at Indian banks, telecom companies, e-commerce
📋
Document processing
Agents read invoices, contracts, forms, and applications — extract structured data, validate against rules, flag exceptions, and route to the right team. Replaces manual data entry at scale.
Infosys deployed 500+ AI agents for document workflows across major clients
📈
Research and competitive intelligence
Agents browse the web, read industry reports, extract key data points, and synthesise findings into structured briefings — replacing hours of manual research per week for analysts and strategy teams.
Used at consulting teams in TCS, Wipro, and Accenture India
🤰
IT operations and monitoring
Agents monitor infrastructure, detect anomalies, correlate alerts across systems, diagnose root causes, and trigger standard remediation scripts — with human approval for anything that affects production.
HCL and Wipro cybersecurity operations — referenced in Module 1.3
🏗
Supply chain and procurement
Agents monitor supplier lead times, flag risks, compare quotes, and draft purchase orders. TCS's agentic AI solutions in supply chain operate inside client value chains, orchestrating actual business transactions.
TCS 150+ specialised agentic AI solutions — supply chain and financial services
The tools that power AI agents in 2026
Tool / Framework
What it does
Who uses it
Claude Code
Agentic coding — reads your codebase, writes code, runs tests, fixes failures autonomously. Leading coding agent in 2026.
Developers at Indian IT companies, individual engineers
Microsoft 365 Copilot
RAG + agent capabilities across Outlook, Teams, Word, Excel. Can summarise meetings, draft emails, update spreadsheets autonomously. Deployed to 330K+ employees at TCS, Infosys, Wipro.
All roles — the most widely deployed agent tool in Indian IT
LangGraph / LangChain
Python frameworks for building multi-step agent workflows. Developers use these to wire together LLMs, tools, and memory.
Developers building custom agent systems
n8n / Zapier
No-code workflow automation. Connect apps, trigger actions based on events, build simple agent-like automations without writing code. Covered in Module 4.4.
Non-developers, analysts, managers building automations
AutoGen (Microsoft)
Multi-agent framework — multiple specialised AI agents that collaborate, debate, and divide work on complex tasks.
Enterprise AI teams building production agent systems
What you need to know about agent risks
Challenges in deploying AI agents include hallucinations, inconsistent outputs, and error cascades in complex workflows — issues that can introduce serious risks, making reliability and governance critical. Most Indian enterprises are implementing strict monitoring, limited autonomy, and clear oversight to prevent operational failures.
The three agent risks that matter for Indian IT professionals:
1. Error cascades — in a multi-step agent workflow, a wrong decision in step 2 can produce a chain of correct-looking actions based on wrong premises. By step 7, the outcome is far from what was intended. This is why human checkpoints at high-stakes steps are non-negotiable in enterprise deployments.
2. Autonomous action without review — an agent that can send emails, update databases, or trigger financial transactions without human approval is a compliance and operational risk. Always know which tools your agent can use autonomously and which require human sign-off.
3. Data access and privacy — agents that can read files, query databases, and call APIs must have carefully scoped permissions. An agent with access to all company data is a security and DPDP Act compliance risk. Principle of least privilege applies — give agents access to only what they need for the specific task.
How to work alongside AI agents — your role in an agentic workflow
The human-in-the-loop is not optional — it is the product. To gain lasting value from AI agents, enterprises must reimagine workflows and upskill employees to work alongside these systems. Success depends on balancing human judgement with automation. The engineers and analysts who understand how to design, monitor, and govern agent workflows are the most valuable people in any AI team.
Your role in an agentic workflow is not to do what the agent does — it is to:
Define the goal clearly — agents are only as good as the task definition they receive. Ambiguous goals produce unpredictable agent behaviour. This is prompt engineering applied to agent instructions.
Set the approval gates — decide which steps in the workflow require human review before the agent proceeds. Sending an email? Human review. Querying a database? Autonomous. Updating a production record? Human review.
Monitor and debug — agents fail silently in ways that are harder to catch than code errors. Reviewing agent logs, spotting unexpected tool calls, and identifying when the agent took a wrong turn are increasingly valued skills.
Validate outputs — the agent’s final output is not automatically correct. Verify against the original goal. Check that each step produced the intended result. This is the human-in-the-loop role that Indian IT companies are actively hiring for.
The three things to remember
KEY TAKEAWAYS — MODULE 4.2
AI agents act, chatbots answer — the defining difference is tool use and multi-step autonomous action. An agent can browse the web, run code, call APIs, and complete a multi-step task without human approval at every step. Understanding this distinction is now a baseline expectation in Indian IT interviews.
The five components are your vocabulary — perception, reasoning, memory, planning, and tool use. When you hear "agentic AI" in a meeting or interview, you can now decompose any agent system into these five components and describe how each works. This is the conceptual fluency that separates informed professionals from passive observers.
Human-in-the-loop is the career opportunity — enterprises are not deploying fully autonomous agents for high-stakes tasks. They are deploying supervised agents with human checkpoints. The professionals who know how to define goals, set approval gates, monitor workflows, and validate outputs are the most valued people in any AI deployment team.
✎
Assignment 4.2 — Map an AI Agent Workflow for a Real Task
This assignment develops the workflow design skill that Indian IT companies look for in AI delivery roles.
Choose a repetitive task in your target domain — something currently done manually that involves multiple steps, multiple systems, or both. Good candidates: weekly report generation, ticket triage, data extraction from documents, candidate screening, status update emails.
Map it as an agent workflow using the five-component framework:
Goal: What is the agent trying to accomplish? (One clear sentence)
Tools needed: Which systems does the agent need to access? (APIs, files, databases, email)
Steps: Break the task into 4–7 numbered steps the agent would take
Human checkpoints: At which steps should a human review before the agent proceeds?
Risk: What could go wrong if the agent makes a wrong decision at step X?
Add to your AI Learning Journal under "Module 4.2 — Agent Workflow Design." This pairs with your RAG use case from Assignment 4.1 — together they show you can think at an AI systems level, not just a tools level.
Being able to design a human-supervised agent workflow on a whiteboard — identifying tools, steps, and approval gates — is a skill that will set you apart in every AI-related placement interview you attend from here on.