The glass box

The judgment, made inspectable.

The defensible core isn't the outreach - it's the judgment about who not to contact. Every gate, tier and number is a pure function you can read. Try it:

Decision coverage - click any case29 cases · every reason, gate, lever, channel & suppression rule · same input → same decision, always

Tier 1 · recover

Channel logic

Tier 2 · lever

Tier 3 · suppress

Gate ordering (edge)

…or construct one by hand

Failure reason

Order value ₹4,500

RiskShield (fraud) · gate ≥ 7815

RTO Intelligence · COD gate ≥ 6425

Local hour · quiet 21-814:00

Contacts in 7 days · cap 30

Decision traceTier 1 · AI voice call

₹4,500RiskShield 15RTO 2514:00freq 0/3contact ✓consent ✓voice-consent ✓DND offsensitive ✗stock ✓serial ✗

0 Idempotency · dedup · completion check
Not yet paid on any session; event is unique (idempotency key clean).
1 Fraud / risk gate (RiskShield)
RiskShield score 15 < 78: not a fraud pattern.
2 Reachability + consent gate
Valid, consented contact; identity not flagged shared / reassigned.
3 Safety + fatigue gate
No sensitivity, fatigue, staleness or quiet-hours conflict.
4 Classify failure reason → tier
Tier 1 · Bank / PSP downtime - Money and intent are present; the rail blipped.
5 Select fix / lever
Alternate method now; retry when the rail is healthy
6 Select channel
AI voice call · Voice eligible: consented, in-hours, high value, economics clear.
7 Control-group assignment
In the contacted arm (holdout 0%).
8 Execute (secure link · grounded message)
Sent via AI voice call with a single-use, current-cart link.
9 Monitor + cancel-on-completion
Watching for completion on any channel; pending outreach cancels on completion.
10 Attribute + cost
Per-attempt cost 6.00; recovery measured vs control.

The numbers

Modelled win rate: 68%Attempt cost: ₹6.00Voice value-gate margin: ₹370Voice economics: clears

AI voice call script

Hi, this is Indie Threads's assistant. Your payment of ₹4,500 didn't go through a moment ago - looks like the session dropped. I can send a fresh secure link to your WhatsApp right now to pay by card or UPI. Shall I?

Rules decide

Fraud, consent, tier, suppression, channel - deterministic and auditable. No LLM decides who to contact.

AI talks

The message & voice script are LLM-written (OpenRouter), grounded in the failure reason, with a template fallback.

Tools decide who/whether; a language model writes the last mile. See it live on the demo - fire a real email and the copy is LLM-generated.

The decision pipeline

Every event runs the gauntlet - suppression happens before any contact

Event

failed / abandoned

gate 0

Idempotency

already paid?

↓ suppress

gate 1

Fraud · RiskShield

card-testing?

↓ suppress

gate 2

Consent + reach

allowed to?

↓ suppress

gate 3

Safety + fatigue

sensitive? quiet hours?

↓ suppress

✓ Eligible

Tier 1 or 2

Left alone, logged: fraud · no consent · already paid · sensitive item · serial-abandoner · out of stock · frequency cap · quiet hours

Classify → Tier

reason → T1 / T2

Pick fix + lever

link · EMI · COD (RTO-gated)

Pick channel

voice* · WhatsApp · SMS

Hold-out

control group

Measure

incremental, vs control

* AI voice is gated by value, consent, DND and lawful hours - and ships behind a legal flag. WhatsApp + SMS work without it.

The recoverability taxonomy

The honest answer to “will they still not buy even if we reach them?”

Failure / abandonment event

↓ routes into one of three tiers ↓

Tier 1 · Recover

Bank downtimeUPI timeoutSession crashTechnical decline

Reason-tailored fix, in seconds

Tier 2 · Lever

Insufficient fundsCOD-seekerPrice shock

EMI / Pay-Later · COD (RTO-gated) · clarify

Tier 3 · Suppress

FraudNo consentAlready paidSensitive item

Do not contact - logged with a reason

Attribution: incremental, not vanity

Gross - “paid after we messaged”₹11,90,400

what spray-and-pray would claim

− Control proves they'd have paid anyway₹3,47,800

subtract the baseline

= Incremental - the only number we report₹8,42,600

the lift we actually caused

Illustrative. In the live demo this is computed every tick from the real contacted vs held-out cohorts.

Unit economics: gate the expensive channel

Voice's extra lift over WhatsApp must out-earn its extra cost. In practice the binding gates are the value floor + consent + DND + lawful hours; the economics below is the sanity floor.

use_voice  iff  value × extra_lift
                > extra_cost × safety_margin

Channel	Cost / attempt	Role
AI voice	₹6.00	flagged · ≥ ₹2,000
WhatsApp	₹0.85	the workhorse
SMS	₹0.18	fallback

Illustrative pass-through costs.

What's real vs what's simulated

An independent concept composing real, separate Cashfree products into one recovery agent. Hover any row for the exact scope. Not affiliated with Cashfree.

Real in this build · live code

Decision engine · traceable per event
Attribution math · holdout + Wald CI
CSV import · recomputes in-browser
Live LLM copy · rendered as a preview

Simulated · not built yet

Events, customers & amounts · synthetic
Every rupee / rate · conditional on priors
Gateway webhooks & WhatsApp/SMS/voice · not wired
RiskShield / RTO scores · synthetic stand-ins

Real Cashfree products it fuses

Cart Recovery · cashfree.com/docs/payments/checkout/cart-recovery
RTO Intelligence · cashfree.com/docs/payments/checkout/cod/rto
RiskShield · cashfree.com/risk-shield-payment-gateway

Benchmarks it reads against

~70% of online shopping carts are abandoned on average (70.22% aggregate of 50 studies, 2006-2025). · Baymard Institute, Sep 2025
Prepaid has overtaken COD in tracked D2C volume (~62% prepaid vs ~38% COD), though COD still leads many tier-2/3 markets. · Unicommerce, CY2024