How Swiggy Is Using AI Agents To Take Commerce Beyond Its App

Swiggy is expanding its commerce services to be accessible through external AI agents, enabling users to discover food, build shopping carts, place orders, and book restaurant tables via AI assistants outside of their own app. At Inc42's inaugural 'The CTO Summit 2026', Swiggy's Chief Technology Officer Madhusudhan Rao explained how the company is making its services available to agents, rather than assuming every customer's shopping journey starts within their app.

Swiggy has implemented four Model Context Protocol (MCP) servers, which act as connectors allowing AI tools to interact with their services. These MCP servers support 66 tools across various segments such as food delivery, Instamart, Dineout, and Scenes, covering areas like discovery, menus, carts, ordering, reservations, and tracking.

By connecting their MCP servers to large language models (LLMs), users can ask the assistant to help plan meals according to a dietary schedule. Rao stated that this can be achieved by leveraging Swiggy's MCP with any LLM of the user's choice. Such integrations allow Swiggy to cater to individual preferences and requirements that would be challenging to accommodate through fixed features within their app.

Instead of developing a separate interface for each use case, Swiggy is making its core services available for AI agents to utilize. Rao emphasized the importance of building simple, reusable capabilities that can be invoked by various services, humans, and AI agents. The company is also reworking its internal AI systems to reduce reliance on a single model.

Rao shared that they previously built a contact-center agent around one model, and when the provider faced capacity constraints, transitioning away from it took nearly a month. This experience led Swiggy to invest in systems for evaluating and experimenting with alternative models, making it easier to change models without rebuilding the entire workflow.

The company now utilizes an LLM gateway to direct tasks to different models, testing alternatives in live operations. They also employ models from various cloud providers, selecting based on availability, response time, and cost. By isolating AI agents in restricted production-like environments, Swiggy aims to limit what agents can access and perform.

Rao explained that they deploy agents in separate production-like fabrics with stringent ingress and egress rules. The company is also addressing questions around agents' identities, permissions, and trust as they expand their use. Internally, Swiggy employs AI to assist delivery partners with onboarding. Their in-house assistant guides riders through the process in their local language, improving rider net promoter scores (NPS) and the onboarding funnel.