Not affiliated with Cashfree.

Retail & D2C paymentsConcept · built on public Cashfree interfaces

AI recovery agents bill per recovered order. Relay Recover proves which ones you caused.

And it stops the contacts you should not make. Every failed payment or dropped checkout passes deterministic gates for fraud, consent, DND, COD risk and quiet hours; the rest get the fix their decline reason calls for; and a randomized holdout measures the lift the fee is charged on.

The problem

Recovery is now sold per outcome. The outcome is counted gross.

Gross is not caused

Some of the customers an agent contacts would have paid anyway. A per-recovery fee charges for them too, and nothing on the invoice says how many.

Generated copy breaks rules

An offer in a transactional message, a free-written SMS outside its DLT template, a cart reminder filed as a utility template, COD offered after the RTO gate said no.

Some contacts cost money

Recovering fraud buys a chargeback; chasing an order that already settled risks a double charge; messaging about a sensitive item can out a customer.

Why now

In 2026, India's payment and checkout platforms started selling recovery agents.

  • Cashfree Relay

    GA 27 Aug 2026. Its Cart Recovery agent calls the buyer with an offer: “you only pay for recovered orders”.

    Gap: Pays and reports on recovered orders - gross. No holdout or incremental measurement is described, and the cart agent leads with an offer.

    cashfree.com/relay · Analytics India Magazine, 27 Aug 2026
  • Razorpay Agent Studio

    Launched 12 Mar 2026 with a voice-led Abandoned Cart Conversion agent.

    Gap: Launch material describes conversions, not a control group, and its sample conversation closes with a loyalty discount.

    razorpay.com/newsroom, 12 Mar 2026
  • GoKwik KwikEngage

    Over 1M AI calls a month; a 24 Sep 2026 report claims a “10X incremental return”.

    Gap: “Incremental” there is voice + WhatsApp versus WhatsApp alone, not versus no contact. No randomized holdout is disclosed.

    medianews4u / Social Samosa, 24 Sep 2026

None publishes a per-decision holdout lift. See the full landscape. Assessed from each vendor's public material; Relay Recover is an independent concept, not affiliated with any company named here.

Where the signal lives

The payment rail knows why it broke. A marketing tool does not.

The payment rail sees, in real time
Decline reason code
bank down · UPI timeout · insufficient funds · limit · technical decline
Fraud score
card-testing and payment-fraud risk (at Cashfree: RiskShield)
COD / return risk
return-to-origin risk (at Cashfree: RTO Intelligence)
Instrument + status
what was tried, and whether the payment is truly final
A marketing tool sees, later
“checkout not completed”
No reason, no risk signal, no idea whether it already settled. So it can only message everyone.
The number to bill on

Gross, minus what would have happened anyway.

Gross - “paid after we messaged”₹20,48,790
what a per-recovery invoice counts
− What the control arm says would have paid anyway₹7,23,195
the baseline
= Incremental - the number to report and bill on₹13,25,595
the lift outreach caused

Reference run: 2,400 synthetic events through the live engine (seed 2026). 95% interval on the incremental figure: ₹10,08,368 to ₹16,42,822. The control room computes this live, every tick.

Upstream

Before recovery comes acceptance.

Relay Recover starts where the gateway gives up: after retries have failed. The upstream half - a decline taxonomy and smart retries, measured on an authorization ledger - is the Acceptance Intelligence console.

Acceptance console
What is real here

The decisions and the maths run live. The events are synthetic.

Every gate, fix, channel, holdout and interval on this site is computed by the engine as you watch. The merchant, customers and outcome rates are simulated until real data is dropped in.

Bring a failed-payment export

Drop a gateway export into the control room: real UPI and card decline codes are mapped and every row gets a decision - contact, hold out, or leave alone, with the rule that decided it. It runs in your browser; nothing is uploaded.