RefundIQ - One stop GST Refund Automation Platform
AI & Accounting

RefundIQ - One stop GST Refund Automation Platform

Author : CA. Kunal Budhwar

Watch on Youtube

USE CASE - AI IN GST REFUND AUTOMATION

RefundIQ - One stop GST Refund Automation Platform

A 4-layer infrastructure that ingests client books and GST returns, runs 23 pre-filing gap checks against the statute, computes the Rule 89 refund working, and prepares the complete RFD-01 pack in the exact format the GST portal accepts.

Author: CA Kunal Budhwar

Problem Statement

Across India, GST refund filings - exports under LUT, exports with payment of IGST, inverted duty structure, SEZ supplies, deemed exports and excess cash-ledger balances - are filed by lakhs of exporters and processed by 50,000-plus CA firms every quarter. Each filing follows the same skeleton: identify the eligible scenario, confirm the two-year time bar under Section 54(1), pull books and GSTR-1 / 2B / 3B for the period, reconcile the triangle, compute Net ITC and Adjusted Total Turnover per Rule 89(4) or 89(5), draft the statement (1A / 2 / 3 / 3A / 4 / 4A / 5B), assemble the Annexure-A declarations and the CA certificate, and file. The logic stays substantially the same; the data does not.

The bottleneck is not the formula - it is what happens before the formula runs. Books rarely reconcile with the portal returns out of the box. ITC claimed in 3B diverges from 2B; outward turnover in 1 diverges from 3B; LUT coverage lapses mid-year; deemed-export invoices get claimed twice (once by supplier, once by recipient); shipping bill numbers and FIRC dates are missing on export invoices; ineligible Section 17(5) ITC slips into the Net ITC pool. RefundIQ is built to surface these gaps before the RFD-01 is filed, not after a deficiency memo (RFD-03) lands.

Technology Used

RefundIQ is a multi-tenant SaaS for CA firms and direct exporters, built on Next.js 15 (App Router) and TypeScript on the web, FastAPI + Python on the compute layer, and Supabase Postgres with row-level security as the per-tenant data store. It consumes Tally XML directly via a thin TDL plugin (Send to RefundIQ), a Zoho Sigma extension for Zoho Books, and a CSV fallback. Returns are pulled from the GST portal by a Playwright bot with in-platform captcha relay - no GSP/ASP dependency. The refund engine is rule-based and deterministic: 23 codified pre-filing gap checks across eligibility, time bar, reconciliation, ITC purity and document completeness, plus verbatim Rule 89(4) / 89(5) formula engines tied to Accepted Architectural Decision Records. An LLM layer (Anthropic Claude, with a Groq Llama 3.3 70B fallback) sits on top, grounded on a curated GST corpus and 12 hard statutory anchors to keep explanations faithful to the bare act.

Solution / Key Features

Layer 1 - Eligibility & Scenario Lock. Maps the engagement to one of 23 refund scenarios (A1–A6 zero-rated, B1 inverted duty, C1/C2 deemed export, D1–D3 cash / excess / wrong-head, and the F/G special-category scenarios), checks the two-year bar under Section 54(1) Explanation (2), and confirms LUT coverage for the relevant period.

Layer 2 - Reconciliation Triangle. Books ⇄ GSTR-1 ⇄ GSTR-3B for outward; Books ⇄ GSTR-2B ⇄ GSTR-3B for ITC; per-invoice match between books purchases and 2B with canonical (supplier_gstin, invoice_no) key. Surfaces GAP-01 / 19 / 20 / 21 / 22 with sign-aware severity - over-claims as blockers (ASMT-10 / DRC-01C risk), under-claims as warnings (Section 16(4) permanent-loss risk).

Layer 3 - Refund Working (shipped). Computes Net ITC and Adjusted Total Turnover per Rule 89(5) for inverted duty (inputs-only per ADR-003) and Rule 89(4) for zero-rated (inputs + input services per ADR-004); also handles the post-2024 mechanical tax-paid path for deemed exports under ADR-012. Emits the maximum refund, the computation breakdown and the supporting per-invoice statement detail every CA needs for the working paper.

Layer 4 - Portal-Format Pack (shipping). Generates the RFD-01 Annexures in the exact GSTN offline-utility layout - Statement 1A, 2, 3, 3A, 4, 4A, 5B, plus Annexure B (ITC ledger) - so the file uploads directly to the portal without re-keying. Annexure-A declarations (passing-on, no-prosecution, Section 16(2)(c), no-drawback) and the CA certificate are auto-drafted from versioned templates in Word format, ready for review and signature.

Tally-native input. Direct XML ingestion via a minimal TDL plugin. No mapping templates, no re-keying.

Portal-faithful output. Statements are written into the exact RFD_STMT01A / RFD_ANNEXURE_B layouts the GST portal accepts - sheet names, header rows, column order, period format (MMYYYY), all preserved.

Deterministic, statute-cited engines. Every formula function carries the governing ADR reference and quotes verbatim Rule 89 / Section 54 text; nothing is a black box.

Reconciliation-first. The triangle (Books / GSTR-1 / GSTR-2B / GSTR-3B) is checked before the RFD-01 is drafted, not after a deficiency memo.

Multi-tenant by construction. Row-level security on every tenant table; cross-tenant isolation suite is part of CI.

AI explanatory layer. LLM grounded on a curated GST corpus and 12 hard statutory anchors (Section 54(14) ₹1,000 floor, 2-year bar, 7-day provisional refund, Section 36 72-month retention, Section 56 interest, Rule 89(4)/(5), Rule 96(10), FEMA realisation) so answers stay faithful to the bare act.

Roadmap

The 23-check engine, formula engines and portal-pack generators are decoupled from the scenario dispatcher - adding a refund track is a matter of plugging in a new generator and a new gap-rule subset. Tracks in the pipeline: full RFD-01 portal auto-fill via Playwright (the CA still presses Submit, EVC only); ICEGATE shipping-bill reconciliation for A1 Rule 96 filings; eMudhra-signed declarations (Phase 2+); white-label client portal + WhatsApp document collection (Plivo); status push to PracticeIQ for firms running both products; and the full Tally TDL with in-Tally pre-flight (Phase 5).

Impact

Pre-filing data clean-up time per refund engagement drops by 70–85%. A scored, partner-ready eligibility-and-gap report - with the Rule 89 working and a portal-ready Annexure pack - is produced in under two minutes from a Tally export and the three return files. The CA firm reclaims senior bandwidth that today goes into chasing exporters for shipping bills and FIRCs; the exporter gets a structured refund-readiness signal before money is locked up; and deficiency memos (RFD-03) drop sharply because the over-claim / under-claim / wrong-category gaps are caught before the filing leaves the firm.

Conclusion

The Rule 89 formula in Indian GST refunds is standardised; the inputs are not. RefundIQ scores eligibility first, reconciles the triangle next, computes the refund under the right sub-rule, and emits the RFD-01 pack in the format the GST portal accepts. Better reconciliation makes better refunds - measurable, visible, and traceable line-by-line to the statute.