Amazon ML Head Rajeev Rastogi on Why Agentic AI Is Replacing the GenAI Hype

What just happened?#

Walk into any tech office in Bengaluru or Hyderabad right now, and you will hear a quiet confession. People are exhausted by basic chatbots. The initial rush of watching an AI draft an email or write a college essay has worn off. By late 2026, businesses want to know what these models can actually finish without someone babysitting them.

That exact frustration took center stage at TechSparks 2026. Speaking at the event, Rajeev Rastogi, Vice President of Machine Learning at Amazon, cut straight through the surrounding marketing noise. His message was clear: enterprise tech is transitioning away from static software and moving toward autonomous, agentic AI systems.

For the past few years, companies rushed to slap chat windows onto their websites. Most found that answering questions is easy, but running an actual business workflow is completely different. Rastogi pointed out that the real shift happening across the enterprise operating layer is not about generating smarter paragraphs. It is about systems that can grasp natural language instructions and carry out actions across different programs on their own.

Instead of an employee opening four browser tabs, copying data between CRM dashboards, and manually filing tickets, autonomous agents are taking on the legwork. It represents a fundamental change in how software gets built and used.

How does this actually work?#

Think about how you book an emergency train ticket on IRCTC. A standard generative AI chatbot is like a well-read friend sitting beside you. If you ask for the fastest train to Madurai, it lists the train numbers, departure times, and food options. But when you ask it to actually book the berth, enter the passenger details, apply a coupon, and complete the UPI payment, it stops cold. It can only talk, not act.

Agentic AI works differently. It combines the language understanding of a large model with execution tools. You give it an objective in plain words, and the agent breaks that goal down into individual steps. It checks seat availability, calls the ticketing API, handles form inputs, flags potential issues, and only pings you when it needs final payment verification.

In a company setting, this changes the whole dynamic. Consider a typical seller support desk on a platform like Amazon or Flipkart. A supplier writes in saying fifty boxes of inventory went missing between a warehouse in Bhiwandi and a fulfillment center near Chennai.

Under the old setup, a human agent had to read the complaint, search ERP logs, track the shipment carrier, pull up the delivery receipt, and manually calculate the reimbursement. A static system could only display the data on a screen. An agentic AI workflow reads the seller's email, pulls the logistics logs via API, compares weight scans from both hubs, confirms the discrepancy, drafts the adjustment note, and queues the payout for manager approval.

The system does not just summarize what went wrong. It actively gathers the evidence and queues up the fix.

What changes for people in India?#

India is the back-office engine of global enterprise tech. Millions of professionals work in IT services, business process management, catalog operations, and customer support. When enterprise software shifts from point-and-click dashboards to autonomous agent loops, the daily routine inside those offices changes fast.

Take e-commerce cataloging as an everyday example. Small sellers across Surat, Tirupur, and Delhi upload thousands of product listings every week. Correcting faulty tax codes, checking image guidelines, and translating product descriptions used to require massive human review teams. Agentic pipelines are stepping into those middle layers, handling the bulk parsing and flagging only the edge cases for human eyes.

This does not mean Indian tech jobs vanish overnight, but the skill floor is moving up. Writing simple code routines or manually moving spreadsheet rows from one portal to another is becoming obsolete. The engineers in demand right now across Indian tech hubs are those who know how to connect models to external APIs, write strict evaluation benches, and build fail-safe guardrails.

For ordinary consumers, this shift should bring relief from terrible customer care bots. We have all suffered through banking and telecom bots that force you to pick between five pre-set options and loop endlessly without solving your issue. As agentic architectures take root, customer support tools will actually have permission to look up your account, cancel a stalled order, and trigger your UPI refund directly.

What should you do now?#

If you build software or manage an operations team, the first priority is stepping away from the surface-level AI tricks. Memorizing prompts to get clever text replies was useful three years ago. Today, that is table stakes.

Start by auditing the repetitive chores your team handles every week. Look for any process where an employee reads an incoming message, checks two internal dashboards, and performs a standard set of clicks. Those multi-app workflows are the primary targets for agentic integration.

Second, focus on clean APIs and structured data. An AI agent is only as good as the internal tools it can reach. If your company records are trapped in locked PDFs or legacy portals without API access, an agent cannot do anything useful. Building well-documented REST APIs with strict input and output schemas is the fastest way to get your business ready for autonomous tooling.

Finally, keep safety boundaries tight. As Rastogi emphasized, handing operations over to autonomous agents requires clear rules. Never give an agent unmonitored write access to production databases or unrestricted financial authority. Keep a human in the loop for high-value approvals, enforce strict API rate limits, and run automated logs on every autonomous step. Letting an agent handle the paperwork makes sense; letting it run wild without supervision is asking for trouble.