Beyond the GenAI hype: Amazon's Rajeev Rastogi bets on agentic AI to boost enterpise operations

“Clearly, the entire operating layer is becoming agentic,” said Rajeev Rastogi, Vice President of Machine Learning at Amazon. He said business technology is going through a major change.

This means moving from software that follows fixed instructions to systems that can understand everyday language and act on their own. Unlike traditional software, which follows a set of steps written in advance, agentic systems can choose different ways to carry out a task.

“Agents prepare plans dynamically to solve a problem... and invoke the appropriate tools,” Rastogi explained during a fireside chat at the 17th edition of YourStory’s main technology startup event, TechSparks.

Instead of needing a person to guide every step, these AI agents understand requests, plan a series of actions, and use specialised digital tools to complete complex tasks from start to finish.

At Amazon, this approach is already used in tools such as Rufus, a shopping assistant that talks with customers. Rufus understands what customers ask for, applies price filters, uses search tools, and helps people compare products and move through checkout. Amazon also has independent assistants for sellers. These tools fill in product listings, track stock levels, and help small businesses understand the rules for exporting goods to other countries.

Beyond the generative AI hype

Although generative artificial intelligence (Gen AI) is attracting a lot of attention, Rastogi stressed that it is not the best solution for every computing problem. Gen AI is good at complex reasoning, summarising text, analysing reviews, and creative jobs such as changing the backgrounds of product images. However, it can struggle when a system must handle huge amounts of work and respond very quickly.

“For certain applications, Gen AI models will never match the performance that you can get from traditional machine learning models,” Rastogi said.

Traditional machine learning uses established methods, such as gradient-boosted decision trees and standard neural networks. These methods can process structured numerical data very quickly.

For large tasks, such as searching through tens or hundreds of billions of products in an online store, generative AI models can cost too much to run and take too long to respond, he said. To work quickly, businesses can use an agentic system to understand what a person wants, while traditional search engines find the needed information at high speed.

When a task does need generative AI, business leaders do not have to use only the biggest AI systems. Frontier models are the largest AI models and need enormous amounts of computing power, which makes them expensive to run.

Rastogi pointed to a cheaper option: “Fine-tuning some of the open-base models... can give you models that perform a lot better than some of these large models out there... at one-fifth or one-tenth of the cost.”

Fine-tuning means taking a smaller model that is publicly available, such as Qwen, and training it further using data from a particular business or task. Businesses can use these improved open models for specialised jobs, such as sorting information into categories or making predictions. They may achieve the same or better accuracy than much larger closed models, such as Claude or ChatGPT, while spending far less on computing.

Unlocking the white spaces

For technology entrepreneurs and people who design systems for large businesses, the biggest opportunities may be in automating specific work processes rather than building basic AI models from scratch. Creating new frontier models requires huge amounts of money and computing equipment, making it difficult to sustain.

Instead, Rastogi said, “building on top of the frontier models, using agentic AI to automate workflows in domains... that’s definitely promising.”

Inspired by coding tools such as Cursor, entrepreneurs can find opportunities in specialised areas such as law, finance, healthcare, agriculture, and human resources. One important area is automating standard operating procedures. These are documents written in everyday language that explain the steps workers should follow to solve complex problems at work.

An agentic AI platform can make unclear instructions easier to understand by using a company's internal information. It can split long procedures into smaller tasks and assign them to groups of specialised AI agents. These agents can then use internal tools, collect information from documents, analyse images, and solve problems automatically.

Looking ahead, AI models will need to change to remain practical and affordable. Today's frontier models face serious physical and financial limits.

“The frontier models today are massive, computationally super expensive, consume a lot of energy and resources... and I don’t think that’s sustainable,” Rastogi said.

Today's large AI models also mainly work by predicting the next word or piece of text in a sequence. They do not truly understand physical laws, the shape and layout of objects in three-dimensional space, or how one event causes another.

Rastogi said future AI systems could use groups of smaller models designed for specific areas, with a central system directing each task to the right model. Instead of storing huge amounts of information about the world inside one giant model, these systems could separate reasoning from information storage and use fast external search engines to find the facts they need.

Over time, the industry could move towards AI models that build an internal understanding of physical spaces and how the world around them behaves. This could provide a practical and affordable way to automate work in businesses.