Cashfree's Relay Goes Live for Every Merchant: AI Agents That Recover Failed Payments, Not Just Flag Them

Most of the agentic-commerce conversation this year has focused on the buying side: AI agents that shop for humans, compare products, and complete purchases inside a chat window. This week, one of India’s largest payments platforms moved the story decisively to the other side of the transaction. Cashfree Payments took Relay, its platform of AI “super agents” for payment operations, from a months-long beta to general availability for every merchant on the platform — and the agents don’t just flag problems for a human to fix. They take the actions themselves.

What Relay actually does

Launched to all Cashfree merchants at the end of August after a beta that began in May 2026, Relay automates the unglamorous operational work that surrounds getting paid — the follow-up tasks that decide whether a completed sale actually becomes collected revenue. The task list reads like a back-office job description:

• Failed payment retries — detecting dropped transactions and re-attempting them under rules the merchant configures
• Abandoned cart recovery — reaching back to shoppers who didn’t finish checkout
• Cash-on-delivery confirmation — verifying COD orders before dispatch, a workflow that matters enormously in Indian e-commerce where return rates on unconfirmed orders can be punishing
• Subscription failure management — catching lapsed recurring payments before they silently churn
• Dispute filing — submitting chargeback disputes ahead of deadlines instead of after they’ve passed

Two design details stand out. First, merchants configure Relay agents by describing an outcome in plain language or by voice — “recover my failed UPI retries after 6 PM” — rather than through rules-engine plumbing, or they can deploy pre-trained turnkey agents within minutes. Second, the agents read merchant transaction records and execute tasks directly rather than surfacing a dashboard of to-dos. Co-founder Reeju Datta framed the product as “an intelligent team member” that takes over follow-up workflows lean founder teams can’t staff.

The numbers Cashfree is claiming

The headline figures are striking, with the caveat that they are company claims rather than independently audited results. According to Cashfree’s data, seven in ten Indian SMBs handle payment operations entirely manually, and the average SMB spends close to 60 hours a week on payment ops. Relay’s stated target is to cut that to under 45 minutes. The company also projects that the agents can recover up to Rs. 20,000 crore (roughly $2.4 billion) annually in lost gross merchandise value across its merchant base — revenue that currently evaporates through failed payments, abandoned carts, and missed dispute windows.

The scale of the platform underneath matters for context. Cashfree, founded in 2015, processes over $80 billion in transactions annually for more than a million businesses and holds all three relevant Reserve Bank of India licenses: Payment Aggregator, Payment Aggregator-Cross Border, and Prepaid Payment Instrument. This is not a startup experimenting on a sandbox — it is regulated payments infrastructure extending itself into agent-driven operations.

Guardrails: approval gates and data boundaries

The most consequential design decisions are not the automation itself but the controls around it. Relay requires explicit merchant approval at two critical points: before any movement of money and before any customer contact, with all agent actions remaining visible to the merchant afterward. That architecture — agents that prepare and execute, humans who authorize the sensitive steps — is precisely the pattern that banks and regulators have been converging on this year as agentic AI enters regulated finance. Google and Singapore’s DBS Bank have published frameworks built on the same logic: approval flows, audit trails, and measurable performance around agent-driven actions.

The second guardrail is architectural. Relay runs entirely on Cashfree’s own infrastructure, and merchant transaction data is not shared with external AI providers. For a payments company operating under RBI supervision, keeping sensitive transaction records inside its own perimeter while still delivering agentic automation is as much a compliance decision as an engineering one — and it preempts the data-residency objection before merchants can raise it.

The business model bet: free now, priced on outcomes later

Relay’s pricing strategy may be its most quietly radical feature. It is free at launch for all merchants, with Cashfree planning to move to outcome-based pricing once it has collected adoption data. In plain terms: the company wants merchants using the product first, then intends to charge in proportion to measurable results — recovered revenue, saved staff hours, fewer manual errors.

This is a meaningful inversion of how enterprise software has historically been sold, and it aligns the vendor’s incentives with the merchant’s ledger. It also reflects a broader trend across AI products this year: when an agent’s output can be tied directly to recovered revenue, per-seat licensing starts to look like leaving money on the table. The risk, of course, is measurement — “recovered” revenue requires a credible counterfactual (would that cart have been recovered anyway?), and merchants should expect the eventual pricing formula to embed some contested assumptions. But as a go-to-market move, free-until-proven is hard to argue with.

Why this matters beyond India

First, it’s a working, production-grade example of agentic payments on the operational side — autonomous execution inside a merchant’s payment stack, including dispute filing, which is exactly the workflow where agent accountability, audit trails, and reputational history start to matter. Most agentic-commerce pilots so far have been demos; Relay is GA on infrastructure processing $80 billion a year.

Second, it signals where the agentic AI market is heading for SMBs. The companies most likely to benefit from task-completing agents are small businesses that can’t afford to hire dedicated operations staff — but they’re also the least able to evaluate, integrate, or supervise agent products. Embedding agents inside a platform the merchant already uses, with approval gates baked in, neatly sidesteps that adoption problem. Expect payments platforms, CRMs, and accounting software globally to study this template.

Third, the timing is not accidental. Agent interoperability and authorization standards — Google’s Agent Payments Protocol, NIST’s early work on agent identity, the A2A/MCP convergence — are being built right now, and Cashfree’s approval-gate architecture is a concrete implementation of the “verifiable authorization” principle those efforts are pushing. When the standards mature, platforms that already enforce human sign-off on money movement and customer contact will be positioned to plug in.

The open questions are real: how reliably the agents perform across Cashfree’s highly diverse million-merchant base, how outcome-based pricing will actually be calculated, and whether 45 minutes a week survives contact with messy reality. But as a marker of where agentic AI is going — from chat windows into the financial back office, with human approval at the points that count — Relay’s general availability is one of the clearest signals yet.

---

Figures are from Cashfree’s announcements as reported by TechGig, Elets BFSI, A2AWire, and Enterprise DNA on August 27–30, 2026. Recovery and time-saving numbers are company projections, not independent measurements.