Indian eCommerce runs on volatility. Prices shift, rankings move, competitors react, and RTO risk changes by the hour. Yet many forecasting systems were built for a world where customer profiles mattered more than marketplace behaviour. That assumption no longer holds. The future of accurate forecasting is moving toward cleaner, privacy safe datasets that capture what buyers do, not who they are. With DPDP reinforcing minimal data practices and platforms tightening access to identifiers, agentic AI is emerging as the architecture that can read these signals in real time and turn them into precise, actionable forecasts.
CEO & Co-founder
Graas.ai
Agentic AI does not try to understand who the customer is. It focuses on what is happening across the marketplace ecosystem. This distinction matters. Brands start their day by checking which listings on Amazon dropped in rank, which Flipkart keywords lost efficiency, which categories spiked, which PIN codes show rising RTO risk, and how competitor prices moved overnight. None of these signals require personal identifiers. They require clarity, speed, and context.
The limitations of identity-heavy forecasting
Traditional forecasting models blended sell-through rates, inventory depth, elasticity curves, promotions, and seasonality. They were useful but slow to adapt. They lived inside dashboards, where analysts interpreted charts and made decisions in weekly or monthly cycles. In fast-moving marketplaces, that delay erodes margin.
These systems were also fragile. When platforms changed APIs or restricted identifiers, models often needed retraining. Accuracy dropped. Engineering costs rose. Forecasting became something brands maintained rather than relied on.
The rise of event-level forecasting
Agentic AI flips this dependency. It operates on real-time event streams such as price changes, stock movements, keyword trends, return reasons, competitor actions and seasonality. These datasets are privacy safe, DPDP aligned, and closer to the operational truth of how Indian marketplaces behave.
A forecasting engine built around event streams reacts in hours instead of days. And in eCommerce, speed is margin. 1% discount change can alter Buy Box visibility. A stockout can push an SKU to page three. An accurate forecast must evolve as fast as the market moves.
Examples of how agentic AI changes forecasting
SKU-level demand forecasting without customer profiles
An agent monitors sell-through velocity, marketplace search patterns, stock depth, competitor pricing, and seasonality. It anticipates surges or dips without accessing customer identifiers. This reduces stockouts, prevents overstocking, and gives brands a clearer view of what to restock and when. Clean signals outperform deep profiling in most scenarios.
Forecasting RTO risk by PIN code
Return to origin is one of Indian commerce’s biggest margin drains. Instead of relying on user histories, agents study courier delays, local weather disruptions, product-category return tendencies, delivery attempt patterns, and historical PIN-code behaviour. This helps brands identify emerging risk pockets early and adjust inventory placement, payment modes, or communication strategies.
Dynamic pricing and elasticity forecasting
Amazon and Flipkart reward relevance and conversion, and price is a major driver of both. Agents observe competitor prices, anonymised demand curves, and ad-response signals to model price sensitivity in real time. They recommend discount bands that protect margin without sacrificing visibility. A human or a dashboard cannot run this pricing loop at marketplace speed. An agent can.
Agentic AI as a new architecture for decision-making
The shift is deeper than improved accuracy. It is architectural. Dashboards sit beside brands. Agents sit inside the workflow. They monitor signals. Interpret shifts. Recommend actions. Execute them when permitted.
Agents do not replace teams. They collapse the gap between knowing and acting, which is where most leakage happens today. They eliminate the latency between insight, decision, and execution.
DPDP’s focus on purpose-bound data processing aligns naturally with event-level forecasting. Systems built on operational signals depend less on personal identifiers, which makes them more resilient when platforms restrict identifiers. This stability reduces retraining cycles and lowers long-term engineering cost.
The future of forecasting is collecting cleaner data and shifting from identity-driven models to signal-driven models. It is about replacing static dashboards with agentic execution. It is about systems that adapt to the speed, volatility, and complexity of Indian marketplaces.
The tools have changed. The terrain has changed. The playbook must change with it.
–Authored by Prem Bhatia, CEO & Co-founder, Graas.ai