For two years, enterprise AI has summarised, drafted and recommended, then left a human to take the next step. Agentic AI removes that step. An agent interprets intent, reaches into enterprise systems, reasons across multiple steps and executes within defined limits.
In India, that shift is moving from pilot to production, and the early evidence suggests the real disruption is not to productivity but to how software is used at all.
The app becomes optional
In January 2026, Swiggy connected Food, Instamart and Dineout to AI assistants through the Model Context Protocol. Assistants such as ChatGPT, Claude and Gemini can now search, compare, build carts, apply offers, place orders, track deliveries and book tables. Instamart was positioned as the first quick-commerce platform globally to integrate MCP, with 40,000-plus SKUs exposed. Authentication runs on OAuth, so credentials never pass to the agent.
The strategic point: the AI becomes the interface, and Swiggy becomes the execution layer behind it. In April, Swiggy’s Builders Club opened that layer to outside developers, with three MCP servers and more than 18 API tools at launch. Swiggy calls it a shift from platform to ecosystem orchestrator.
Retail is following. Reliance Retail-backed Fynd says its agent has handled more than 4.3 million interactions across web, mobile and WhatsApp. BigBasket piloted grocery purchases inside ChatGPT, settled through UPI Reserve Pay on Razorpay’s stack. The shopping site starts to look like a backend.
Agents move onto the factory floor and into the back office
Tata Steel says it deployed 300-plus specialised agents in nine months, part of 860 models and agents across its value chain in FY2025-26. Its digital assistant resolves more than 70% of routine HR tickets autonomously, and agents that triage customer complaints have cut average turnaround by 50%, the company reports. On the plant floor, Safety EyeQ flags hazards from live video, while Asset Sphere agents generate proactive maintenance plans.
Mahindra runs agents at both ends of the value chain: a self-healing paint shop and an agentic maintenance system in manufacturing, and WhatsApp agents that have handled about 400,000 customer conversations, from lead nurturing to test-drive bookings.
HDFC Bank is building a unified platform on MCP, Agentic Studio and Agentic Mesh. As of April, five use cases were in production and 14 in development, targeting faster turnaround and first-time-right outcomes. Razorpay’s Agent Studio, launched in March, targets operational leakage: failed subscription payments, disputes, abandoned carts, risky cash-on-delivery orders and settlement analysis.
Similarly, Flipkart’s payments platform, Super.money is rolling out agents that can autonomously shop for products and buy gold, and it has about 20 million monthly active users. The gold purchases execute automatically once prices fall to levels the customer sets. The agents are meant to expand later into bill payments and investments. This is the strongest new example. It is a live agent moving real money under a customer-set rule, which fits your delegated-trust argument well.
The pattern matters. A chatbot saves an employee a few minutes. An agent watching equipment, safety conditions or payment failures intervenes before the loss occurs. Value shifts from assistance to operational intervention. It also explains where enterprises are starting: workflows with clear baselines, such as failed payments, ticket resolution and equipment downtime, rather than open-ended assistants.
Payments and trust decide whether this scales
An agent can find a restaurant and fill a cart. Someone still has to pay, and India’s payment rails were designed around a human approving each transaction.
Pine Labs’ P3P protocol lets an agent complete a UPI payment after upfront consumer authorisation, with delegated permissions, spending controls and auditability. Mastercard demonstrated authenticated, tokenised agentic transactions in India this year with banks, aggregators and merchants including Swiggy, Instamart, Tira and Zepto. NPCI’s AtOM platform goes further, using agent-to-agent communication to coordinate changes across NPCI, banks and payment providers.
Autonomy also changes the risk equation. A generative model that gives a wrong answer is a quality problem. An agent that gives a wrong answer and acts on it is an incident. The governance questions are blunt:
1. What is this agent allowed to do, and who authorised it?
2. Which systems and data can it touch, and which can it not?
3. Can we reconstruct every decision, and can we stop it?
The most important layer in the agentic stack may not be the LLM. It may be identity and permissions: a digital equivalent of power of attorney, where authority is delegated, bounded, logged and revocable.
In practice, that means treating agents as a new class of user, with their own scoped credentials, spending limits, a complete action log and a working kill switch. The enterprises that build these controls before extending autonomy, rather than after, will be the ones able to scale.
Why India has an edge, and where the moat sits
India does not have to build agent infrastructure from scratch. UPI, Aadhaar-based identity, digital KYC, account aggregators, ONDC and API-first fintech already exist; agents can plug into them.
The winners will not necessarily own the best model. Tata Steel’s advantage is decades of industrial data. Swiggy’s is its merchant, inventory and fulfilment network. Razorpay and Pine Labs sit inside payment workflows. The LLM may be interchangeable. The enterprise context is not.
Enterprise software follows the same logic. Applications become capabilities that agents invoke. The application remains the system of record and execution; the agent becomes the system of interaction. That puts a premium on exposing core workflows as governed, well-documented APIs, as Swiggy has done with its MCP servers. A business an agent cannot invoke is a business an agent will not choose.
The bottom line
India moved from digitisation to platformisation, and platforms produced APIs. Agents now connect those APIs to real-world execution. The next generation of Indian digital businesses may be built less around applications people use and more around capabilities that agents invoke.