“Clearly the entire operating layer is becoming agentic,” said Rajeev Rastogi, Vice President of Machine Learning at Amazon, pointing to a decisive transformation in business technology.
This marks a transition from static software to autonomous systems capable of understanding natural language instructions. Unlike conventional software that follows rigid, pre-programmed scripts, agentic systems possess dynamic execution paths.
“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 flagship tech-startup event, TechSparks.
Instead of relying on human intervention at every step, these autonomous agents evaluate queries, formulate a sequence of actions, and call specialised digital tools to complete complex tasks end-to-end.
Within Amazon, this operating layer is active in systems like Rufus, a conversational shopping assistant that parses customer requests, applies price filters, invokes search engines, and guides users through product comparisons and checkout journeys. Similarly, autonomous seller assistants pre-fill product listings, monitor inventory levels, and guide small businesses through international export compliance.
Beyond the generative AI hype
Despite intense market enthusiasm surrounding generative artificial intelligence (Gen AI), Rastogi emphasised that generative models are not a universal cure for every computing challenge. Generative AI excels at complex reasoning, text summarisation, review analysis, and creative tasks such as editing product image backgrounds. However, it struggles with high-volume, low-latency applications.
“For certain applications, Gen AI models will never match the performance that you can get from traditional machine learning models,” Rastogi pointed out.
Traditional machine learning relies on classical algorithms such as gradient-boosted decision trees or standard neural networks that process structured numerical data with extreme speed.
For large-scale operational needs such as ecommerce product search handling tens or hundreds of billions of items, generative models suffer from prohibitive computing costs and slow response times, he said. High-performance architectures therefore use agentic orchestrators to parse human intent while relying on underlying traditional search engines for lightning-fast data retrieval.
Where generative capabilities are genuinely required, enterprise leaders do not need to rely exclusively on massive foundation systems. Frontier models, which represent the largest, most computationally heavy artificial intelligence architectures, carry substantial operational overheads.
Rastogi highlighted a far more economical alternative, noting, "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 involves taking a smaller, publicly available open-source model, such as Qwen, and training it on specific proprietary application data. By adopting fine-tuned open models, businesses can achieve equal or superior accuracy for specialised classification or prediction tasks compared to giant closed models like Claude or ChatGPT, while dramatically lowering their computational expenses.
Unlocking the white spaces
For technology entrepreneurs and enterprise architects, the most compelling opportunities lie in vertical workflow automation rather than attempting to construct foundational models. Developing new frontier models requires unsustainable financial capital and computational infrastructure.
Instead, Rastogi advises that "building on top of the frontier models, using agentic AI to automate workflows in domains... that’s definitely promising.”
Taking inspiration from coding tools like Cursor, major opportunities exist across specialised domains including law, finance, healthcare, agriculture, and human resources. A particularly valuable white space lies in automating standard operating procedures, which are natural language documents detailing step-by-step instructions for human workers to resolve complex operational issues.
An agentic automation platform can clarify poorly written instructions using internal knowledge bases, break down long procedures into manageable sub-tasks, and delegate them to networks of specialised sub-agents that automatically browse internal tools, extract document data, analyse images, and resolve issues.
Looking ahead, the underlying design of AI models must evolve to remain sustainable. Current frontier models face severe physical and financial constraints.
“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 noted.
Furthermore, today’s monolithic models function primarily as statistical next-token prediction engines, guessing the most likely next word in a sequence without a genuine understanding of physical laws, 3D spatial geometry, or cause and effect.
According to Rastogi, to overcome these boundaries, the future landscape will transition toward networks of smaller, domain-specific models directed by central routing orchestrators. Rather than storing vast amounts of world knowledge inside giant model parameters, future AI architectures will separate core reasoning capabilities from data storage, relying on external, high-speed search engines for information retrieval.
Eventually, the industry will advance towards genuine world models that possess an internal understanding of physical spaces and environmental behaviour, establishing a practical, cost-effective framework for enterprise automation.