AI agents for business in India: 20 that companies actually run

Most lists of AI agents for business in India are lists of software. This one is a list of jobs. We went looking for agents that Indian companies, and global platforms available in India, have publicly said they are running: from the lead qualifier on WhatsApp to the bot that matches invoices before your accountant looks at them. Each entry below has a dated, sourced example.

Two cautions before you read on. First, nearly every figure here is reported by the company itself, so treat it as a claim worth testing, not a guarantee. Second, the agents that work are narrow. They do one job well, inside rules someone wrote, and they hand over when the rules run out. (If the words “chatbot”, “agent” and “agentic application” still blur together, start with our plain-language explainer.)

That is why every agent on this list comes with two things: a named owner in your team, and a review step. Those are our recommendations, based on how we build these systems, not a description of each company’s internal process.

Sales and leads: where AI agents for business in India start

1. The WhatsApp lead qualifier

It answers enquiries at any hour, asks a few qualifying questions, recommends products and books appointments. On 14 May 2026, Meta launched Business AI inside the WhatsApp Business app in India, working in major Indian languages; Meta says one seller, The Purple Sunset, closes 6 to 7 orders a day through it. Owners can take over any chat or switch the AI off.

Owner: sales lead. Review step: read a sample of chats daily and take over anything about price, discounts or delivery dates.

2. The voice relationship and cross-sell agent

It calls existing customers in their own language, checks in and spots the next product they might need. Mahindra Finance and Sarvam reported in September 2026 that their voice agents had completed over one crore calls across 12 Indian languages, covering sales, collections and employee feedback, with every call recorded and analysed.

Owner: head of sales. Review step: a person listens to a weekly sample and approves any change to the offer script.

3. The property enquiry voice agent

Real estate runs on phone calls, and this agent takes the routine ones. NoBroker built ConvoZen to handle its own call volume; Business Today reported on 24 February 2026 that about 30 to 35% of NoBroker’s 6 to 6.5 lakh daily conversations were already handled by voice AI. The co-founders said complex and sensitive cases would stay with human agents.

Owner: sales operations manager. Review step: escalation rules reviewed monthly; any complaint call goes straight to a person.

4. The deal analyser

It reads your pipeline and suggests a win probability, the next best action and a follow-up for each deal. Zoho announced a Deal Analyser agent on 17 July 2025 at Zoholics India in Bengaluru, alongside a Revenue Growth Specialist for upsell and cross-sell.

Owner: sales manager. Review step: suggestions are advice; a person decides which follow-ups are sent.

5. The shopping and checkout agent

This one moves from recommending to paying. On 9 October 2025, NPCI and Razorpay announced a pilot with OpenAI (reported by Reuters) to let users shop and pay by UPI inside ChatGPT, with BigBasket among the first merchants and UPI Reserve Pay letting users set aside funds for a specific merchant. It was a pilot at the time of reporting.

Owner: e-commerce head. Review step: spending limits set by the customer, and a person reviews disputed or unusual orders.

Customer support

6. The support voice bot

It picks up the phone, understands the question and resolves the routine ones. TechCrunch reported on 26 November 2024 that Meesho’s generative AI voice bot was handling about 60,000 calls a day in English and Hindi; Meesho claimed 95% resolution, with 5% needing a human, and said its support staff moved to complex queries and seller support.

Owner: customer support head. Review step: the 5% that escalate are reviewed weekly to fix the bot, not just the ticket.

7. The support chat and ticket agent

It reads incoming chats and tickets, answers the common ones and routes the rest. Zomato launched Nugget on 17 February 2025, saying it resolves up to 80% of queries on its own and was already handling over 15 million support interactions a month across Zomato, Blinkit and Hyperpure.

Owner: support team lead. Review step: refunds and goodwill credits above a set amount need a person’s approval.

8. The travel service virtual agent

Flight status, baggage, check-in, loyalty points: the questions that flood an airline’s contact centre. Microsoft’s Air India case study, published 15 November 2024, says the assistant covers over 1,300 topics, fully automates 97% of sessions and escalates to contact centre staff when it detects a customer needs more help.

Owner: contact centre manager. Review step: escalation triggers and answers on policy topics reviewed whenever a rule changes.

9. The insurance policy-servicing bot

It handles policy queries, changes, renewals and even takes a customer from proposal to payment on one call. Business Today reported on 3 August 2026 that about 70% of ICICI Lombard’s inbound calls now go through a bot or its WhatsApp chatbot, up from roughly 10% a year earlier, with around a million inbound and outbound calls a month.

Owner: head of customer service. Review step: anything touching a claim decision or a policy exclusion goes to a licensed person.

10. The student doubt-solver

Ed-tech has its own version of support: questions at midnight before an exam. PhysicsWallah said in its results coverage on 9 December 2025 that its AI platforms, including AI Guru and automated doubt-solving and grading, had resolved more than 100 million student queries, and its grading engine had checked over 700,000 subjective answer sheets.

Owner: academic head. Review step: faculty spot-check answers by subject, and any grade a student disputes is rechecked by a teacher.

Human in the loop: support agents are judged by the cases they hand over, not the ones they close. Read the escalations every week. They tell you where the agent is guessing, and where your own policies are unclear.

Finance and back office

11. The collections reminder agent

It calls or messages before and after a due date, in the borrower’s language, and logs what was promised. The same Mahindra Finance and Sarvam deployment covers the full collections cycle, from pre-due reminders to later-stage conversations, connected to the lender’s CRM and loan servicing systems. Collections calls reach people under financial stress, so tone, consent and a quick route to a human matter as much as volume.

