Cashfree Payments made Relay, a set of AI agents that handle the chores around collecting a payment, available to all its merchants on August 26, 2026, after a beta that began in May. The Indian provider, licensed by the Reserve Bank of India as a payment aggregator, lists five uses: retrying failed payments, following up on abandoned carts, confirming cash-on-delivery orders, recovering failed subscription payments and filing disputes before the deadline.
Founded in 2015 and based in Bengaluru, Cashfree says it serves more than 800,000 businesses and processes more than $80 billion a year. Relay is free at launch, and the company plans to move to outcome-based pricing.
Indian acquirers work on thin margins, and UPI carries no MDR
The merchant discount rate, or MDR, is the fee a merchant pays on each electronic payment it accepts. In India it is zero on UPI, the dominant instant payment system, while the same merchant pays 1.5% to 2% on a credit card and up to 0.90% on a debit card. Finance Minister Nirmala Sitharaman cited those figures on August 10, 2026, during the debate on the Taxation and Other Laws (Amendment) Bill, 2026, which the Rajya Sabha, Parliament’s upper house, passed that day. The bill makes a fee on UPI possible without imposing one. The decision rests with the steering committee headed by the National Payments Corporation of India (NPCI).
The ₹20,000 crore target, or ₹200 billion, needs a benchmark. In July 2026, UPI processed 2,366 crore transactions (23.66 billion) worth ₹29.9 lakh crore, which puts the goal at about 0.7% of a single month on that one system. Business Standard ties the target to agents that nudge buyers about to abandon a purchase, by voice call, by text or with a discount. Cashfree has not said over what period it aims to hit it.
What each agent touches in the payment chain
Retrying a failed payment generates a new authorization request. The declined transaction is closed. The next attempt carries its own identifier, goes back to the issuing bank and may be declined for a different reason. The number of retries and their spacing are set by the rules of the network used.
Recovering a failed subscription payment runs through a mandate, the standing authorization a payer gives for an amount to be debited when due, within the cap and frequency the payer set. The agent does not re-create that authorization. It either replays a debit under an existing mandate or asks the customer for a new one.
Filing a dispute enters a process whose deadlines are set elsewhere. On cards, the networks, led by Visa and Mastercard , publish the filing windows and the reason codes they accept. Evidence filed after the deadline forfeits the amount without any review of the merits, however strong the merchant’s case. On UPI, the National Payments Corporation of India runs its own procedures. The agent handles the filing calendar. The substance of the case stays with the merchant.
Abandoned-cart follow-up leaves the payment system and enters customer relations. Mayank Juneja, director of engineering at Cashfree Payments, described the practice it replaces: “Today, merchants have a growth team who call customers to understand why they abandoned their cart or if they faced issues during payment. These teams could give a customer a discount. These functions would now be powered by an AI agent.” The agent keeps the same pricing lever. Outreach to customers and pricing policy remain the merchant’s call.
| Agent | Action | What it commits |
|---|---|---|
| Failed payment | New authorization | The issuer and the network |
| Abandoned cart | Call or message, possible discount | The customer relationship and prices |
| Cash-on-delivery order | Confirmation before dispatch | Logistics |
| Failed subscription | Replayed debit or new mandate | The mandate |
| Dispute | Filing before the deadline | The network or NPCI |
Razorpay’s Vulcan tackles the same squeeze from the other end
Razorpay, Cashfree’s direct rival in Indian acquiring, unveiled Vulcan in August 2026, calling it “India’s first transformer-based AI foundation model for payments,” built, trained and hosted in India. Founder Harshil Mathur says it was trained on 4 billion payments and 3 trillion data points, and uses about 3,000 signals per transaction. Built with NVIDIA and AWS, it covers routing, fraud, risk and personalization: decisions made during the transaction itself.
The two products do not overlap. Vulcan works during authorization, where the route is chosen and a decline is decided. Relay works afterward, on what the decline leaves unresolved. Both answer the same constraint: processing margins remain thin for Indian acquirers, as Business Standard notes in presenting the agent as a new revenue line. MediaNama points out that Razorpay has not published the methodology, baseline or sample period behind its beta results.
Outcome-based pricing will hinge on attribution
Cashfree has yet to build its outcome-based pricing. Co-founder Reeju Datta cites carts converted with the agent’s help and disputes resolved as the parameters. Charging for an outcome requires attributing it, and a cart converted after an automated follow-up might have converted anyway. Business Standard reports another potential use: loan recovery.