AI Recovery PRD

Priority: P0 — Razorpay AI Revenue Recovery Buildathon MVP Status: Active implementation Source of truth: Documentation/requirements/24CE064_Razorpay_AI_Revenue_Recovery_Module_Wise_Requirements_authoritative.docx Scope rule: This PRD is derived only from that authoritative module-wise requirements document.

1. Product outcome

InvoNotify helps an Indian B2B merchant recover overdue revenue safely. It turns an at-risk obligation into a traceable, bounded workflow:

Revenue signal → Recovery case → Score → Diagnose → Decide → Guardrail → Execute → Outcome → Audit

The demo must prove business impact across a batch, not merely recommend an action. Its headline measures are revenue at risk, recovered revenue, recovery rate, AI-versus-baseline incremental recovery, and blocked / escalated actions.

2. Users and demo account

| User | Job to be done | Access | |---|---|---| | Merchant admin | Run a portfolio sweep, approve exceptional actions, view outcomes | Full dashboard | | Operator | Review a case’s evidence and execute approved work | Recovery dashboard | | Reviewer / judge | Reproduce the complete demo without real customer data | Seeded demo workspace | | Customer | Pay a single invoice without seeing merchant data | Public, token-based payment page |

The seeded account is intentionally a local demo identity, not a real Razorpay login and not a production credential:

ID:       razorpay
Password: razorpay

The visible ID maps only in non-production to the non-routable local address razorpay@invo-notify.test. It is never a real Razorpay account.

Razorpay is used only with merchant-provided Test Mode keys. No keys, payment credentials, or real customer data are committed to the repository.

3. P0 scope — what is demonstrably working

The product deliberately goes deep on B2B receivables rather than presenting shallow versions of every Razorpay direction.

• A — Ingestion (P0): signed Razorpay webhook intake plus tenant-scoped,

versioned v1 single/batch JSON or CSV ingestion, semantic validation, quarantine, persisted evidence, central recovery-case linkage, deduplication, and safe replay.

• B — Risk detection (P0): TypeScript logistic risk model, feature

contributions, expected recovery calculation, and model-version evidence.

• C — Diagnosis (P0): rules-first root-cause taxonomy, confidence,

replaceable evidence, and human-review fallback for unknown causes.

• D — Enrichment (P0): customer payment history, CIBIL proxy, VIP and

opt-out attributes, invoice balance, and due age.

• E — Decision (P0): typed rules/LLM decision agent with a fixed,

validated action allow-list.

• F — Execution (P0): payment-link and reminder adapters, simulated

failure fallback, and idempotent action evidence.

• G — Promise-to-pay (P1): scoped promise records, low-confidence review,

verified-payment reconciliation, and missed-promise escalation.

• H — Guardrails and audit (P0): deterministic policy checks, terminal

case protection, contact limits/cooldowns, approvals, guardrail evaluations, and append-only audit evidence.

• I — Analytics (P0): tenant-scoped recovery overview, funnel,

intervention/root-cause analytics, case drill-down IDs, and explicit simulated-versus-confirmed provenance.

4. Acceptance criteria for the P0 demo

• A demo batch contains diverse overdue invoices, including low risk, high value,

opt-out, low-value, partial-payment and already-recovered examples.

• A recovery sweep produces a risk score, expected recovery, recommended action,

policy verdict and audit record for every eligible case.

• An action cannot run for a paid, disputed, opted-out, throttled, or stopped

case. High-value or high-risk money actions require human approval.

• Razorpay Test Mode Payment Links can be configured through environment

variables; without keys, the simulation path still proves the workflow.

• A payment webhook is signature checked and idempotent, updates the invoice,

writes one durable recovery-settlement record, and closes the recovery case only once the invoice is fully paid.

• The dashboard drills from KPI → case → decision / guardrail / execution

evidence, and labels simulated outcomes honestly.

• The included test suite covers risk scoring, decision fallback, policy stopping

rules, rate limits and recovery flow.

5. Non-goals and safety boundaries

• This is not a fake Razorpay service or a place to store bank cards, UPI PINs,

customer tokens, or Razorpay credentials.

• The LLM may explain or rank an allowed action; it never calls a payment API or

bypasses the policy engine.

• The shipped metrics are simulation-backed until real, consented merchant

outcomes are available. The UI and docs must keep that label visible.

• WhatsApp, voice, live mandate retries and automatic promise extraction remain

out of P0 until consent, contact-window and provider controls are implemented.

6. Demo runbook

1. Run ./scripts/local-db.sh start, or configure another local PostgreSQL

database and the core environment variables.

1. Set DATABASE_URL and DIRECT_URL to that database, run migrations, then

pnpm ai:seed.

1. Run pnpm ai:verify to reproduce the safe, no-side-effect policy walkthrough.
2. Start the app and sign in with the demo account above.
3. Open AI Recovery, inspect the seeded portfolio, then select **Run Safe

Demo**. It persists recommendations and audit evidence without contacting a customer or provider.

1. Open cases to show differing decisions: a payment link, an approval request,

an opt-out block and a cost-to-recover stop.

1. Use Razorpay Test Mode only when keys/webhook configuration are present;

otherwise use the included deterministic simulation to show reproducible recovery outcomes.

1. Open Analytics and show recovered revenue, recovery rate, baseline

comparison, policy-block count and an audit drill-down.

7. Definition of done

P0 is complete only when the demo account, seeded dataset, recovery sweep, guardrails, audit drill-down and metrics all run from one documented workflow; tests and type checks pass; and every externally-integrated result is labeled as either Razorpay Test Mode or simulated.