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

Agentic AI is software that understands a natural-language request, creates a plan, and chooses tools to complete the task. Its path can change as it works. Ordinary software follows rigid, pre-programmed instructions and usually needs people to guide each step. This matters because agents can handle more complicated operations with less human intervention.

For example, an agent may receive a shopping request, apply price filters, search products, compare options, and guide checkout. It does not simply run one fixed script. It interprets the request, sequences actions, and invokes specialised systems, such as search engines or internal databases.

Amazon is already using this model in Rufus and seller assistants. The article presents agentic AI as a shift toward an operating layer that can coordinate work. It also notes that agents should work alongside traditional machine-learning and search systems, rather than replace every existing technology.

Amazon uses AI agents as practical operating tools, not merely chat interfaces. These systems interpret requests, decide which steps are needed, and connect to specialised digital tools. That can reduce human intervention in shopping and seller workflows while allowing the system to complete several linked tasks.

Rufus is the clearest example. It parses a customer’s request, applies price filters, invokes search engines, and guides the customer through product comparisons and checkout. Autonomous seller assistants work on the business side. They pre-fill product listings, monitor inventory levels, and guide small businesses through international export compliance.

The key mechanism is orchestration. An agent understands intent and dynamically creates a sequence of actions, while underlying tools perform specific jobs. Amazon’s examples show how agentic systems can connect natural-language requests with search, product data, inventory information, and compliance processes.

The operational scale can reach tens or hundreds of billions of products in ecommerce search. That is a huge retrieval workload, especially when users expect answers quickly. The article does not provide a multiplier comparing this workload with tasks generative AI handles well. It does make the important distinction that generative AI is better suited to reasoning, summarisation, review analysis, and creative editing.

At this scale, a system must retrieve relevant information rapidly across an enormous catalogue. Generative models can become slow and expensive when asked to perform that basic retrieval directly. A traditional search engine is designed for high-volume lookup and can return results with much lower delay.

The practical implication is architectural, not simply numerical. An agent can understand what the shopper wants, then hand the retrieval task to a conventional search system. This division lets enterprises support massive operational workloads without forcing a generative model to do every computation.

Enterprise systems divide the work because different technologies are good at different jobs. Generative AI can interpret complex language and reason about a request. Traditional machine-learning models process structured numerical data quickly. Search engines handle large-scale retrieval with low latency. Using one model for everything would ignore these differences and increase cost or delay.

An agentic orchestrator acts as the coordinator. It parses the user’s intent and decides which tools to call. In ecommerce, it can understand a shopping request, apply filters, and then send the retrieval task to a traditional search engine. The search engine can scan enormous product collections more efficiently than a generative model.

This approach reflects the article’s central warning about GenAI hype. Generative models are not a universal cure. Enterprises can reserve them for reasoning, summarisation, and creative work, while established models and search systems handle fast, repetitive operations. The result is a more practical operating layer.

Fine-tuning means adapting a smaller open-base model, such as Nyaya, with proprietary application data. For specialised classification or prediction, that model may match or outperform a giant closed model while costing one-fifth or one-tenth as much. This could make advanced AI more affordable for enterprise teams and reduce dependence on the largest systems.

The mechanism is targeted training. Instead of using a frontier model built for many broad tasks, a business teaches a smaller model its own data and use case. The resulting system focuses its capacity on a particular application. That can improve accuracy for specialised work while requiring less computation.

The likely enterprise effect described in the article is lower computational expense and operational overhead. The source does not state a specific energy reduction, so energy savings cannot be quantified from it. Still, choosing a smaller model can make deployment more economical and may support wider use of AI in business operations.

Standard operating procedures are natural-language documents that explain how workers should resolve complex issues. They are a promising target because many business processes already exist as written instructions, yet still require people to interpret and execute them. Automating these procedures could improve consistency and reduce manual work across law, finance, healthcare, agriculture, and human resources.

An agentic platform can clarify poorly written instructions using internal knowledge bases. It can split a long procedure into smaller tasks and assign them to specialised sub-agents. Those sub-agents may browse internal tools, extract information from documents, analyse images, and resolve issues. The system therefore converts prose into coordinated action.

Rastogi argues that building new frontier models demands unsustainable financial capital and computing infrastructure. Building on existing frontier models to automate specialised workflows is more promising. The opportunity lies in solving concrete enterprise problems rather than competing to create another massive general-purpose model.

Each component has a distinct job. Traditional machine-learning models process structured numerical data quickly, using methods such as gradient-boosted decision trees or standard neural networks. Search engines retrieve information from very large collections with low latency. An orchestration system interprets intent, prepares a plan, and invokes the right tools. Together, they divide work according to capability.

In Amazon’s shopping example, an orchestrator can understand a customer’s request and apply its conditions. A search engine then retrieves products rapidly. Another model or tool could classify data, analyse reviews, or process images. The agent coordinates these steps instead of performing every task itself.

This design can be more practical than one giant model. It avoids using expensive generative computation for routine retrieval and structured processing. The article points toward networks of specialised sub-agents and models for enterprise workflows. Future AI may therefore depend on cooperation among focused systems, although the article does not specify a final architecture.