Owner: collections head. Review step: compliance reviews scripts before launch, and any hardship or dispute is routed to a person the same day.

12. The chargeback and dispute responder

It gathers order, delivery and payment evidence and drafts a response to a payment dispute. Razorpay launched Agent Studio on 12 March 2026 at FTX’26 in Bengaluru, with a Dispute Responder Agent among the first agents, built using Anthropic’s Claude Agent SDK.

Owner: finance or payments operations manager. Review step: a person approves responses on high-value disputes before they are filed.

13. The reconciliation bot (bank and GST)

This is the least glamorous agent and often the most useful. Much of it is plain matching rather than AI: Tally’s banking and accounting page describes “one-click download, reconcile and auto-accept for matched invoices” under its Connected IMS feature, and Razorpay’s Agent Studio launch describes matching bank statements against settlements in its Agentic Dashboard. The agent’s real job is the leftovers: the mismatched GSTIN, the short payment, the UPI credit with no invoice number.

Owner: finance controller or your CA. Review step: the accountant approves every adjustment before it is posted or filed.

14. The KYC and document verification agent

It reads a PAN card, GSTIN certificate or utility bill, checks it and flags anything that does not match. Zoho’s July 2025 launch included India-specific agents for PAN, Voter ID, Udyog Aadhar, GSTIN, driving licence, LPG connection and electricity bills. CEO Mani Vembu said the agents run on a low-code platform “so that there is a human in the loop for verification and modification.”

Owner: onboarding or compliance lead. Review step: a person clears every flagged document; nothing is rejected by the agent alone.

Operations

15. The delivery confirmation agent

For D2C brands, a failed cash-on-delivery order is lost money twice. Shiprocket’s April 2023 product update describes Delivery Boost, an AI-backed system that asks the customer on WhatsApp whether they can receive the order or want it reattempted later; if there is no reply, the seller can call from the dashboard.

Owner: fulfilment manager. Review step: someone checks the daily list of non-responders and high-value orders before reattempts are booked.

16. The network and IT operations agent

It watches systems, diagnoses routine faults and drafts the fix. On 17 February 2026, Infosys and Anthropic announced a partnership starting with a centre of excellence for telecom, where agents are meant to modernise network operations and service delivery, before moving into financial services and manufacturing.

Owner: IT or network operations head. Review step: any change to a live system needs an engineer’s approval and a rollback plan.

Human in the loop: the rule that keeps operations agents safe is simple. They can read, diagnose and draft as much as they like. Anything that changes a live system, moves money or cannot be undone waits for a named person to approve it.

Marketing

17. The ad creative and catalogue agent

It generates backgrounds, resizes images, writes variations and adapts ads to brand colours and tone. At Cannes Lions in June 2025, Meta announced new generative tools for Advantage+, and Storyboard18 reported a BigBasket catalogue ad test with 2.7% higher click-through and 3.4% lower cost per install than the control.

Owner: marketing manager. Review step: a senior designer or brand lead approves every new creative before it runs. (This is the part of our own work where AI does the heavy lifting and a senior person signs off.)

18. The abandoned-cart recovery agent

It contacts a shopper who left checkout, asks why, and sends a payment link. Razorpay’s Agent Studio launch in March 2026 included an Abandoned Cart Conversion Agent working by voice or message, built with partners including Nugget by Zomato and SuperU.

Owner: growth or e-commerce lead. Review step: marketing approves any discount the agent may offer, with a hard cap, and checks opt-in before anyone is contacted.

HR

19. The recruiter sourcing agent

It turns a job description and the hiring manager’s notes into a shortlist, and drafts outreach. LinkedIn said on 3 September 2025 that Hiring Assistant, its first AI agent for recruiters, would be available globally in English by the end of that month; early users reported saving over four hours per role.

Owner: talent acquisition lead. Review step: a recruiter reviews every shortlist for fit and bias, and no candidate is rejected by the agent alone.

20. The employee workflow agent

One place for employees to ask for something, with the agent doing the steps across HR, IT and finance. Darwinbox launched Super Agent on 18 September 2025, with examples such as shift swaps and intern offboarding across payroll, IT access and badges; the company said it “keeps humans in the loop for sensitive steps” and keeps a full audit trail.

Owner: HR operations head. Review step: pay, exits and access changes need a manager’s approval in the system.

What the list does not show

Every example above is a company describing its own success, usually at launch. You rarely read about the agent that was quietly switched off. And the numbers are not comparable: a 95% resolution rate on routine calls says nothing about how an agent would handle your contracts or your GST notices.

Two more Indian realities. First, consent: an agent that messages customers or reads their documents is processing personal data under the DPDP Act and Rules, so design consent, retention and deletion in from day one (we covered this for WhatsApp in our automation guide). Second, language: the deployments that scale here, from Meesho to Mahindra Finance, work in Hindi and regional languages, not only English.

How to pick your first one

Do not start from this list. Start from your own week. Write down the tasks your team repeats, how long each takes, and what it costs if one goes wrong. The right first agent is high on repetition and low on risk in draft: enquiry replies, payment reminders, reconciliation mismatches. The wrong first agent is anything that quotes a price, rejects a person or moves money on its own.

Then give it one owner, one review step and one number to watch, run it for a few weeks with a person approving every output, and widen it only when the review shows it is reliable. If it does not pay back, stop it; most AI pilots that fail fail on that discipline, not on the model.

If you would like a second pair of eyes on which agent to build first, and where the human sign-off should sit, we offer a free consultation. No pitch deck required.