How we keep this accurate.

Our reference data sources

We maintain a live mirror of the following sources:

• Appendix 4R (DTA RoDTEP rates) and Appendix 4RE (AA and EOU RoDTEP rates): DGFT publications. Updated within 24 hours of any notification.
• ITC HS (Indian Customs Tariff Schedule): CBIC. Refreshed annually with any chapter-level amendments.
• Customs Tariff Notifications: CBIC daily RSS, indexed by HS chapter.
• RoSCTL rate card: Ministry of Textiles. Tracked per notification.
• Drawback All Industry Rate schedule: CBIC annual notification, with mid-year amendments tracked.
• Sector-specific council circulars: APEDA, MPEDA, EEPC, GJEPC, Pharmexcil, EPCH, HEPC, CLE.
• RBI Master Direction on Exports: Monitored for changes to EDPMS, eBRC, and FEMA tolerance windows.

How we audit AI classification quality

Every HS classification produced by our model has three layers of validation:

Layer 1: schema validation. The model output is parsed and validated against a strict JSON schema. Hallucinated HS codes (codes that do not exist in ITC HS) are rejected at this layer.

Layer 2: rate cross-check. The rate the model cites for the suggested HS code is cross-checked against our reference copy of Appendix 4R or 4RE. If the model says "RoDTEP rate 2.4 percent for HS 6109.10" and our reference says 2.6 percent, the suggestion is flagged for review.

Layer 3: confidence calibration. Below 80 percent confidence, the suggestion is moved to manual review rather than presented as a primary recommendation. Above 80 percent, the suggestion is shown with the reasoning trace, the cited Appendix row, and a "common confusion" note about adjacent HS codes that this code is sometimes wrongly classified as.

We measure model performance on a 1,800-bill historical test set provided by our beta CA partners. We retrain or re-prompt when top-1 accuracy at 90 percent confidence drops below 96 percent.

What happens when we are wrong

ShippingBill.ai is informational software. We do not file your shipping bills, we do not represent you at customs, and we do not provide legal advice. The accept-or-override decision on every HS classification is yours (and your CA's). The filing decision is yours.

That said, our commitment is real. If a classification we surfaced was accepted by your CA and subsequently rejected by customs, and the rejection was due to a defect in our classification logic (not in your data or the CA's review), we will:

• Help you draft the response or revision filing at no additional charge.
• Refund the proportionate annual subscription tied to the affected period.
• Publish a post-mortem of the defect in our changelog so other customers benefit from the fix.

Liability is otherwise capped at total fees paid in the preceding 12 months, per our terms (see /terms).

How we handle DGFT trade notice changes

Our DGFT scraper runs every 30 minutes against the DGFT trade notices RSS and the CBIC notifications RSS. When a new notice is detected:

1. The notice is parsed and tagged by scheme, HS chapter, and effective date.
2. We identify all users with shipping bills filed in the prior 45 days whose claim is affected.
3. We compute the financial delta per affected user (per shipping bill).
4. We push WhatsApp + email alerts to each affected user the same day, with the rupee impact and suggested action.
5. The change is logged in our public changelog at /changelog.

Data residency, retention, and deletion

All customer data is stored in Supabase Mumbai region (ap-south-1) for DPDP compliance. We retain shipping bill, classification, and reconciliation data for the duration of your subscription plus 90 days for re-activation ease. On cancellation, you can request immediate deletion, which we honour within 7 working days, or wait for the 90-day automatic purge.

You can export your entire data at any time, in CSV and JSON, from /app/settings.

What is on our public changelog

We publish every product change (feature, fix, model update) at /changelog with the date, the change description, and any user-visible effect. Material changes to classification logic, rate references, or scheme eligibility are also pushed via WhatsApp to all affected users.